window.ICML_PAPERS["68689"] = [{"id":"61007","en":"Local Redundancy: An Information-Theoretic Measure of Plasticity from Synthetic Memorization","ko":"국소 중복성: 합성 암기로부터 본 가소성의 정보이론적 척도","authors":"Jiaxuan Cheng","pos":"#4604","link":"https://openreview.net/forum?id=ucbH88BgIk","abs":"

Plasticity—a neural network's ability to adapt to new tasks—is critical for continual and transfer learning. Existing measures, such as effective rank, dead neuron fraction, and weight norm, lack theoretical grounding and correlate poorly with performance on new tasks. We introduce local redundancy, an information-theoretic measure derived from universal compression theory. We define local redundancy as the worst-case redundancy of a local model family—parameters in an infinitesimal neighborhood along gradient directions—and show this is a principled measure of plasticity. Although local redundancy is intractable to compute exactly, we prove that the expected squared gradient norm on a synthetic memorization task provides an efficiently computable lower bound. Experiments on continual image classification and time series transfer learning demonstrate that local redundancy predicts downstream performance better than existing measures and enables pretraining checkpoint selection where validation loss plateaus.

","absKo":"Plasticity, 즉 neural network가 새로운 task에 적응하는 능력은 continual learning과 transfer learning에 매우 중요하다. effective rank, dead neuron fraction, weight norm과 같은 기존 지표들은 이론적 근거가 부족하고 새로운 task 성능과의 상관도도 낮다. 우리는 universal compression theory로부터 유도된 정보이론적 척도인 local redundancy를 제안한다. local redundancy를 gradient directions를 따라 무한소 이웃에 있는 parameters로 이루어진 local model family의 worst-case redundancy로 정의하고, 이것이 plasticity의 원리적 측정치임을 보인다. local redundancy는 정확히 계산하기에는 다루기 어렵지만, synthetic memorization task에서의 expected squared gradient norm이 효율적으로 계산 가능한 lower bound를 제공함을 증명한다. continual image classification과 time series transfer learning 실험은 local redundancy가 기존 지표보다 downstream performance를 더 잘 예측하며, validation loss가 plateau를 이루는 상황에서 pretraining checkpoint selection을 가능하게 함을 보여준다."},{"id":"61161","en":"VEQ: Modality-Adaptive Quantization for MoE Vision-Language Models","ko":"VEQ: MoE Vision-Language Model을 위한 Modality-Adaptive Quantization","authors":"Guangshuo Qin, Zhiteng Li, Zheng Chen, Weihang Zhang, Linghe Kong, Yulun Zhang","pos":"#4605","link":"https://openreview.net/forum?id=tA8TFhuBDC","abs":"

Mixture-of-Experts(MoE) Vision-Language Models(VLMs) offer remarkable performance but incur prohibitive memory and computational costs, making compression essential. Post-Training Quantization (PTQ) is an effective training-free technique to address the massive memory and computation overhead. Existing quantization paradigms fall short as they are oblivious to two critical forms of heterogeneity: the inherent discrepancy between vision and language tokens, and the non-uniform contribution of different experts. To bridge this gap, we introduce Visual Expert Quantization (VEQ), a dual-aware quantization framework designed to simultaneously accommodate cross-modal differences and heterogeneity between experts. Specifically, VEQ incorporates 1)Modality-expert-aware Quantization, which utilizes expert activation frequency to prioritize error minimization for pivotal experts, and 2)Modality-affinity-aware Quantization, which constructs an enhanced Hessian matrix by integrating token-expert affinity with modality information to guide the calibration process. Extensive experiments across diverse benchmarks verify that VEQ consistently outperforms state-of-the-art baselines. Specifically, under the W3A16 configuration, our method achieves significant average accuracy gains of 2.04\\% on Kimi-VL and 3.09\\% on Qwen3-VL compared to the previous SOTA quantization methods, demonstrating superior robustness across various multi-modal tasks.

","absKo":"

Mixture-of-Experts(MoE) Vision-Language Models(VLMs)는 뛰어난 성능을 제공하지만, 막대한 memory와 computational cost를 초래하므로 compression이 필수적이다. Post-Training Quantization (PTQ)은 이러한 거대한 memory 및 computation overhead를 해결하기 위한 효과적인 training-free 기법이다. 기존 quantization paradigm은 두 가지 중요한 heterogeneity를 간과한다는 점에서 한계가 있다. 하나는 vision token과 language token 사이의 본질적인 차이이고, 다른 하나는 서로 다른 expert의 비균일한 기여도이다. 이러한 간극을 메우기 위해, 우리는 cross-modal 차이와 expert 간 heterogeneity를 동시에 수용하도록 설계된 dual-aware quantization framework인 Visual Expert Quantization (VEQ)을 제안한다. 구체적으로 VEQ는 1)Modality-expert-aware Quantization을 포함하는데, 이는 expert activation frequency를 활용해 핵심 expert의 error minimization를 우선시하며, 2)Modality-affinity-aware Quantization을 포함하는데, 이는 token-expert affinity와 modality 정보를 통합하여 강화된 Hessian matrix를 구성하고 calibration 과정을 안내한다. 다양한 benchmark에 걸친 광범위한 실험은 VEQ가 SOTA baseline을 일관되게 능가함을 검증한다. 특히 W3A16 configuration에서 우리의 방법은 이전 SOTA quantization method와 비교해 Kimi-VL에서 평균 정확도를 2.04\\%, Qwen3-VL에서 3.09\\% 향상시켰으며, 다양한 multi-modal task 전반에서 우수한 robustness를 보여준다.

"},{"id":"64932","en":"Geometric Flow Grounding: A Unified Manifold Decoupling Framework for Dynamics Discovery and Verification","ko":"Geometric Flow Grounding: 동역학 발견 및 검증을 위한 통합 Manifold Decoupling 프레임워크","authors":"Chang Yu, Yuxuan Luo, Yixuan Du, Yuqing Zhou, Siyuan Li, Jingbo Zhou, jiawei jiang, Zhen Lei, Stan Z Li","pos":"#1001","link":"https://openreview.net/forum?id=IYSBJvVRTx","abs":"

Modeling complex dynamics from observational data is fundamental to scientific discovery and artificial intelligence. However, existing approaches ranging from Neural ODEs to diffusion models are often plagued by the entanglement of static state representations and instantaneous motion, leading to accumulated errors and off-manifold hallucinations where predicted trajectories violate intrinsic geometric constraints. To address this, we propose Geometric Flow Grounding, a unified framework that enforces dynamic evolution strictly along the tangent bundle of the learned data manifold via a differentiable Neural Tangent Projection Layer. By geometrically decoupling state representation from tangential dynamics, our method generalizes across diverse data regimes. In the context of scientific discovery, we demonstrate that the projection layer eliminates numerical aliasing in sparse dynamical systems and recovers interpretable gene regulatory motifs from single-cell data by disentangling states from developmental velocities. Bridging to trustworthy AI, we further repurpose the geometric projection residual as a zero-shot metric for deepfake video detection, identifying generative inconsistencies against the implicit flow of pre-trained world models. Our results establish manifold-constrained projection as a universal operator for both discovering natural laws and verifying synthetic content.

","absKo":"

관측 데이터로부터 복잡한 dynamics를 모델링하는 것은 과학적 발견과 artificial intelligence의 근본 문제다. 그러나 Neural ODEs부터 diffusion model에 이르는 기존 접근법은 정적 상태 표현과 순간적 운동의 얽힘으로 인해 흔히 누적 오차와 off-manifold hallucination을 야기하며, 이로 인해 예측 궤적이 내재된 기하학적 제약을 위반한다. 이를 해결하기 위해 우리는 학습된 data manifold의 tangent bundle을 따라 동적 진화를 엄격히 강제하는 미분 가능한 Neural Tangent Projection Layer를 통해 구현되는 통합 프레임워크, Geometric Flow Grounding을 제안한다. 상태 표현과 접선 dynamics를 기하학적으로 분리함으로써, 우리 방법은 다양한 데이터 영역 전반으로 일반화된다. scientific discovery의 맥락에서, 우리는 projection layer가 sparse dynamical system에서 numerical aliasing을 제거하고, single-cell data에서 상태와 developmental velocity를 분리하여 해석 가능한 gene regulatory motif를 복원함을 보인다. trustworthy AI로 확장하여, 우리는 geometric projection residual을 deepfake video detection을 위한 zero-shot metric으로 재활용하여, pre-trained world model의 implicit flow에 비추어 generative inconsistency를 식별한다. 우리의 결과는 manifold-constrained projection이 자연 법칙을 발견하고 합성 콘텐츠를 검증하는 데 모두 유효한 보편적 operator임을 확립한다.

"},{"id":"61436","en":"LLM4Cov: Execution-Grounded Agent Learning for High-Coverage Hardware Verification","ko":"LLM4Cov: High-Coverage Hardware Verification을 위한 Execution-Grounded Agent Learning","authors":"Hejia Zhang, Zhongming Yu, Chia-Tung Ho, Mark Ren, Brucek Khailany, Jishen Zhao","pos":"#1003","link":"https://openreview.net/forum?id=q8AIRg06GX","abs":"

Execution-grounded LLM agents offer a promising paradigm for learning from tool feedback, but such feedback is often expensive and slow to obtain, making online reinforcement learning (RL) impractical. High-coverage hardware verification exemplifies this challenge due to its reliance on industrial simulators and non-differentiable execution signals. We propose LLM4Cov, an offline agent-learning framework that models verification as memoryless state transitions guided by deterministic evaluators. Building on this formulation, we introduce execution-validated data curation, policy-aware agentic data synthesis, and worst-state-prioritized sampling to enable scalable learning under execution constraints. We further curate a reality-aligned benchmark adapted from an existing verification suite through a revised evaluation protocol. Using the proposed pipeline, a compact 4B-parameter model achieves 69.2\\% coverage pass rate under agentic evaluation, outperforming its teacher by 5.3\\% and demonstrating competitive performance against models an order of magnitude larger.

","absKo":"

Execution-grounded LLM agent는 tool feedback으로부터 학습하는 유망한 패러다임을 제공하지만, 이러한 feedback은 얻는 데 비용이 많이 들고 느려 online reinforcement learning(RL)을 비현실적으로 만든다. High-coverage hardware verification은 industrial simulator와 non-differentiable execution signal에 의존한다는 점에서 이러한 문제를 잘 보여준다. 우리는 verification을 deterministic evaluator가 이끄는 memoryless state transition으로 모델링하는 offline agent-learning framework인 LLM4Cov를 제안한다. 이 formulation 위에서, 우리는 execution-validated data curation, policy-aware agentic data synthesis, 그리고 worst-state-prioritized sampling을 도입하여 execution constraint 하에서의 확장 가능한 학습을 가능하게 한다. 또한 기존 verification suite를 수정된 evaluation protocol로 재구성하여 현실에 맞춘 benchmark를 구축한다. 제안한 pipeline을 사용하면, compact한 4B-parameter model이 agentic evaluation에서 69.2\\% coverage pass rate를 달성하며, teacher를 5.3\\% 앞서고, 수십 배 더 큰 model들과 경쟁력 있는 성능을 보인다.

"},{"id":"65113","en":"LMM4-IC4K: A Large Multimodal Model Powered Integrated Circuit Footprint Geometry Understanding","ko":"LMM4-IC4K: Large Multimodal Model 기반 집적 회로 풋프린트 기하 구조 이해","authors":"Yida Wang, Taiting Lu, Runze Liu, Lanqing Yang, Zhe Chen, Yuehai Wang, Yixin Liu, Kaiyuan Lin, XiaoMeng Chen, Dian Ding, Yijie Li, Yifan Yang, Yi-Chao Chen, yincheng jin, Mahanth Gowda","pos":"#1004","link":"https://openreview.net/forum?id=GfJao7wf4s","abs":"

Printed-Circuit-board (PCB) footprint geometry labeling of integrated circuits (IC) is essential in defining the physical interface between components and the PCB layout, requiring precise visual perception. However, the unstructured nature of footprint drawings and abstract diagram annotations prevents direct IC footprint parsing and automated package geometry labeling methods from developing. Existing Large Multimodal Models (LMMs) struggle with inaccurate geometric perception, limiting their effectiveness in this task. To address these challenges, we propose LMM4-IC4K, a novel framework that treats IC mechanical drawings as images and leverages LMMs for structured geometric interpretation. To support such a framework, we introduce ICGeo8K, a multi-modal dataset with 8,608 labeled samples, including 4138 real-world IC footprint samples and 4470 synthetically generated samples. We further present a two-stage training framework to fine-tune LMMs for IC footprint labeling. Extensive experiments demonstrate that our model outperforms state-of-the-art LMMs on the proposed benchmark. The accurate translation of footprint diagrams enabled by LMM4-IC4K contributes to advancing automation and standardization within the PCB industry.

","absKo":"

집적회로(IC)의 Printed-Circuit-board (PCB) footprint geometry labeling은 구성 요소와 PCB layout 사이의 물리적 인터페이스를 정의하는 데 필수적이며, 정밀한 시각적 인식을 요구한다. 그러나 footprint 도면의 비정형성 및 추상적인 diagram annotation 때문에 직접적인 IC footprint parsing과 자동 package geometry labeling 방법의 개발이 어렵다. 기존의 Large Multimodal Models (LMMs)는 부정확한 geometric perception으로 인해 이 task에서의 효과가 제한된다. 이러한 문제를 해결하기 위해 우리는 IC 기계 도면을 이미지로 취급하고 LMM을 활용하여 구조화된 geometric interpretation을 수행하는 새로운 프레임워크인 LMM4-IC4K를 제안한다. 이러한 프레임워크를 지원하기 위해, 우리는 8,608개의 labeled sample을 포함하는 multi-modal dataset ICGeo8K를 도입한다. 여기에는 4,138개의 실제 IC footprint sample과 4,470개의 합성 생성 sample이 포함된다. 또한 IC footprint labeling을 위해 LMM을 fine-tune하는 2단계 training framework를 제시한다. 광범위한 실험은 우리의 모델이 제안한 benchmark에서 state-of-the-art LMM을 능가함을 보여준다. LMM4-IC4K가 가능하게 한 footprint diagram의 정확한 변환은 PCB 산업 내 자동화와 표준화의 발전에 기여한다.

"},{"id":"62251","en":"Recovering Policy-Induced Errors: Benchmarking and Trajectory Synthesis for Robust GUI Agents","ko":"정책 유도 오류 회복: Robust GUI Agents를 위한 벤치마킹과 궤적 합성","authors":"Tianpeng Bu, Xin Liu, Qihua Chen, Hao Jiang, Shurui Li, hongtao duan, Lu Jiang, lulu hu, Bin Yang, Minying Zhang","pos":"#1006","link":"https://openreview.net/forum?id=iJwDylm93H","abs":"

While GUI agents have advanced rapidly, they often lack the robustness to recover from their own errors, hindering real-world deployment. To bridge this gap at both the evaluation and data levels, we introduce GUI-RobustEval and propose Robustness-driven Trajectory Synthesis. GUI-RobustEval containing 1,216 executable test cases that systematically measure error recovery capabilities across a broad and realistic spectrum of error modes. At the data level, RoTS is a scalable synthesis framework that creates 800k high quality data via a tree-based pipeline that proactively discovers diverse error modes and synthesizes corresponding recovery steps. Our two models, RoTS-7B and RoTS-32B, fine-tuned on our dataset, both demonstrate significant gains on GUI-RobustEval and traditional GUI benchmarks. Notably, RoTS-32B achieves state-of-the-art performance on OSWorld, with a 47.4% success rate and a 33.8% All-Pass@4 score, indicating that the enhanced long-horizon error recovery ability synergistically boosts robustness and overall performance. We will release our benchmark, dataset, and models to facilitate future research.

","absKo":"

GUI agent는 빠르게 발전해 왔지만, 종종 스스로의 오류를 복구하는 견고성이 부족해 실제 배포를 가로막는다. 평가와 데이터 양쪽 수준에서 이 간극을 메우기 위해, 우리는 GUI-RobustEval을 도입하고 Robustness-driven Trajectory Synthesis를 제안한다. GUI-RobustEval은 다양한 실제적 오류 모드 전반에서 오류 복구 능력을 체계적으로 측정하는 1,216개의 실행 가능한 테스트 케이스로 구성된다. 데이터 측면에서 RoTS는 tree-based 파이프라인을 통해 80만 개의 고품질 데이터를 생성하는 확장 가능한 합성 프레임워크로, 다양한 오류 모드를 선제적으로 발견하고 그에 대응하는 복구 단계를 합성한다. 우리의 데이터셋으로 fine-tuning한 두 모델 RoTS-7B와 RoTS-32B는 모두 GUI-RobustEval과 전통적인 GUI benchmark에서 유의미한 성능 향상을 보인다. 특히 RoTS-32B는 OSWorld에서 47.4% success rate와 33.8% All-Pass@4 score를 기록하며 state-of-the-art 성능을 달성했는데, 이는 향상된 장기적 오류 복구 능력이 견고성과 전체 성능을 시너지 있게 끌어올린다는 점을 보여준다. 우리는 향후 연구를 촉진하기 위해 benchmark, dataset, model을 공개할 예정이다.

"},{"id":"65853","en":"SE-GA: Memory-Augmented Self-Evolution for GUI Agents","ko":"SE-GA: GUI 에이전트를 위한 메모리 증강 자기진화","authors":"shilong jin, Lanjun Wang, Zhuosheng Zhang","pos":"#1007","link":"https://openreview.net/forum?id=9S9V5ntQJM","abs":"

Autonomous Graphical User Interface (GUI) agents often struggle with multi-step tasks due to constrained context windows and static policies that fail to adapt to dynamic environments. To address these limitations, this work proposes the Self-Evolving GUI Agent (SE-GA), a novel framework that integrates hierarchical memory structures with an iterative self-improvement mechanism. At the core of our approach is Test-Time Memory Extension (TTME), which facilitates long-term planning by dynamically retrieving episodic, semantic, and experiential memories to provide salient contexts during inference. To ensure continuous learning, we introduce Memory-Augmented Self-Evolution (MASE), which is a training pipeline that adopts the data collected by TTME to stabilize and enhance the agent's foundational policy. Extensive evaluations across both offline and online benchmarks demonstrate SE-GA achieves state-of-the-art performance, reaching success rates of 89.0\\% on ScreenSpot and 75.8\\% on the challenging AndroidControl-High dataset. Furthermore, significant improvements on the AndroidWorld benchmark highlight the superior generalization to dynamic environments.

","absKo":"

Autonomous Graphical User Interface (GUI) agent는 제한된 context window와 동적 환경에 적응하지 못하는 static policy 때문에 다단계 작업에서 자주 어려움을 겪는다. 이러한 한계를 해결하기 위해, 본 연구는 계층적 memory structure와 반복적 self-improvement 메커니즘을 통합한 새로운 프레임워크 Self-Evolving GUI Agent (SE-GA)를 제안한다. 우리 접근의 핵심에는 Test-Time Memory Extension (TTME)이 있으며, 이는 episodic, semantic, experiential memory를 동적으로 검색하여 추론 시 중요한 context를 제공함으로써 장기 계획을 가능하게 한다. 지속적인 학습을 보장하기 위해, 우리는 Memory-Augmented Self-Evolution (MASE)을 도입한다. 이는 TTME가 수집한 데이터를 활용하여 agent의 foundational policy를 안정화하고 강화하는 training pipeline이다. 오프라인과 온라인 benchmark 전반에 대한 광범위한 평가에서 SE-GA는 state-of-the-art 성능을 달성했으며, ScreenSpot에서 89.0\\%, 까다로운 AndroidControl-High 데이터셋에서 75.8\\%의 성공률을 기록했다. 더 나아가 AndroidWorld benchmark에서의 유의미한 향상은 동적 환경에 대한 우수한 일반화 능력을 강조한다.

"},{"id":"64033","en":"SKETCH: Semantic Key-Point Conditioning for Long-Horizon Vessel Trajectory Prediction","ko":"SKETCH: 장기 예측 선박 궤적을 위한 시맨틱 키포인트 조건화","authors":"Linyong Gan, Zimo Li, Wenxin XU, Xingjian Li, Jianhua Huang, Enmei Tu, Shuhang Chen","pos":"#1008","link":"https://openreview.net/forum?id=RYTZ3Y2mBJ","abs":"

Accurate long-horizon vessel trajectory prediction remains challenging due to compounded uncertainty from complex navigation behaviors and environmental factors. Existing methods often struggle to maintain global directional consistency, leading to drifting or implausible trajectories when extrapolated over long time horizons. To address this issue, we propose a semantic-key-point-conditioned trajectory modeling framework, in which future trajectories are predicted by conditioning on a high-level Next Key Point (NKP) that captures navigational intent. This formulation decomposes long-horizon prediction into global semantic decision-making and local motion modeling, effectively restricting the support of future trajectories to semantically feasible subsets. To efficiently estimate the NKP prior from historical observations, we adopt a pretrain-finetune strategy. Extensive experiments on real-world AIS data demonstrate that the proposed method consistently outperforms state-of-the-art approaches, particularly for long travel durations, directional accuracy, and fine-grained trajectory prediction.

","absKo":"

정확한 long-horizon vessel trajectory prediction은 복잡한 navigation behavior와 환경 요인으로 인한 복합적인 uncertainty 때문에 여전히 어렵습니다. 기존 방법은 global directional consistency를 유지하는 데 종종 실패하여, 긴 시간 범위로 extrapolation할 때 drift하거나 그럴듯하지 않은 trajectory를 생성합니다. 이 문제를 해결하기 위해 우리는 semantic-key-point-conditioned trajectory modeling framework를 제안하며, 여기서 future trajectory는 navigational intent를 포착하는 고수준 Next Key Point (NKP)에 조건화하여 예측됩니다. 이 formulation은 long-horizon prediction을 global semantic decision-making과 local motion modeling으로 분해하여, future trajectory의 support를 semantic적으로 feasible한 subset으로 효과적으로 제한합니다. historical observation으로부터 NKP prior를 효율적으로 추정하기 위해 우리는 pretrain-finetune strategy를 채택합니다. 실제 AIS data에 대한 광범위한 실험은 제안한 방법이 특히 긴 travel duration, directional accuracy, fine-grained trajectory prediction에서 state-of-the-art 접근법을 일관되게 능가함을 보여 줍니다.

"},{"id":"66800","en":"“Do Diffusion Models Dream of Electric Planes?” Discrete and Continuous Simulation-Based Inference for Aircraft Design","ko":"\"Diffusion 모델은 전기 비행기 꿈을 꾸는가?\" 항공기 설계를 위한 이산 및 연속 Simulation-Based Inference","authors":"Aurelien Ghiglino, Daniel Elenius, Anirban Roy, Ramneet Kaur, Manoj Acharya, Colin Samplawski, Brian Matejek, Susmit Jha, Juan Alonso, Adam Cobb","pos":"#1009","link":"https://openreview.net/forum?id=0COq6zy5kN","abs":"

In this paper, we generate conceptual engineering designs of electric vertical take-off and landing (eVTOL) aircraft. We follow the paradigm of simulation-based inference (SBI), whereby we look to learn a posterior distribution over the full eVTOL design space. To learn this distribution, we must sample over discrete aircraft configurations (topologies) and their corresponding set of continuous parameters. Therefore, we introduce a hierarchical probabilistic model consisting of two diffusion models. The first model leverages recent work on Riemannian Diffusion Language Modeling (RDLM) and Unified World Models (UWMs) to enable us to sample topologies from a discrete and continuous space. For the second model we introduce a masked diffusion approach to sample the corresponding parameters conditioned on the topology. We show our approach successfully rediscovers known trends and captures governing physical laws in aircraft design.

","absKo":"

이 논문에서는 electric vertical take-off and landing (eVTOL) aircraft의 개념적 engineering design을 생성한다. 우리는 simulation-based inference (SBI) 패러다임을 따르며, 전체 eVTOL design space에 대한 posterior distribution을 학습하고자 한다. 이 분포를 학습하기 위해서는 discrete aircraft configuration(topology)와 그에 대응하는 continuous parameter 집합을 모두 샘플링해야 한다. 따라서 우리는 두 개의 diffusion model로 구성된 hierarchical probabilistic model을 도입한다. 첫 번째 model은 Riemannian Diffusion Language Modeling (RDLM)과 Unified World Models (UWMs)에 대한 최근 연구를 활용하여 discrete 및 continuous space로부터 topology를 샘플링할 수 있게 한다. 두 번째 model에서는 topology를 조건으로 해당 parameter를 샘플링하기 위해 masked diffusion 접근법을 도입한다. 우리는 제안한 접근이 항공기 설계에서 알려진 경향을 성공적으로 재발견하고 지배적인 물리 법칙을 포착함을 보인다.

"},{"id":"62527","en":"Beyond Static Endpoints: Tool Programs as an Interface for Flexible Agentic Web Services","ko":"정적 엔드포인트를 넘어서: 유연한 Agentic Web Service를 위한 Tool Program 인터페이스","authors":"Mugeng Liu, Shuoqi Li, Yixuan Zhang, Yun Ma","pos":"#1014","link":"https://openreview.net/forum?id=fmwNYkYRN6","abs":"

In the agentic web era, LLM-based agents increasingly invoke web services as tools, yet most interfaces are still exposed as \\emph{static endpoints}. As tasks grow into long-horizon workflows with loops, conditionals, joins, and retries, agents must externalize control flow into stepwise calls and multi-round reasoning, causing excessive network turns, over-/under-fetching, and brittle recovery under partial failures. We argue that the bottleneck is representational: endpoint sequences are a poor interface for expressing tool intent. We present ToolPro, which represents an agent’s tool intent as an \\emph{executable tool program}---a composable program that compactly encodes multi-step service interactions with explicit effect types. ToolPro combines (i) constraint-guided program construction with compiler/runtime feedback to improve executability of LLM-produced code, (ii) effect-aware replay that provides exactly-once semantics for state-modifying calls across iterative repair and re-execution, and (iii) a profile-driven consolidation policy that decides when program execution outperforms stepwise calling. We instantiate ToolPro over MCP-style services with WebAssembly sandboxing. Across three real applications and six workflows, ToolPro reduces end-to-end latency by up to 37.2\\% and client-side traffic by up to 85.3\\%, with larger gains under higher network latency and increased workflow complexity. We believe that ToolPro lays a foundation for agent-facing service interfaces in the emerging agentic web.

","absKo":"

agentic web 시대에 들어서면서 LLM-based agent가 web service를 tool로 호출하는 일이 점점 많아지고 있지만, 대부분의 interface는 여전히 \\emph{static endpoint}로 노출되어 있다. 작업이 loop, conditional, join, retry를 포함하는 장기적 workflow로 커질수록, agent는 control flow를 stepwise call과 multi-round reasoning으로 외부화해야 하며, 그 결과 과도한 network turn, over-/under-fetching, 그리고 부분 실패 상황에서 취약한 복구가 발생한다. 우리는 병목이 representation에 있다고 주장한다. 즉, endpoint sequence는 tool intent를 표현하기에 부적절한 interface이다. 우리는 agent의 tool intent를 \\emph{executable tool program}으로 표현하는 ToolPro를 제안한다. 이는 explicit effect type을 사용해 multi-step service interaction을 간결하게 인코딩하는 composable program이다. ToolPro는 (i) compiler/runtime feedback을 이용한 constraint-guided program construction으로 LLM이 생성한 code의 실행 가능성을 높이고, (ii) iteration 기반 repair 및 재실행 전반에 걸쳐 state-modifying call에 대해 exactly-once semantics를 제공하는 effect-aware replay를 제공하며, (iii) program execution이 stepwise calling보다 유리한 시점을 결정하는 profile-driven consolidation policy를 결합한다. 우리는 WebAssembly sandboxing과 함께 MCP-style service 위에 ToolPro를 구현한다. 세 개의 실제 application과 여섯 개의 workflow에서 ToolPro는 end-to-end latency를 최대 37.2\\%까지, client-side traffic을 최대 85.3\\%까지 줄였으며, network latency가 높고 workflow complexity가 증가할수록 더 큰 개선을 보였다. 우리는 ToolPro가 emerging agentic web에서 agent-facing service interface의 기반을 마련한다고 본다.

"},{"id":"65440","en":"Generative Inverse Design with Abstention via Diagonal Flow Matching","ko":"대각 Flow Matching을 통한 기권 기능이 있는 생성적 역설계","authors":"Miguel de Campos, Werner Krebs, Hanno Gottschalk","pos":"#1100","link":"https://openreview.net/forum?id=DdUriffJhf","abs":"Inverse design aims to find design parameters $x$ achieving target performance $y^*$. Generative approaches learn bidirectional mappings between designs and labels, enabling diverse solution sampling. However, standard conditional flow matching (CFM), when adapted to inverse problems by pairing labels with design parameters, exhibits strong sensitivity to their arbitrary ordering and scaling, leading to unstable training. We introduce Diagonal Flow Matching (Diag-CFM), which resolves this through a zero-anchoring strategy that pairs design coordinates with noise and labels with zero, making the learning problem provably invariant to coordinate permutations. This yields order-of-magnitude improvements in round-trip accuracy over CFM and invertible neural network baselines across design dimensions up to $P{=}100$. We develop two architecture-intrinsic uncertainty metrics, Zero-Deviation and Self-Consistency, that enable three practical capabilities: selecting the best candidate among multiple generations, abstaining from unreliable predictions, and detecting out-of-distribution targets; consistently outperforming ensemble and general-purpose alternatives across all tasks. We validate on airfoil, gas turbine combustor, and an analytical benchmark with scalable design dimension.","absKo":"Inverse design은 target performance $y^*$를 달성하는 design parameter $x$를 찾는 것을 목표로 한다. Generative approach는 design과 label 사이의 bidirectional mapping을 학습하여 다양한 해의 샘플링을 가능하게 한다. 그러나 표준 conditional flow matching(CFM)을 inverse problem에 맞게 design parameter와 label을 짝짓는 방식으로 적용하면, 이들의 임의적인 ordering과 scaling에 매우 민감하게 반응하여 불안정한 training을 유발한다. 우리는 Diagonal Flow Matching(Diag-CFM)을 도입하여, design coordinate를 noise와 짝짓고 label을 zero와 짝짓는 zero-anchoring strategy를 통해 이 문제를 해결하며, 학습 문제가 coordinate permutation에 대해 provably invariant하도록 만든다. 그 결과, 최대 $P{=}100$의 design dimension에 걸쳐 CFM 및 invertible neural network baseline 대비 round-trip accuracy가 자릿수 수준으로 향상된다. 또한 우리는 Zero-Deviation과 Self-Consistency라는 두 개의 architecture-intrinsic uncertainty metric을 개발하여, 여러 generation 중 최적 후보 선택, 신뢰할 수 없는 예측에 대한 abstention, out-of-distribution target 탐지라는 세 가지 실용적 기능을 가능하게 하며, 모든 task에서 ensemble 및 general-purpose 대안보다 일관되게 우수함을 보인다. 우리는 airfoil, gas turbine combustor, 그리고 확장 가능한 design dimension을 갖는 analytical benchmark에서 이를 검증한다."},{"id":"63179","en":"Active Timepoint Selection for Learning Measure-Valued Trajectories","ko":"측정값 궤적 학습을 위한 능동적 시점 선택","authors":"Nicolas Huynh, Mihaela van der Schaar","pos":"#1103","link":"https://openreview.net/forum?id=ZLraoxBly2","abs":"

Inferring continuous probability paths from sparse snapshots is a fundamental challenge in domains like single-cell biology, where high-fidelity data acquisition is often destructive and constrained by prohibitive sequencing costs. This motivates the need for active learning strategies to strategically select optimal measurement times. However, designing active learning policies for this setting remains an open problem: the target objects reside on the infinite dimensional Wasserstein space where standard Euclidean metrics are ill-defined, and current interpolation methods lack epistemic uncertainty quantification. We introduce a framework which extends active experimentation to the space of measures. By leveraging Linearized Optimal Transport (LOT), we map distributional snapshots into a tangent space amenable to Gaussian Process modeling, allowing us to construct a tractable probabilistic surrogate for the underlying probability path. This yields a geometric acquisition function that iteratively selects measurement times to minimize uncertainty. Empirical results demonstrate that our strategy outperforms uncertainty-agnostic baselines on both synthetic and real-world datasets.

","absKo":"

희소한 snapshot으로부터 continuous probability path를 추론하는 것은 single-cell biology와 같은 분야에서 근본적인 과제이다. 이 분야에서는 고충실도 데이터 수집이 종종 파괴적이며, 비싼 sequencing cost 때문에 제약을 받는다. 이는 최적의 measurement time을 전략적으로 선택하기 위한 active learning strategy의 필요성을 낳는다. 그러나 이러한 설정에서의 active learning policy를 설계하는 일은 여전히 열린 문제로 남아 있다. 목표 대상이 표준 Euclidean metric이 정의되기 어려운 infinite dimensional Wasserstein space 위에 존재하고, 현재의 interpolation method에는 epistemic uncertainty quantification이 부족하기 때문이다. 우리는 measure의 공간으로 active experimentation을 확장하는 framework를 제안한다. Linearized Optimal Transport (LOT)를 활용해 distribution snapshot을 Gaussian Process modeling에 적합한 tangent space로 매핑하고, 이를 통해 underlying probability path에 대한 tractable probabilistic surrogate를 구성한다. 그 결과, uncertainty를 최소화하도록 measurement time을 반복적으로 선택하는 geometric acquisition function을 얻는다. 실험 결과는 우리의 전략이 synthetic 및 real-world dataset 모두에서 uncertainty-agnostic baseline보다 우수함을 보여준다.

"},{"id":"63701","en":"4DPC$^2$hat: Towards Dynamic Point Cloud Understanding with Failure-Aware Bootstrapping","ko":"4DPC$^2$hat: 실패 인지 부트스트래핑을 활용한 동적 포인트 클라우드 이해를 향하여","authors":"Xindan Zhang, Weilong Yan, YUFEI SHI, Xuerui Qiu, Tao He, Ying Li, Ming Li, Hehe Fan","pos":"#1104","link":"https://openreview.net/forum?id=UUJQGmt5GS","abs":"Point clouds provide a compact and expressive representation of 3D objects, and have recently been integrated into multimodal large language models (MLLMs). However, existing methods primarily focus on static objects, while understanding dynamic point cloud sequences remains largely unexplored. This limitation is mainly caused by the lack of large-scale cross-modal datasets and the difficulty of modeling motions in spatio-temporal contexts. To bridge this gap, we present 4DPC$^2$hat, the *first* MLLM tailored for dynamic point cloud understanding. To this end, we construct a large-scale cross-modal dataset 4DPC$^2$hat-200K via a meticulous two-stage pipeline consisting of topology-consistent 4D point construction and two-level captioning. The dataset contains over 44K dynamic object sequences, 700K point cloud frames, and 200K curated question–answer (QA) pairs, supporting inquiries about counting, temporal relationship, action, spatial relationship, and appearance. At the core of the framework, we introduce a Mamba-enhanced temporal reasoning MLLM to capture long-range dependencies and dynamic patterns among a point cloud sequence. Furthermore, we propose a *failure-aware* bootstrapping learning strategy that iteratively identifies model deficiencies and generates targeted QA supervision to continuously strengthen corresponding reasoning capabilities. Extensive experiments demonstrate that our 4DPC$^2$hat significantly improves action understanding and temporal reasoning compared with existing models, establishing a strong foundation for 4D dynamic point cloud understanding.","absKo":"Point cloud는 3D object를 compact하고 expressive하게 표현할 수 있으며, 최근 multimodal large language model(MLLM)에 통합되었다. 그러나 기존 방법은 주로 static object에 초점을 맞추고 있어, dynamic point cloud sequence를 이해하는 문제는 여전히 충분히 탐구되지 않았다. 이러한 한계는 대규모 cross-modal dataset의 부족과 spatio-temporal context에서 motion을 모델링하는 어려움에서 주로 비롯된다. 이 격차를 메우기 위해 우리는 dynamic point cloud understanding에 특화된 *최초의* MLLM인 4DPC$^2$hat을 제안한다. 이를 위해 topology-consistent 4D point construction과 two-level captioning으로 구성된 정교한 two-stage pipeline을 통해 대규모 cross-modal dataset 4DPC$^2$hat-200K를 구축한다. 이 dataset은 44K개 이상의 dynamic object sequence, 700K개의 point cloud frame, 그리고 200K개의 정제된 question–answer(QA) pair를 포함하며, counting, temporal relationship, action, spatial relationship, appearance에 대한 질의를 지원한다. framework의 핵심에서는 point cloud sequence 내의 장기 의존성과 동적 패턴을 포착하기 위해 Mamba-enhanced temporal reasoning MLLM을 도입한다. 또한 우리는 모델의 결함을 반복적으로 식별하고 해당 reasoning capability를 지속적으로 강화할 targeted QA supervision을 생성하는 *failure-aware* bootstrapping learning strategy를 제안한다. 광범위한 실험은 우리 4DPC$^2$hat이 기존 모델에 비해 action understanding과 temporal reasoning을 크게 향상시킴을 보여 주며, 4D dynamic point cloud understanding을 위한 강력한 foundation을 확립한다."},{"id":"65379","en":"Flow for Future: Geometric SE(3)-Equivariant Flow Matching for 3D Trajectory Prediction","ko":"미래를 위한 Flow: 3D 궤적 예측을 위한 기하학적 SE(3)-Equivariant Flow Matching","authors":"Junwei Wu, Yihang Liu, Ruixuan Yu, Jian Sun","pos":"#1310","link":"https://openreview.net/forum?id=EBujA4tldV","abs":"

Predicting 3D geometric trajectory requires capturing complex spatiotemporal dependencies while preserving physical symmetries. While flow matching offers a powerful generative paradigm, extending it to SE(3)-equivariant dynamics is challenging due to the inherent gap between deterministic history and stochastic evolving flows. To address this, we introduce GSE-Flow, an SE(3)-equivariant flow matching framework. We first propose a Coherent Sequence Encoding and Time-Modulated Embedding strategy that unifies historical and evolving streams, incorporating velocity and flow time via equivariant affine transformations to guide continuous evolution. We further design a Geometry-Feature Tensorization mechanism that projects node states into a tensor product space, enabling Context-Flow Fusion to guide trajectory evolution with historical context. GSE-Flow guarantees theoretical SE(3)-equivariance and achieves SOTA accuracy on MD17, MD22, and CMU MoCap benchmarks for geometric trajectory prediction, while demonstrating generality by enhancing deterministic baselines.

","absKo":"

3D geometric trajectory를 예측하려면 physical symmetry를 보존하면서 복잡한 spatiotemporal dependency를 포착해야 한다. flow matching은 강력한 generative paradigm이지만, deterministic history와 stochastic evolving flow 사이의 본질적인 간극 때문에 이를 SE(3)-equivariant dynamics로 확장하는 것은 어렵다. 이를 해결하기 위해, 우리는 SE(3)-equivariant flow matching framework인 GSE-Flow를 제안한다. 먼저, historical stream과 evolving stream을 하나로 통합하는 Coherent Sequence Encoding과 Time-Modulated Embedding 전략을 제안하며, equivariant affine transformation을 통해 velocity와 flow time을 통합하여 continuous evolution을 유도한다. 이어서 node state를 tensor product space로 사상하는 Geometry-Feature Tensorization 메커니즘을 설계하여, Context-Flow Fusion이 historical context를 활용해 trajectory evolution을 안내할 수 있도록 한다. GSE-Flow는 이론적으로 SE(3)-equivariance를 보장하며, geometric trajectory prediction에서 MD17, MD22, CMU MoCap benchmark 전반에 걸쳐 SOTA 정확도를 달성한다. 또한 deterministic baseline을 향상시키는 방식으로 일반성도 입증한다.

"},{"id":"66135","en":"Physics-informed coarsening for multigrid graph neural networks surrogates","ko":"멀티그리드 그래프 신경망 대리 모델을 위한 물리 정보 기반 조대화","authors":"Amir Bazzi, Ramy Nemer, Alves José, Elie Hachem","pos":"#1500","link":"https://openreview.net/forum?id=6aFl5SegYt","abs":"

Learning-based surrogates for partial differential equations have recently matched the accuracy of classical solvers while achieving orders-of-magnitude speedups, predominantly in fluid settings and structured geometries. In contrast, robust surrogates for deformable solids remain underexplored, despite the presence of nonlinear elasticity, plasticity, and transient behavior that challenge standard architectures. We introduce a multigrid graph neural network for solid mechanics that couples an encoder-processor-decoder backbone with a physics-informed coarsening strategy. Instead of downsampling via geometric heuristics, our method scores nodes using a residual-based measure of local physical activity and preferentially retains regions of high strain or stress concentration, allocating multiscale capacity where it is most needed. This preserves long-range interactions through hierarchical message passing while improving stability over long rollouts. We evaluate on multiple datasets covering linear, nonlinear, and transient regimes, and observe consistent gains in accuracy and rollout stability compared to standard sampling baselines. Our results highlight the importance of physics-informed coarsening for scalable surrogate modeling in solid mechanics.

","absKo":"

부분 미분 방정식에 대한 learning-based surrogate는 최근 classical solver와 맞먹는 정확도를 달성하면서도 수십 배에서 수백 배의 속도 향상을 보였으며, 주로 fluid 설정과 structured geometry에서 두드러졌다. 반면 deformable solid를 위한 robust surrogate는 nonlinear elasticity, plasticity, transient behavior가 standard architecture에 도전이 됨에도 불구하고 아직 충분히 연구되지 않았다. 우리는 solid mechanics를 위한 multigrid graph neural network를 제안하며, encoder-processor-decoder backbone과 physics-informed coarsening strategy를 결합한다. 기하학적 heuristic으로 downsampling하는 대신, 우리 방법은 residual-based measure로 local physical activity를 평가하고, 변형률(strain)이나 응력(stress) 집중이 높은 영역을 우선적으로 보존하여 multiscale capacity를 가장 필요한 곳에 할당한다. 이는 hierarchical message passing을 통해 장거리 상호작용을 유지하면서도 긴 rollout에서의 안정성을 향상시킨다. 우리는 linear, nonlinear, transient regime을 포괄하는 여러 dataset에서 평가하였고, standard sampling baseline과 비교해 정확도와 rollout 안정성에서 일관된 향상을 관찰했다. 우리의 결과는 solid mechanics에서 scalable surrogate modeling을 위해 physics-informed coarsening이 중요함을 강조한다.

"},{"id":"61127","en":"DOCKSMITH: Scaling Reliable Coding Environments via an Agentic Docker Builder","ko":"DOCKSMITH: Agentic Docker Builder를 통한 신뢰 가능한 코딩 환경 확장","authors":"Jiaran Zhang, Lu Ma, Yanhao Li, Fanqi Wan, DI QI, Xin Wu, Zhewei Huang, Liangyu Chen, YINGWEI MA, Qi Han, Xiangyu Zhang","pos":"#202","link":"https://openreview.net/forum?id=tRbgtWwmHB","abs":"

Reliable Docker-based environment construction is a dominant bottleneck for scaling execution-grounded training and evaluation of software engineering agents. We introduce DockSmith, a specialized agentic Docker builder designed to address this challenge. DockSmith treats environment construction not merely as a preprocessing step, but as a core agentic capability that exercises long-horizon tool use, dependency reasoning, and failure recovery, yielding supervision that transfers beyond Docker building itself. DockSmith is trained on large-scale, execution-grounded Docker-building trajectories produced by a SWE-Factory--style pipeline augmented with a loop-detection controller and a cross-task success memory. Training a 30B-A3B model on these trajectories achieves open-source state-of-the-art performance on Multi-Docker-Eval, with 39.72\\% Fail-to-Pass and 58.28\\% Commit Rate. Moreover, DockSmith improves out-of-distribution performance on SWE-bench Verified, SWE-bench Multilingual, and Terminal-Bench 2.0, demonstrating the broader agentic benefits of environment construction. Our model and Docker-building trajectories are publicly available at https://huggingface.co/collections/8sj7df9k8m5x8/docksmith.

","absKo":"

신뢰할 수 있는 Docker 기반 환경 구축은 소프트웨어 엔지니어링 에이전트의 execution-grounded training과 evaluation을 확장하는 데 있어 지배적인 병목이다. 우리는 이 문제를 해결하기 위해 특화된 agentic Docker builder인 DockSmith를 소개한다. DockSmith는 environment construction을 단순한 preprocessing step이 아니라, 장기적 tool use, dependency reasoning, failure recovery를 수행하는 핵심 agentic capability로 다루며, 그 결과 Docker building 자체를 넘어 전이되는 supervision을 생성한다. DockSmith는 loop-detection controller와 cross-task success memory로 보강된 SWE-Factory 스타일 pipeline이 생성한 대규모 execution-grounded Docker-building trajectory로 학습된다. 이러한 trajectory로 30B-A3B 모델을 학습하면 Multi-Docker-Eval에서 open-source state-of-the-art 성능을 달성하며, 39.72\\% Fail-to-Pass와 58.28\\% Commit Rate를 기록한다. 또한 DockSmith는 SWE-bench Verified, SWE-bench Multilingual, Terminal-Bench 2.0에서 out-of-distribution 성능을 향상시켜, environment construction이 지닌 더 넓은 agentic 이점을 보여준다. 우리 모델과 Docker-building trajectory는 https://huggingface.co/collections/8sj7df9k8m5x8/docksmith 에서 공개되어 있다.

"},{"id":"66245","en":"Olaf-World: Orienting Latent Actions for Video World Modeling","ko":"Olaf-World: 비디오 World Modeling을 위한 잠재 행동 방향 설정","authors":"Yuxin Jiang, Yuchao Gu, Ivor Tsang, Mike Zheng Shou","pos":"#2811","link":"https://openreview.net/forum?id=5TiuerrwR8","abs":"Scaling action-controllable world models is limited by the scarcity of action labels. While latent action learning promises to extract control interfaces from unlabeled video, learned latents often fail to transfer across contexts: they entangle scene-specific cues and lack a shared coordinate system. This occurs because standard objectives operate only *within* each clip, providing no mechanism to align action semantics across contexts. Our key insight is that although actions are unobserved, their *semantic effects* are observable and can serve as a shared reference. We introduce **Seq$\\Delta$-REPA**, a sequence-level control-effect alignment objective that anchors integrated latent action to temporal feature differences from a frozen, self-supervised video encoder. Building on this, we present **Olaf-World**, a pipeline that pretrains action-conditioned video world models from large-scale passive video. Extensive experiments demonstrate that our method learns a more structured latent action space, leading to stronger zero-shot action transfer and more data-efficient adaptation to new control interfaces than state-of-the-art baselines.","absKo":"행동 제어 가능한 world model의 확장은 action label의 희소성에 의해 제약을 받는다. latent action learning은 라벨이 없는 비디오에서 control interface를 추출할 수 있다는 가능성을 제시하지만, 학습된 latent는 종종 맥락 간 전이에서 실패한다. 즉, 장면 특유의 단서를 얽어내고 공유 좌표계를 갖지 못한다. 이는 표준 objective가 각 clip 내부에서만 작동하며, 맥락 간 action semantics를 정렬할 메커니즘을 제공하지 않기 때문이다. 우리의 핵심 통찰은 action은 관측되지 않더라도 그 semantic effect는 관측 가능하며, 공유 reference로 사용할 수 있다는 점이다. 우리는 **Seq$\\Delta$-REPA**를 제안하는데, 이는 통합된 latent action을 고정된 self-supervised video encoder의 temporal feature difference에 정렬시키는 sequence-level control-effect alignment objective이다. 이를 바탕으로, 대규모 passive video로부터 action-conditioned video world model을 사전학습하는 파이프라인인 **Olaf-World**를 제시한다. 광범위한 실험은 우리의 방법이 더 구조화된 latent action space를 학습하며, 그 결과 state-of-the-art baseline보다 더 강한 zero-shot action transfer와 더 데이터 효율적인 새로운 control interface 적응을 달성함을 보여준다."},{"id":"64025","en":"Interleaved Selective State Space Models for Efficient WiFi-Based 3D Multi-Person Pose Estimation","ko":"효율적인 WiFi 기반 3D 다중 인물 자세 추정을 위한 교차 삽입 선택적 State Space Models","authors":"Quang-Anh N.D., Kok-Seng Wong","pos":"#3011","link":"https://openreview.net/forum?id=RbeAuiypKb","abs":"

WiFi-based human pose estimation offers privacy-preserving and occlusion-robust sensing, but current Transformer-based approaches suffer from quadratic complexity and lack explicit inductive biases for Channel State Information structure. We propose WiFi-Mamba, the first State Space Model architecture for WiFi-based 3D multi-person pose estimation. Our approach introduces three key contributions: (1) a Dual-Stream Selective State Space Model that processes amplitude and phase through parallel pathways with cross-stream state coupling to respect their distinct physical properties, (2) Selective State Attention for pose query decoding with SSM-derived sequential context, and (3) Persistent SSM Memory for temporal consistency across frames without recurrent memory explosion. Extensive experiments on the Person-in-WiFi 3D dataset, covering both single-person and multi-person scenarios, demonstrate 16-27% MPJPE reduction across varying numbers of persons while using only 4.4% of baseline parameters (2.14M vs. 48.2M), achieving superior efficiency-accuracy trade-offs particularly beneficial for edge deployment in privacy-sensitive continuous monitoring scenarios.

","absKo":"

WiFi 기반 human pose estimation은 privacy-preserving이고 occlusion에 강한 sensing을 제공하지만, 현재 Transformer 기반 접근법은 quadratic complexity를 가지며 Channel State Information 구조에 대한 explicit inductive bias가 부족합니다. 우리는 WiFi 기반 3D multi-person pose estimation을 위한 첫 State Space Model 아키텍처인 WiFi-Mamba를 제안합니다. 우리의 접근은 세 가지 핵심 기여를 도입합니다: (1) amplitude와 phase를 각각의 물리적 특성을 존중하도록 parallel pathway로 처리하고 cross-stream state coupling을 통해 연결하는 Dual-Stream Selective State Space Model, (2) SSM-derived sequential context를 활용한 pose query decoding을 위한 Selective State Attention, (3) recurrent memory explosion 없이 frame 간 temporal consistency를 유지하는 Persistent SSM Memory. Person-in-WiFi 3D dataset에서 single-person과 multi-person 시나리오를 모두 포함한 광범위한 실험은, 다양한 인원 수에 걸쳐 MPJPE를 16-27% 감소시키면서 baseline parameter의 단 4.4%만 사용함(2.14M vs. 48.2M)을 보여 주었고, 특히 privacy에 민감한 continuous monitoring 시나리오의 edge deployment에 유용한 우수한 efficiency-accuracy trade-off를 달성했습니다.

"},{"id":"60537","en":"Precision-Induced Miscalibration: Understanding and Correcting Confidence Distortion in Quantized Neural Networks","ko":"Precision-Induced Miscalibration: Quantized Neural Network의 Confidence Distortion 이해와 보정","authors":"Jiawei Gu, Fengyuan Nie, Hao Tang, Yanpeng Sun","pos":"#3014","link":"https://openreview.net/forum?id=zCkFbxKeF5","abs":"Low-precision arithmetic is pervasive in neural network training and deployment, yet its effect on prediction \\textit{confidence}, not just accuracy, remains unexamined. We show that the softmax function amplifies logit-space quantization errors in an input-dependent manner: confidence distortion scales with the product of precision-dependent error bound $\\epsilon$ and logit norm, peaking when the model is confident but not saturated. This explains why identical models report different confidence values across precisions, a phenomenon we term \\textit{Precision Split}. During training, the same mechanism causes gradient underflow: when logit margins exceed a precision-dependent threshold, gradients vanish and samples silently stop contributing to learning. Since logit norm serves as a computable proxy for precision-induced risk, we propose Precision-Aware Confidence Scaling (PACS), which applies sample-adaptive temperature inversely related to this risk, with sub-one-percent overhead and no full-precision computation required. On ImageNet with mixed-precision ResNet-50, PACS reduces Expected Calibration Error from 5.82\\% to 1.92\\% while maintaining accuracy, with consistent improvements across architectures, precision formats, and modalities.","absKo":"Low-precision arithmetic는 neural network training과 deployment 전반에 널리 사용되지만, accuracy뿐 아니라 prediction \\textit{confidence}에 미치는 영향은 아직 충분히 검토되지 않았다. 우리는 softmax function이 logit-space quantization error를 input-dependent 방식으로 증폭시킨다는 점을 보인다. 즉, confidence distortion은 precision-dependent error bound $\\epsilon$과 logit norm의 곱에 비례해 커지며, model이 confident하지만 아직 saturated되지 않았을 때 가장 크다. 이는 동일한 model이 precision에 따라 서로 다른 confidence 값을 보고하는 현상을 설명하며, 우리는 이를 \\textit{Precision Split}이라 부른다. training 중에는 같은 메커니즘이 gradient underflow를 일으켜, logit margin이 precision-dependent threshold를 넘으면 gradient가 사라지고 sample이 조용히 learning에 기여하지 않게 된다. logit norm이 precision-induced risk의 계산 가능한 proxy로 작동하므로, 우리는 이 risk에 반비례하는 sample-adaptive temperature를 적용하는 Precision-Aware Confidence Scaling (PACS)를 제안한다. PACS는 full-precision computation 없이 동작하며 오버헤드는 1% 미만이다. mixed-precision ResNet-50을 사용한 ImageNet에서 PACS는 accuracy를 유지하면서 Expected Calibration Error를 5.82\\%에서 1.92\\%로 낮추며, architecture, precision format, modality 전반에서 일관된 개선을 보인다."},{"id":"61820","en":"Unbiased Reward Modeling from Implicit Preference","ko":"암묵적 선호로부터의 편향 없는 Reward Modeling","authors":"Eric Wang, Haocheng Yang, Licheng Pan, Lei Shen, Xiaoxi Li, Yinuo Wang, Zhichao Chen, Yuan Lu, Haoxuan Li, Zhouchen Lin","pos":"#305","link":"https://openreview.net/forum?id=membbcuXeR","abs":"

Despite the success of reinforcement learning from human feedback (RLHF) in aligning language models, current reward modeling heavily relies on explicit preference data with high collection costs. In this work, we study implicit reward modeling---learning reward models from implicit human feedback---which offers a cost-effective alternative. We identify two fundamental challenges in implicit reward modeling: (1) Implicit preference data lacks definitive negative samples, which makes standard positive-negative classification methods inapplicable; (2) Implicit preference data suffers from user preference bias, where different responses have different propensities to elicit user feedback actions, which exacerbates the difficulty of distinguishing definitive negative samples. To address these challenges, we propose ImplicitRM, which aims to learn unbiased reward models from implicit preference data. ImplicitRM stratifies training samples into four latent groups via a stratification model. Building on this, it derives a learning objective through likelihood maximization, which we prove is theoretically unbiased, effectively resolving both challenges. Experiments demonstrate that ImplicitRM learns accurate reward models across implicit preference datasets. Code is available at https://anonymous.4open.science/r/ImplicitRM-5FB3.

","absKo":"

language model을 정렬하는 데 있어 reinforcement learning from human feedback(RLHF)가 성공을 거두었음에도 불구하고, 현재의 reward modeling은 높은 수집 비용을 수반하는 명시적 preference data에 크게 의존한다. 본 연구에서는 암묵적 human feedback으로부터 reward model을 학습하는 implicit reward modeling을 연구하며, 이는 비용 효율적인 대안이 된다. 우리는 implicit reward modeling에서 두 가지 근본적 도전을 식별한다: (1) 암묵적 preference data에는 명확한 negative sample이 부족하여 표준 positive-negative classification 방법을 적용할 수 없다; (2) 암묵적 preference data는 user preference bias의 영향을 받는데, 서로 다른 response가 user feedback action을 유발할 propensity가 서로 달라 definitive negative sample을 구분하는 어려움을 가중시킨다. 이러한 도전을 해결하기 위해 우리는 암묵적 preference data로부터 unbiased reward model을 학습하는 것을 목표로 하는 ImplicitRM을 제안한다. ImplicitRM은 stratification model을 통해 training sample을 네 개의 latent group으로 층화한다. 이를 바탕으로 likelihood maximization을 통한 learning objective를 도출하며, 우리는 이것이 이론적으로 unbiased함을 증명하여 두 가지 도전을 효과적으로 해소함을 보인다. 실험은 ImplicitRM이 암묵적 preference dataset 전반에서 정확한 reward model을 학습함을 보여준다. 코드: https://anonymous.4open.science/r/ImplicitRM-5FB3

"},{"id":"65416","en":"From Prior to Pro: Efficient Skill Mastering via Distribution Contractive RL Finetuning","ko":"사전정보에서 전문가로: 분포 수축 RL 파인튜닝을 통한 효율적 스킬 습득","authors":"Zhanyi Sun, shuran song","pos":"#3812","link":"https://openreview.net/forum?id=DsXN7VUwA3","abs":"

We introduce Distribution Contractive Reinforcement Learning (DICE-RL), a framework that uses reinforcement learning (RL) as a “distribution contractor” to refine pretrained generative robot policies. DICE-RL turns a pretrained behavior prior into a high-performing “pro” policy by amplifying high-success behaviors from online feedback. We pretrain a diffusion-based policy for broad behavioral coverage, then finetune it with a stable, sample-efficient residual off-policy RL framework that combines selective behavior regularization with value-guided action selection. Extensive experiments and analyses show that DICE-RL reliably improves performance with strong stability and sample efficiency, enabling mastery of complex long-horizon manipulation skills both in simulation and on a real robot. Project website: dice-rl-anonymous.github.io.

","absKo":"

우리는 Distribution Contractive Reinforcement Learning (DICE-RL)을 소개한다. 이는 reinforcement learning (RL)을 “distribution contractor”로 사용하여 pretrained generative robot policy를 정교화하는 framework이다. DICE-RL은 online feedback에서 성공률이 높은 behavior를 증폭시켜, pretrained behavior prior를 높은 성능의 “pro” policy로 전환한다. 우리는 broad behavioral coverage를 위해 diffusion-based policy를 pretrain한 뒤, selective behavior regularization과 value-guided action selection을 결합한 안정적이고 sample-efficient한 residual off-policy RL framework로 이를 finetune한다. 광범위한 실험과 분석은 DICE-RL이 강한 stability와 sample efficiency로 성능을 안정적으로 향상시키며, simulation과 실제 robot 모두에서 복잡한 long-horizon manipulation skill을 숙달하게 함을 보여준다. Project website: dice-rl-anonymous.github.io.

"},{"id":"60785","en":"Discrete Survival Knowledge Distillation for Competing Risks Analysis","ko":"Competing Risks Analysis를 위한 Discrete Survival Knowledge Distillation","authors":"Feiyang Deng, Lingfeng Luo, Di Wang, Qinmengge Li, Lingxuan Kong, Kevin He","pos":"#3910","link":"https://openreview.net/forum?id=wjsUR2vppX","abs":"

Accurate prediction in survival analysis with competing risks is challenged by rare event rates and limited effective sample sizes. Knowledge distillation offers a promising way to transfer information from an external teacher to improve a local student, but existing methods are overwhelmingly developed for uncensored outcomes and do not directly extend to survival analysis, where censored observations provide only partial information. Moreover, prior work often assumes that teacher and student share identical outcome definitions, whereas in competing risks settings, they may differ in outcome granularity and event definitions, further complicating knowledge transfer. To address these gaps, we propose DiSKD (Discrete Survival Knowledge Distillation), a deep learning framework for discrete-time competing risks that integrates teacher predictions via a cause-specific, time-dependent Kullback--Leibler divergence. DiSKD enables flexible and privacy-preserving transfer without requiring raw data sharing, remains robust to model misspecification or outcome-definition heterogeneity, and adaptively weights teacher guidance by emphasizing compatible teachers while down-weighting less relevant ones. Simulation studies and real-world applications demonstrate improved discrimination and calibration.

","absKo":"

competing risks가 있는 survival analysis에서의 정확한 예측은 희귀한 event rate와 제한된 effective sample size 때문에 어렵다. Knowledge distillation은 외부 teacher로부터 정보를 전달하여 local student를 개선하는 유망한 방법이지만, 기존 방법은 대부분 uncensored outcome을 대상으로 개발되었으며, censored observation이 부분적인 정보만 제공하는 survival analysis로는 직접 확장되지 않는다. 더 나아가, 기존 연구는 teacher와 student가 동일한 outcome definition을 공유한다고 가정하는 경우가 많지만, competing risks 설정에서는 outcome granularity와 event definition이 서로 다를 수 있어 knowledge transfer가 더욱 복잡해진다. 이러한 간극을 해결하기 위해, 우리는 cause-specific하고 time-dependent한 Kullback--Leibler divergence를 통해 teacher prediction을 통합하는 discrete-time competing risks용 deep learning framework인 DiSKD(Discrete Survival Knowledge Distillation)를 제안한다. DiSKD는 raw data 공유 없이도 유연하고 privacy-preserving한 transfer를 가능하게 하며, model misspecification이나 outcome-definition 이질성에도 견고하고, 호환되는 teacher를 강조하고 덜 관련 있는 teacher의 가중치를 낮추는 방식으로 teacher guidance를 적응적으로 가중한다. 시뮬레이션 연구와 실제 응용은 discrimination과 calibration의 향상을 보여준다.

"},{"id":"62694","en":"Signature-Informed Transformer for Asset Allocation","ko":"자산 배분을 위한 Signature-Informed Transformer","authors":"Yoontae Hwang, Stefan Zohren","pos":"#502","link":"https://openreview.net/forum?id=eBM5ALLJNx","abs":"

Modern deep learning for asset allocation typically separates forecasting from optimization. We argue this creates a fundamental mismatch where minimizing prediction errors fails to yield robust portfolios. We propose the Signature Informed Transformer to address this by unifying feature extraction and decision making into a single policy. Our model employs path signatures to encode complex path dependencies and introduces a specialized attention mechanism that targets geometric asset relationships. By directly minimizing the Conditional Value at Risk we ensure the training objective aligns with financial goals. We prove that our attention module rigorously amplifies signature derived signals. Experiments across diverse equity universes show our approach significantly outperforms both traditional strategies and advanced forecasting baselines. The code is available at: https://anonymous.4open.science/r/Signature-Informed-Transformer-For-Asset-Allocation-DB88

","absKo":"

자산 배분을 위한 현대적 deep learning은 대개 forecasting과 optimization을 분리한다. 우리는 이러한 분리가 근본적인 불일치를 만들어, prediction error를 최소화하는 것이 robust portfolio로 이어지지 않는다고 주장한다. 이를 해결하기 위해 feature extraction과 decision making을 하나의 policy로 통합하는 Signature Informed Transformer를 제안한다. 우리의 모델은 path signature를 사용해 복잡한 path dependency를 인코딩하고, geometric한 asset 관계를 대상으로 하는 specialized attention mechanism을 도입한다. Conditional Value at Risk를 직접 최소화함으로써 training objective가 금융 목표와 정렬되도록 한다. 우리는 attention module이 signature에서 도출된 signal을 엄밀하게 증폭함을 증명한다. 다양한 equity universe에 대한 실험은 우리의 접근이 전통적 전략과 고급 forecasting baseline 모두를 크게 능가함을 보여준다. 코드는 다음에서 제공된다: https://anonymous.4open.science/r/Signature-Informed-Transformer-For-Asset-Allocation-DB88

"},{"id":"65297","en":"MarketSim: Simulating Stock Markets with Large-Scale Generative Agents","ko":"MarketSim: 대규모 생성 에이전트로 주식 시장 시뮬레이션하기","authors":"Jinghua Piao, zhentao liu, Cheng Huang, Huang Jiarui, Songwei Li, Wang Ranran, Yong Li","pos":"#604","link":"https://openreview.net/forum?id=EzpJxPDqXB","abs":"

Stock markets are one of the most complex systems in the modern world, where prices emerge from billions of decentralized interactions among heterogeneous participants in an ever-evolving information landscape. While high-fidelity simulation is important for understanding market dynamics, existing approaches face a persistent trade-off between structural and behavioral fidelity. To this end, we propose MarketSim, a large-scale stock market simulation framework with generative agents. MarketSim introduces a hierarchical multi-agent architecture that decouples strategic reasoning from high-frequency execution, enabling LLM agents to operate in a nanosecond-resolution, NASDAQ-like continuous double auction market. Building on this, we simulate over 15,000 heterogeneous market participants whose interactions shape and are shaped by an evolving market environment grounded in more than 12k real-world news articles, policy documents, and earnings reports. To evaluate our proposed MarketSim, we develop a comprehensive benchmark that includes stocks from 8 GICS sectors and 3 representative real-world scenarios, along with 5 stylized facts for market complexity and 5 price-related statistical metrics. Extensive experiments demonstrate that MarketSim not only captures key complexity properties of real-world markets, but also outperforms state-of-the-art baselines in tracking high-frequency price dynamics with an average MAPE of 3.48%, providing a scalable testbed for market analysis.

","absKo":"

주식 시장은 현대 세계에서 가장 복잡한 시스템 중 하나로, 가격은 끊임없이 진화하는 정보 환경 속에서 이질적인 참여자들 간 수십억 개의 분산 상호작용에서 emerge한다. 시장 동역학을 이해하기 위해서는 high-fidelity simulation이 중요하지만, 기존 접근법은 structural fidelity와 behavioral fidelity 사이의 지속적인 trade-off에 직면한다. 이를 위해 우리는 generative agent를 갖춘 대규모 주식 시장 시뮬레이션 프레임워크인 MarketSim을 제안한다. MarketSim은 전략적 추론과 고주파 execution을 분리하는 hierarchical multi-agent architecture를 도입하여, LLM agent가 nanosecond 해상도의 NASDAQ 유사 continuous double auction market에서 동작할 수 있게 한다. 이를 바탕으로, 우리는 12k개 이상의 실제 뉴스 기사, 정책 문서, 실적 보고서를 토대로 구성된 진화하는 시장 환경의 영향을 주고받는 15,000명 이상의 이질적 시장 참여자를 시뮬레이션한다. 제안한 MarketSim을 평가하기 위해, 우리는 8개의 GICS sector에 속한 주식과 3개의 대표적인 실제 시나리오를 포함하고, 시장 복잡성을 위한 5개의 stylized fact와 가격 관련 5개의 통계 metric을 포함하는 종합 benchmark를 개발한다. 광범위한 실험은 MarketSim이 실제 시장의 핵심 복잡성 특성을 포착할 뿐 아니라, 평균 MAPE 3.48%로 고주파 가격 동역학 추적에서 state-of-the-art baseline을 능가하며, 시장 분석을 위한 확장 가능한 testbed를 제공함을 보여준다.

"},{"id":"63647","en":"GeoReward: Mitigating Contextual Variable Overestimation in Vision-Language Models for Cross-Market Preference Prediction","ko":"GeoReward: 시장 간 선호도 예측을 위한 vision-language model의 문맥 변수 과대추정 완화","authors":"Shuo Liu, Huixiang.Cai, Weiru Zhang, Xiaoyi Zeng","pos":"#1000","link":"https://openreview.net/forum?id=V2yikPhdXE","abs":"

Vision-language models (VLMs) excel in many multimodal tasks but remain prone to a subtle yet impactful failure mode: they tend to overestimate dominant visual-textual cues while underestimating sparse but decision-critical contextual variables. This issue, which we term Contextual Variable Overestimation (CVE), becomes particularly evident in real-world applications such as predicting advertisement image preferences across diverse geographic markets. For instance, when a VLM (e.g., Qwen2-VL) is asked to choose between two product images tailored for different countries (e.g., Korea vs. France), it often defaults to a consistent output (e.g., always selects “A”), ignoring ground-truth regional variations. This collapse occurs because pervasive high-volume signals, such as product attributes and dense image patches, overwhelm the few but critical tokens that encode market-specific context (e.g., country names). To address CVE, we first collect a new multimodal dataset of real advertising creatives and their click-through performance across multiple countries. We then introduce GeoReward, a reward model designed to predict ad image preferences across diverse geographic markets. GeoReward integrates three purpose-built mechanisms: (1) Market-Aware Retrieval Augmentation, which retrieves and injects region-aligned preference signals during training to sharpen localization awareness. (2) Context-Guided Visual Modulation, a lightweight adapter that dynamically adjusts visual representations using textual country embeddings, enabling fine-grained regional adaptation. (3) Selective Sensitivity Loss, an objective that applies heightened penalties for context-specific mispredictions, sharpening the model's focus on critical variables. Furthermore, we demonstrate how GeoReward can guide the fine-tuning of RL for a VLM to generate background designs for text-to-image models (e.g., SDXL), producing market-aware advertising creatives. Experiments validate that our framework mitigates CVE and outperforms existing baselines. This work not only diagnoses a systematic bias in VLMs toward dominant perceptual features but also delivers a targeted solution for applications where sparse contextual variables govern decision-making.

","absKo":"

Vision-language models (VLMs)는 많은 multimodal task에서 뛰어난 성능을 보이지만, 미묘하지만 영향력이 큰 failure mode에 여전히 취약하다. 즉, 지배적인 visual-textual cue를 과대평가하는 반면, 희소하지만 의사결정에 결정적인 contextual variable은 과소평가하는 경향이 있다. 우리는 이 문제를 Contextual Variable Overestimation (CVE)라고 부르며, 다양한 geographic market에서 advertisement image preference를 예측하는 것과 같은 실제 응용에서 특히 두드러진다. 예를 들어, VLM(e.g., Qwen2-VL)에게 Korea와 France처럼 서로 다른 국가에 맞춘 두 product image 중 하나를 고르게 하면, ground-truth regional variation을 무시한 채 종종 일관된 출력(e.g., 항상 “A” 선택)으로 수렴한다. 이러한 collapse는 product attribute와 dense image patch처럼 광범위한 high-volume signal이 country name처럼 market-specific context를 인코딩하는 소수이지만 중요한 token을 압도하기 때문에 발생한다. CVE를 해결하기 위해, 우리는 먼저 여러 국가에 걸친 실제 advertising creative와 그 click-through performance를 담은 새로운 multimodal dataset을 수집한다. 그다음, 다양한 geographic market에서 ad image preference를 예측하도록 설계된 reward model인 GeoReward를 소개한다. GeoReward는 세 가지 특화 메커니즘을 통합한다. (1) Market-Aware Retrieval Augmentation은 training 중 region-aligned preference signal을 retrieval하여 주입함으로써 localization awareness를 sharpen한다. (2) Context-Guided Visual Modulation은 textual country embedding을 사용해 visual representation을 동적으로 조정하는 lightweight adapter로, 세밀한 regional adaptation을 가능하게 한다. (3) Selective Sensitivity Loss는 context-specific misprediction에 대해 더 강한 penalty를 적용하는 objective로, 중요한 variable에 대한 모델의 집중을 더욱 sharpen한다. 더 나아가, GeoReward가 VLM의 RL fine-tuning을 지도하여 text-to-image model(e.g., SDXL)을 위한 background design을 생성하게 함으로써, market-aware advertising creative를 만들 수 있음을 보인다. 실험 결과는 우리의 framework가 CVE를 완화하고 기존 baseline을 능가함을 검증한다. 이 연구는 지배적인 perceptual feature를 향한 VLM의 체계적 bias를 진단할 뿐 아니라, 희소한 contextual variable이 의사결정을 좌우하는 응용을 위한 정밀한 해결책도 제시한다.

"},{"id":"60833","en":"Global Merger-Arbitrage Forecasting with Language Models","ko":"Language Model을 이용한 글로벌 Merger-Arbitrage 예측","authors":"Hinal Jajal, Michał Mucha, Charles Sweat, Chris Pulman, Charlie Flanagan, Peter Anderson","pos":"#1002","link":"https://openreview.net/forum?id=wDaX9bFufS","abs":"

Prior work on judgmental forecasting with large language models (LLMs) has focused on broad, mixed‑topic question banks and shallow context (e.g., short news snippets). We study a specialized, high‑stakes financial setting: forecasting M&A outcomes for merger arbitrage. Using rich textual evidence, with context engineering informed by veteran merger-arb specialists, we construct an LLM‑based forecasting system and finetune the model using outcome-conditioned gold reasoning traces. The system outputs probabilistic forecasts over closing at announced terms, higher bid, and deal termination. On an out-of-sample set of more than 400 large deals spanning 42 countries, our finetuned system outperforms a variety of frontier models and market-based baselines, using a Brier score weighted by the P&L impact of each deal.

","absKo":"

대규모 언어 모델(LLMs)을 활용한 judgmental forecasting에 관한 기존 연구는 폭넓고 mixed-topic인 question bank와 얕은 context(예: 짧은 뉴스 스니펫)에 초점을 맞춰 왔다. 우리는 보다 특화되고 고위험인 금융 환경, 즉 merger arbitrage를 위한 M&A outcome forecasting을 연구한다. 숙련된 merger-arb 전문가들의 경험을 반영한 context engineering과 풍부한 텍스트 증거를 활용하여, 우리는 LLM 기반 forecasting system을 구축하고 outcome-conditioned gold reasoning trace로 모델을 finetune한다. 이 시스템은 announced terms로의 closing, higher bid, deal termination에 대한 probabilistic forecast를 출력한다. 42개 국가에 걸친 400건 이상의 대형 거래로 이루어진 out-of-sample set에서, 우리 finetuned system은 각 거래의 P&L impact로 가중한 Brier score를 사용했을 때, 다양한 frontier model과 market-based baseline을 상회하는 성능을 보인다.

"},{"id":"62547","en":"Landmark-Guided Policy Optimization for Multi-Objective Language Model Selection","ko":"Landmark-Guided Policy Optimization으로 다중 목적 Language Model Selection 수행","authors":"Marcio Monteiro, Weichen Li, Puyu Wang, Marius Kloft, Sophie Fellenz","pos":"#1005","link":"https://openreview.net/forum?id=fexmo79Kxo","abs":"

Selecting a pretrained large language model (LLM) to fine-tune for a task-specific dataset can be time-consuming and costly. With several candidate models available to choose from, varying in size, architecture, and pretraining data, finding the best model for a specific task often involves extensive trial and error. In addition, the \"best\" model may not necessarily be the one with the lowest test loss, as practical considerations such as deployment costs, inference throughput, and limited search budgets might also play crucial roles. To address this, we introduce LAMPS (LAnguage Model Pareto Selection), a novel and open-source multi-objective AutoML framework that meta-learns a resource allocation policy to efficiently identify (or approximate) the Pareto front of candidate LLMs for a task-specific dataset. It is based on two key ideas: (1) landmark fine-tuning, which generates early performance indicators of the candidate models, and (2) meta-learning via reinforcement learning, which learns an effective selection policy from historical performance data (a meta-dataset). Our results show that, on held-out datasets, LAMPS reduces search time by an average of 73\\% compared to exhaustive search, while still covering more than 99\\% of the optimal target space hypervolume.

","absKo":"

특정 task-specific dataset에 대해 fine-tuning할 pretrained large language model(LLM)을 선택하는 일은 시간도 많이 들고 비용도 많이 든다. 크기, architecture, pretraining data가 서로 다른 여러 candidate model 중에서 최적의 모델을 찾으려면, 보통 광범위한 시행착오가 필요하다. 또한 \"최고\"의 모델이 반드시 test loss가 가장 낮은 모델인 것은 아니며, 배포 비용, inference throughput, 제한된 search budget 같은 현실적인 고려사항도 중요한 역할을 할 수 있다. 이를 해결하기 위해 우리는 LAMPS(LAnguage Model Pareto Selection)를 제안한다. 이는 task-specific dataset에 대해 candidate LLM들의 Pareto front를 효율적으로 식별(또는 근사)하기 위해 resource allocation policy를 meta-learn하는 새롭고 open-source인 multi-objective AutoML framework이다. 이 framework는 두 가지 핵심 아이디어에 기반한다: (1) candidate model의 초기 성능 지표를 생성하는 landmark fine-tuning, (2) historical performance data(meta-dataset)로부터 효과적인 selection policy를 학습하는 reinforcement learning 기반 meta-learning. 우리의 결과는 held-out dataset에서 LAMPS가 exhaustive search와 비교해 평균 73\\%의 search time을 줄이면서도, 최적 target space hypervolume의 99\\% 이상을 여전히 커버함을 보여준다.

"},{"id":"60834","en":"Drop-in Circulant Structural Priors for Transformer Decoding of Cyclic Codes","ko":"Cyclic Code의 Transformer Decoding을 위한 Drop-in Circulant Structural Prior","authors":"Shuai Xiao, Weijun Fang, Qiaosheng Zhang","pos":"#1101","link":"https://openreview.net/forum?id=wDFrxk7306","abs":"

While Transformer-based architectures have revolutionized neural decoding, existing models often treat codes as generic sequences, ignoring their inherent algebraic properties. In this paper, we take a step toward bridging these two domains by proposing a novel decoding approach that integrates the algebraic structure of cyclic codes into Transformer-based decoders. Leveraging the inherent cyclic properties, we introduce interpretable error correction patterns and inter-node relationship hypotheses that link the structural characteristics of the codes to the model parameters. Building on these insights, we design a plug-and-play, flexibly deployable decoding method tailored for cyclic codes. Experimental results show that our method achieves an average reduction in bit error rate (BER) by an order of magnitude, while also reducing the total number of parameters by approximately 97%. Additional comparative experiments validate our proposed conjectures and highlight a promising pathway for bridging classical coding theory and modern Transformer-based decoding architectures.

","absKo":"

Transformer 기반 architecture가 neural decoding을 혁신했지만, 기존 model은 code를 일반적인 sequence로 취급하여 그 고유한 algebraic property를 무시하는 경우가 많다. 본 논문에서는 cyclic code의 algebraic structure를 Transformer-based decoder에 통합하는 새로운 decoding approach를 제안함으로써 이 두 영역을 연결하는 한 걸음을 내딛는다. 내재된 cyclic property를 활용하여, 우리는 code의 structural characteristic과 model parameter를 연결하는 해석 가능한 error correction pattern과 inter-node relationship hypothesis를 도입한다. 이러한 통찰을 바탕으로, cyclic code에 맞춘 plug-and-play 방식의 유연하게 배포 가능한 decoding method를 설계한다. 실험 결과, 제안 방법은 bit error rate(BER)를 평균적으로 한 자릿수 차수(order) 수준으로 감소시키는 동시에 total number of parameters를 약 97% 줄인다. 추가적인 비교 실험은 우리가 제안한 conjecture를 검증하고 classical coding theory와 modern Transformer-based decoding architecture를 연결하는 유망한 경로를 보여 준다.

"},{"id":"64895","en":"Factorized Scheduling Principle: Learning Interpretable and Transferable Policies via Structured Additive Functions","ko":"Factorized Scheduling 원칙: 구조화된 Additive Function을 통한 해석 가능하고 이전 가능한 Policy 학습","authors":"Hong Je-Gal, Hyun-Suk Lee","pos":"#119","link":"https://openreview.net/forum?id=ImK3lNmGTW","abs":"

Scheduling problems arise from repeatedly selecting one item from a set of candidates based on their states. These problems often reduce to assigning priority scores and choosing the highest-ranked item. In this work, we propose a factorized scheduling principle (FSP) framework to learn interpretable and transferable scheduling rules. The FSP framework represents system states as condition distributions and decomposes a global scheduling principle into additive univariate and pairwise components with identifiability constraints. The scheduling principle enables the framework to maintain a simple priority-based structure during deployment. This principle is learned by using a policy-based objective combined with a temporal-difference signal defined on the condition distribution. Experiments on synthetic and realistic scheduling tasks demonstrate the FSP framework's strong performance, interpretability, and zero-shot generalization across different system scales.

","absKo":"

Scheduling problem은 상태에 기반해 후보 집합에서 매번 하나의 item을 반복적으로 선택하는 문제에서 발생한다. 이러한 문제는 보통 priority score를 할당하고 가장 높은 순위의 item을 고르는 것으로 귀결된다. 본 연구에서는 해석 가능하고 transfer 가능한 scheduling rule을 학습하기 위한 factorized scheduling principle (FSP) framework를 제안한다. FSP framework는 system state를 condition distribution으로 표현하고, global scheduling principle을 identifiable constraint가 있는 additive univariate 및 pairwise component로 분해한다. 이 scheduling principle은 deployment 동안에도 단순한 priority-based structure를 유지하게 한다. 이 원리는 condition distribution 위에 정의된 temporal-difference signal과 policy-based objective를 결합해 학습된다. synthetic 및 realistic scheduling task에 대한 실험은 FSP framework가 다양한 system scale 전반에서 강력한 성능, 해석 가능성, 그리고 zero-shot generalization을 보임을 입증한다.

"},{"id":"61526","en":"CoEvol-NO: State and Coordinate Co-Evolution with an Error-Driven Predictor-Corrector Paradigm for Neural Operator Transformer","ko":"CoEvol-NO: Neural Operator Transformer를 위한 Error-Driven Predictor-Corrector 패러다임 기반 State와 Coordinate Co-Evolution","authors":"Jianqiao Zeng, Ruocheng Wang, Yanzhi Liu, Hao Xiong, Junchi Yan","pos":"#1600","link":"https://openreview.net/forum?id=pG9Wn3Aheq","abs":"

Despite the fast progress in neural operator learning, long-sequence modeling still is a standing challenge whereby latent states have been introduced with techniques well derived. Diverging from existing methods that treat latent states as transient variables or decoupled representations, CoEvol-NO introduces a {persistent state} to establish a {co-evolutionary framework}, where the latent state and mesh sequence are updated jointly and bidirectionally. Inspired by classical numerical methods, we model the layer-wise state evolution as a {Predictor-Corrector (PC)} process. Specifically, a Predictor'' generates a tentative target, followed by aCorrector'' that refines the persistent state via an {error-driven update mechanism}. Furthermore, our theoretical analysis reveals that the widely used \\textit{direct substitution} and \\textit{residual update} paradigms are essentially {first-order approximations} of this error-driven correction under different loss assumptions. We theoretically prove that CoEvol-NO achieves strict {linear time complexity}. Extensive experiments on five standard benchmarks and two large-scale industrial design tasks demonstrate that CoEvol-NO consistently achieves {state-of-the-art (SOTA)} performance.

","absKo":"신경 operator learning의 빠른 발전에도 불구하고, long-sequence modeling은 여전히 지속적인 도전 과제로 남아 있으며, 이를 위해 latent state가 잘 정립된 여러 기법과 함께 도입되어 왔다. latent state를 일시적인 변수나 분리된 representation으로 다루는 기존 방법들과 달리, CoEvol-NO는 {persistent state}를 도입하여 {co-evolutionary framework}를 확립하며, 여기서 latent state와 mesh sequence는 양방향으로 공동 갱신된다. 고전적인 numerical method에서 영감을 받아, 우리는 layer-wise state evolution을 {Predictor-Corrector (PC)} process로 모델링한다. 구체적으로 Predictor''가 잠정적인 target을 생성하고, 이어서Corrector''가 {error-driven update mechanism}을 통해 persistent state를 정교화한다. 또한 우리의 이론 분석은 널리 사용되는 \\textit{direct substitution}과 \\textit{residual update} 패러다임이 서로 다른 loss assumption 하에서 본질적으로 이러한 error-driven correction의 {first-order approximation}임을 보여준다. 우리는 CoEvol-NO가 엄밀한 {linear time complexity}를 달성함을 이론적으로 증명한다. 다섯 개의 표준 benchmark와 두 개의 대규모 산업 설계 과제에서 수행한 광범위한 실험은 CoEvol-NO가 일관되게 {state-of-the-art (SOTA)} 성능을 달성함을 보여준다.

"},{"id":"64385","en":"InfoPO: Information-Driven Policy Optimization for User-Centric Agents","ko":"InfoPO: 사용자 중심 에이전트를 위한 정보 기반 정책 최적화","authors":"Fanqi Kong, Jiayi Zhang, Mingyi Deng, Chenglin Wu, Yuyu Luo, Bang Liu","pos":"#2201","link":"https://openreview.net/forum?id=O7UHBoLrx4","abs":"

Real-world user requests to LLM agents are often underspecified. Agents must interact to acquire missing information and make correct downstream decisions. However, current multi-turn GRPO-based methods often rely on trajectory-level reward computation, which leads to credit assignment problems and insufficient advantage signals within rollout groups. A feasible approach is to identify valuable interaction turns at a fine granularity to drive more targeted learning. To address this, we introduce InfoPO, which frames multi-turn interaction as a process of active uncertainty reduction and computes an information-gain reward that credits turns whose feedback measurably changes the agent’s subsequent action distribution compared to a masked-feedback counterfactual. It then combines this signal with task outcomes via an adaptive variance-gated fusion to identify information importance while maintaining task oriented goal direction. Across diverse tasks including intent clarification, collaborative coding, and tool-augmented decision making, InfoPO consistently outperforms prompting and multi-turn RL baselines. It also demonstrates robustness under user simulator shifts and generalizes effectively to environment interactive tasks. Overall, InfoPO provides a principled and scalable mechanism for optimizing complex agent user collaboration.

","absKo":"

실제 사용자 요청은 LLM agent에 대해 종종 underspecified 되어 있다. agent는 누락된 정보를 획득하고 이후 단계의 올바른 결정을 내리기 위해 상호작용해야 한다. 그러나 현재의 multi-turn GRPO 기반 방법은 trajectory-level reward computation에 의존하는 경우가 많아, rollout group 내에서 credit assignment 문제와 부족한 advantage signal을 초래한다. 보다 세밀한 granularity에서 가치 있는 interaction turn을 식별하여 더 목표 지향적인 학습을 유도하는 것이 실현 가능한 접근법이다. 이를 위해 우리는 InfoPO를 제안한다. InfoPO는 multi-turn interaction을 active uncertainty reduction 과정으로 정식화하고, masked-feedback counterfactual과 비교했을 때 feedback이 agent의 이후 action distribution을 유의미하게 바꾸는 turn에 보상하는 information-gain reward를 계산한다. 이후 이 신호를 task outcome과 adaptive variance-gated fusion으로 결합하여, task-oriented goal direction을 유지하면서 information importance를 식별한다. intent clarification, collaborative coding, tool-augmented decision making을 포함한 다양한 task 전반에서 InfoPO는 prompting 및 multi-turn RL baseline을 일관되게 능가한다. 또한 user simulator shift 하에서도 강건함을 보이며, environment interactive task로 효과적으로 일반화된다. 전체적으로 InfoPO는 복잡한 agent-user collaboration을 최적화하기 위한 원리적이고 확장 가능한 메커니즘을 제공한다.

"},{"id":"60960","en":"Bridging Functional Correctness and Runtime Efficiency Gaps in LLM-Based Code Translation","ko":"LLM 기반 Code Translation에서 Functional Correctness와 Runtime Efficiency 격차 연결","authors":"Longhui Zhang, Jiahao Wang, Chenhao Hu, Bingyu Liang, Jing Li, Min zhang","pos":"#2612","link":"https://openreview.net/forum?id=v0vCc5nFRT","abs":"

While large language models (LLMs) have greatly advanced the functional correctness of automated code translation systems, the runtime efficiency of translated programs has received comparatively little attention. With the waning of Moore’s law, runtime efficiency has become increasingly important for program quality, alongside functional correctness. Our preliminary study reveals that LLM-translated programs often run slower than human-written ones, and this issue cannot be remedied through prompt engineering alone. Therefore, our work proposes SwiftTrans, a code translation framework comprising two key stages: (1) Multi-Perspective Exploration, where MpTranslator leverages parallel in-context learning (ICL) to generate diverse translation candidates; and (2) Difference-Aware Selection, where DiffSelector identifies the optimal candidate by explicitly comparing differences between translations. We further introduce Hierarchical Guidance for MpTranslator and Ordinal Guidance for DiffSelector, enabling LLMs to better adapt to these two core components. To support the evaluation of runtime efficiency in translated programs, we extend existing benchmarks, CodeNet and F2SBench, and introduce a new benchmark, SwiftBench. Experimental results across all three benchmarks show that SwiftTrans achieves consistent improvements in both correctness and runtime efficiency.

","absKo":"

large language model (LLM)은 자동 코드 번역 시스템의 functional correctness를 크게 향상시켰지만, 번역된 프로그램의 runtime efficiency는 비교적 거의 주목받지 못했다. Moore's law의 둔화와 함께 runtime efficiency는 functional correctness와 더불어 program quality에서 점점 더 중요해지고 있다. 우리의 예비 연구는 LLM이 번역한 프로그램이 사람이 작성한 프로그램보다 종종 더 느리게 실행되며, 이 문제는 prompt engineering만으로는 해결되지 않음을 보여준다. 따라서 우리의 작업은 SwiftTrans를 제안하는데, 이는 두 가지 핵심 단계로 구성된 code translation framework이다: (1) Multi-Perspective Exploration에서는 MpTranslator가 parallel in-context learning (ICL)을 활용해 다양한 번역 후보를 생성하고, (2) Difference-Aware Selection에서는 DiffSelector가 번역들 간의 차이를 명시적으로 비교하여 최적 후보를 식별한다. 우리는 또한 MpTranslator를 위한 Hierarchical Guidance와 DiffSelector를 위한 Ordinal Guidance를 도입하여 LLM이 이 두 핵심 구성요소에 더 잘 적응하도록 한다. 번역된 프로그램의 runtime efficiency 평가를 지원하기 위해, 기존 benchmark인 CodeNet과 F2SBench를 확장하고 새로운 benchmark인 SwiftBench를 도입한다. 세 benchmark 전반의 실험 결과, SwiftTrans가 correctness와 runtime efficiency 모두에서 일관된 향상을 달성함을 보인다.

"},{"id":"62804","en":"SemRep: Code Transformation with Semantics-Preserving Representations","ko":"SemRep: 의미 보존 Representation을 활용한 코드 변환","authors":"Weichen Li, Jiamin Song, Bogdan Stoica, Arav Dhoot, Gabriel Ryan, Shengyu Fu, Kexin Pei","pos":"#4014","link":"https://openreview.net/forum?id=d81PBCmA4s","abs":"

Code transformation is a foundational capability in the software development process, where its effectiveness relies on constructing a high-quality code representation to characterize the input code semantics and guide the transformation. Existing approaches treat code transformation as an end-to-end learning task, leaving the construction of the representation needed for semantic reasoning implicit in model weights or relying on expensive compiler-level abstractions. We present SemRep, a framework that improves code transformation through generative code representation learning. Our key insight is to employ the semantics-preserving transformations as the intermediate representation, which will be used to train the model as a generative task, and also guide the subsequent instruction-specific code transformations. Across general code editing and CUDA kernel optimization, SemRep outperforms the strong closed-weight baselines by 6.9% and 43% in correctness, 13.9% in generalization and 6.7% in robustness. Combined with an evolutionary coding agent, SemRep finds optimizations that 685B larger-weight baselines fail to discover while achieving the same performance with 25% less inference compute.

","absKo":"

Code transformation은 software development process의 근간이 되는 능력으로, 그 효과는 input code semantics를 특징짓고 transformation을 안내하는 고품질 code representation을 구성하는 데 달려 있습니다. 기존 접근법은 code transformation을 end-to-end learning task로 다루며, semantic reasoning에 필요한 representation 구성을 model weight 안에 암묵적으로 두거나, 비용이 큰 compiler-level abstraction에 의존합니다. 우리는 generative code representation learning을 통해 code transformation을 개선하는 framework인 SemRep을 제시합니다. 우리의 핵심 통찰은 semantics-preserving transformation을 intermediate representation으로 사용하여, 이를 모델을 generative task로 학습하는 데 활용하고 이후 instruction-specific code transformation도 안내하는 것입니다. 일반적인 code editing과 CUDA kernel optimization 전반에서 SemRep은 강력한 closed-weight baseline보다 correctness에서 6.9%와 43%, generalization에서 13.9%, robustness에서 6.7% 향상됩니다. evolutionary coding agent와 결합했을 때, SemRep은 685B 더 큰 weight를 가진 baseline이 발견하지 못한 optimization을 찾으면서도, inference compute를 25% 덜 사용해 동일한 performance를 달성합니다.

"},{"id":"64007","en":"SoftBinary Coding: A New Information-Theoretic Paradigm for Neural Compression via Fast Channel Simulation","ko":"SoftBinary Coding: 빠른 채널 시뮬레이션을 통한 신경 압축을 위한 새로운 정보이론적 패러다임","authors":"Ezgi Ozyilkan, Sharang Sriramu, Elza Erkip, Aaron Wagner, Jona Ballé","pos":"#4015","link":"https://openreview.net/forum?id=RmlisaChid","abs":"

Neural compression is currently dominated by Nonlinear Transform Coding (NTC), which maps data to real-valued latents via continuous transforms. Despite its success, NTC suffers from train-test mismatch due to non-differentiable quantization, a ''smoothness bias'' inherent in continuous transforms that precludes optimality for certain sources, and a loss of ''shaping gain\" due to the complexity of including high-dimensional vector quantization. We propose SoftBinary Coding (SBC), an end-to-end learning paradigm that bypasses these limitations by using a stochastic binary latent space. In the spirit of vector quantization, SBC employs discrete representations and compresses them through a novel fast binary channel simulation scheme, for which we provide a proof of rate optimality. Experimental gains on information-theoretic sources provide both theoretical and practical closure to NTC's limitations, establishing discrete binary structures as a viable path toward reaching optimal rate--distortion bounds. Surprisingly, SBC also achieves state-of-the-art performance on vector quantization of i.i.d. sources, exceeding Trellis Coded Quantization of the Gaussian source.

","absKo":"

Neural compression은 현재 Nonlinear Transform Coding(NTC)이 지배하고 있으며, 이는 연속 변환을 통해 데이터를 real-valued latent로 매핑한다. 성공에도 불구하고 NTC는 미분 불가능한 quantization으로 인한 train-test mismatch, 특정 source에 대해 최적성을 가로막는 continuous transform 고유의 \"smoothness bias\", 그리고 고차원 vector quantization을 포함시키는 복잡성 때문에 발생하는 \"shaping gain\" 손실을 겪는다. 우리는 stochastic binary latent space를 사용하여 이러한 한계를 우회하는 end-to-end learning paradigm인 SoftBinary Coding(SBC)을 제안한다. vector quantization의 정신에 따라 SBC는 discrete representation을 사용하고, 새로운 fast binary channel simulation scheme을 통해 이를 압축하며, 이에 대해 rate optimality의 증명을 제시한다. information-theoretic source에 대한 실험적 향상은 NTC의 한계에 대한 이론적·실용적 마무리를 모두 제공하며, discrete binary structure가 optimal rate--distortion bound에 도달하는 실행 가능한 경로임을 확립한다. 놀랍게도 SBC는 i.i.d. source의 vector quantization에서도 state-of-the-art 성능을 달성하며, Gaussian source의 Trellis Coded Quantization을 능가한다.

"},{"id":"66062","en":"Agent JIT Compilation for Latency-Optimizing Computer-Use Agent Planning and Scheduling","ko":"지연 최적화 컴퓨터 사용 에이전트 계획 및 스케줄링을 위한 에이전트 JIT 컴파일","authors":"Caleb Winston, Ron Yifeng Wang, Azalia Mirhoseini, Christoforos Kozyrakis","pos":"#4100","link":"https://openreview.net/forum?id=7NiATKcx1B","abs":"Computer-use agents (CUA) automate tasks specified with natural language such as \"order the cheapest item from Taco Bell\", by generating sequences of calls to tools such as click, type, and scroll on a browser. Current CUA implementations follow a sequential fetch-screenshot-execute loop where each iteration requires an LLM call, resulting in high latency and frequent errors from incorrect tool use. We present agent JIT compilation, an alternative that compiles task descriptions directly into executable code that is free to include LLM calls, tool calls, and parallelization. Our approach comprises three components: (1) JIT-Planner, which generates multiple code plans from a task, validates each against tool specifications, and selects the minimum-cost candidate; (2) JIT-Scheduler, which explores parallelization strategies via Monte Carlo cost estimation from learned latency distributions; and (3) an invariant-enforcing tool protocol specifying pre/postcondition state requirements that reduces the rate of generating plans with incorrect tool use. Evaluation across 5 applications demonstrates that JIT-Planner achieves $10.4\\times$ speedup and $+28\\%$ accuracy improvement over Browser-Use, while JIT-Scheduler achieves $2.6\\times$ speedup and $+9\\%$ accuracy improvement over OpenAI CUA.","absKo":"Computer-use agent(CUA)는 \"order the cheapest item from Taco Bell\"과 같이 자연어로 지정된 task를 자동화하며, browser에서 click, type, scroll과 같은 tool 호출 시퀀스를 생성한다. 현재의 CUA 구현은 각 iteration마다 LLM 호출이 필요한 순차적 fetch-screenshot-execute loop를 따르므로, 지연 시간이 높고 잘못된 tool use로 인한 오류가 빈번하다. 우리는 agent JIT compilation을 제안한다. 이는 task description을 LLM 호출, tool 호출, parallelization을 포함할 수 있는 실행 가능한 code로 직접 컴파일하는 대안이다. 우리의 접근법은 세 가지 구성요소로 이루어진다: (1) JIT-Planner, task로부터 여러 code plan을 생성하고, 각 plan을 tool specification과 대조해 검증한 뒤, 최소 비용 후보를 선택한다; (2) JIT-Scheduler, 학습된 latency distribution으로부터 Monte Carlo cost estimation을 통해 parallelization 전략을 탐색한다; (3) 잘못된 tool use를 포함한 plan 생성률을 낮추는, pre/postcondition state requirement를 명시한 invariant-enforcing tool protocol. 5개 application에 대한 평가에서 JIT-Planner는 Browser-Use 대비 10.4\\times speedup과 +28\\% accuracy 향상을 달성하고, JIT-Scheduler는 OpenAI CUA 대비 2.6\\times speedup과 +9\\% accuracy 향상을 달성한다."},{"id":"64508","en":"OmniSapiens: A Foundation Model for Social Behavior Processing via Heterogeneity-Aware Relative Policy Optimization","ko":"OmniSapiens: 이질성 인식 상대 정책 최적화를 통한 사회적 행동 처리를 위한 파운데이션 모델","authors":"Keane Ong, Sabri Boughorbel, Luwei Xiao, Chanakya Ekbote, David Dai, Ao Qu, Jingyao Wu, Rui Mao, Ehsan Hoque, Erik Cambria, Gianmarco Mengaldo, Paul Pu Liang","pos":"#605","link":"https://openreview.net/forum?id=MrupCemBMy","abs":"

To develop socially intelligent AI, existing approaches typically model behavioral dimensions (e.g., affective, cognitive, or social attributes) in isolation. Although useful, this task-specific modeling increases training costs and limits generalization across behavioral settings. Recent reasoning RL methods facilitate training a single unified model across multiple behavioral tasks, but do not explicitly address learning across different heterogeneous behavioral data. To address this gap, we introduce Heterogeneity-Aware Relative Policy Optimization (HARPO), a RL method that balances leaning across heterogeneous tasks and samples. This is achieved by modulating advantages to ensure that no single task or sample carries disproportionate influence during policy optimization. Using HARPO, we develop and release OMNISAPIENS-7B 2.0, a foundation model for social behavior processing. Relative to existing behavioral foundation models, OMNISAPIENS-7B 2.0 achieves the strongest performance across behavioral tasks, with gains of up to +16.85% and +9.37% on multitask and held-out settings respectively, while producing more explicit and robust reasoning traces. We also validate HARPO against recent RL methods, where it achieves the most consistently strong performance across behavioral tasks.

","absKo":"

사회적으로 지능적인 AI를 개발하기 위해, 기존 접근법은 보통 behavioral dimension(예: affective, cognitive, 또는 social attribute)을 서로 분리된 채로 모델링한다. 이러한 방식은 유용하지만, task-specific modeling은 training cost를 증가시키고 behavioral setting 간 generalization을 제한한다. 최근의 reasoning RL 방법은 하나의 통합 모델을 여러 behavioral task에 걸쳐 학습할 수 있게 하지만, 서로 이질적인 behavioral data 전반에서의 학습을 명시적으로 다루지는 않는다. 이 공백을 메우기 위해, 우리는 이질적인 task와 sample 사이에서 학습을 균형 있게 조정하는 RL 방법인 Heterogeneity-Aware Relative Policy Optimization (HARPO)를 소개한다. 이는 advantage를 조절하여 policy optimization 동안 어떤 단일 task나 sample도 과도한 영향을 갖지 않도록 하는 방식으로 달성된다. HARPO를 사용하여, 우리는 social behavior processing을 위한 foundation model인 OMNISAPIENS-7B 2.0을 개발하고 공개한다. 기존 behavioral foundation model과 비교할 때, OMNISAPIENS-7B 2.0은 behavioral task 전반에서 가장 강력한 성능을 달성하며, multitask 및 held-out setting에서 각각 최대 +16.85%와 +9.37%의 향상을 보이는 동시에 더 명시적이고 견고한 reasoning trace를 생성한다. 또한 우리는 HARPO를 최근 RL 방법들과 비교 검증했으며, behavioral task 전반에서 가장 일관되게 강한 성능을 달성함을 확인했다.

"},{"id":"63832","en":"Reading Between the Tokens: Improving Preference Predictions through Mechanistic Forecasting","ko":"토큰 사이를 읽기: mechanistic forecasting을 통한 preference prediction 개선","authors":"Sarah Ball, Simeon Allmendinger, Frauke Kreuter, Niklas Kühl","pos":"#606","link":"https://openreview.net/forum?id=TJvRtbspej","abs":"

Large language models are increasingly used to predict human preferences in both scientific and business endeavors, yet current approaches rely exclusively on analyzing model outputs without considering the underlying mechanisms. Using election forecasting as a test case, we introduce mechanistic forecasting, a method that demonstrates that probing internal model representations offers a fundamentally different---and sometimes more effective--- approach to preference prediction. Examining over 24 million configurations across 7 models, 6 national elections, multiple persona attributes, and prompt variations, we systematically analyze how demographic and ideological information activates latent party-encoding components within the respective models. We find that leveraging this internal knowledge via mechanistic forecasting (opposed to solely relying on surface-level predictions) can improve prediction accuracy. The effects vary across demographic versus opinion-based attributes, political parties, national contexts, and models. Our findings demonstrate that the latent representational structure of LLMs contains systematic, exploitable information about human preferences, establishing a new paradigm for using language models in social science prediction tasks.

","absKo":"

대규모 언어 모델은 과학 및 비즈니스 활동에서 인간의 선호를 예측하는 데 점점 더 많이 사용되고 있지만, 현재 접근법은 underlying mechanism을 고려하지 않고 model output만 분석하는 데 의존한다. 선거 예측을 test case로 사용하여, 우리는 mechanistic forecasting를 도입한다. 이는 내부 model representation을 probing하는 것이 선호 예측에 대해 근본적으로 다르고 때로는 더 효과적인 접근임을 보여주는 방법이다. 7개 model, 6개 국가 선거, 다양한 persona attribute와 prompt variation에 걸친 2,400만 개 이상의 configuration을 살펴보며, 우리는 demographic 및 ideological information이 각 model 내부의 latent party-encoding component를 어떻게 활성화하는지 체계적으로 분석한다. 우리는 mechanistic forecasting을 통해 이러한 internal knowledge를 활용할 때(표면 수준 예측에만 의존하는 것과 달리) prediction accuracy를 향상시킬 수 있음을 발견한다. 이러한 효과는 demographic attribute와 opinion-based attribute, political party, national context, model에 따라 다르게 나타난다. 우리의 발견은 LLM의 latent representational structure가 인간 선호에 대한 체계적이고 활용 가능한 정보를 담고 있음을 보여주며, social science prediction task에서 language model을 사용하는 새로운 paradigm을 확립한다.

"},{"id":"64764","en":"Reduction of Probabilistic Chemical Reaction Networks","ko":"확률적 화학 반응 네트워크의 축소","authors":"Mauricio Montes, Grégoire Sergeant-Perthuis","pos":"#1400","link":"https://openreview.net/forum?id=K8TajLXmiD","abs":"

Programming adaptive behaviors at the cellular level is a long-standing goal that raises the question of how probabilistic computation can be implemented in biochemical systems. Chemical reaction networks (CRNs) provide such a substrate and have been shown to realize probabilistic models, including hidden Markov models and factor graphs, with dynamics reproducing Bayesian inference and belief propagation. However, encoding these algorithms typically requires prohibitively large reaction networks, and classical CRN reduction techniques don't apply. By embedding CRNs into factor graphs in a structure- and dynamics-preserving manner, we leverage recent factor-graph reduction results to obtain significantly smaller CRNs, a gain we demonstrate numerically.

","absKo":"

세포 수준에서 adaptive behavior를 프로그래밍하는 것은 오랜 목표이며, 이는 biochemical system에서 probabilistic computation을 어떻게 구현할 수 있는지에 대한 질문을 제기한다. Chemical reaction networks (CRNs)는 그러한 substrate를 제공하며, hidden Markov model과 factor graph를 포함한 probabilistic model을 구현할 수 있고, 그 dynamics는 Bayesian inference와 belief propagation을 재현하는 것으로 나타났다. 그러나 이러한 algorithm을 encoding하려면 대개 지나치게 큰 reaction network가 필요하며, 고전적인 CRN reduction 기법은 적용되지 않는다. 우리는 CRN을 structure와 dynamics를 보존하는 방식으로 factor graph에 embedding함으로써, 최근의 factor-graph reduction 결과를 활용하여 훨씬 더 작은 CRN을 얻는다. 우리는 이를 수치적으로 입증한다.

"},{"id":"61540","en":"RetrOrchestrator: A Multi-Step Retrosynthesis Agent Dynamically Orchestrating Single-Step Transition Models","ko":"RetrOrchestrator: Single-Step Transition Models를 동적으로 조율하는 Multi-Step Retrosynthesis Agent","authors":"Liao Chang, Luotian Yuan, Yiping Ke, Ying Wei","pos":"#1401","link":"https://openreview.net/forum?id=p6gN6f8pdy","abs":"

Multi-step retrosynthesis planning is a fundamental challenge in organic chemistry, defined by its enormous search space. Existing methods typically formulate it as a Markov Decision Process (MDP) with a fixed choice of transition model (i.e., a single-step retrosynthesis model), and focus on improving how to search through better policies and value functions. However, how the transition space itself is navigated remains largely unexplored. This limitation is particularly urgent given our observation of pronounced skill disparity among single-step prediction models: different models exhibit substantially different performance across molecule states. Motivated by this observation, we introduce RetrOrchestrator, an LLM-powered agent that explicitly accounts for model skill disparity by reframing retrosynthesis planning as a Partially Observable Markov Decision Process (POMDP). By regarding each single-step prediction model as a tool, we further propose a scaffold-aware reinforcement learning algorithm to optimize navigation policy within the transition space. As a result, RetrOrchestrator jointly searches which molecule to expand and which single-step model to apply for the molecule at the current step. Empirically, RetrOrchestrator significantly outperforms static baselines on the Retro*-190 benchmark, achieving a state-of-the-art 94.21\\% success rate as well as a Pareto front in both wallclock time and number of model queries.

","absKo":"

Multi-step retrosynthesis planning은 방대한 search space로 정의되는 organic chemistry의 근본적 도전 과제다. 기존 방법은 보통 이를 고정된 transition model 선택, 즉 single-step retrosynthesis model을 사용한 Markov Decision Process(MDP)로 정식화하고, 더 나은 policy와 value function을 통해 어떻게 search할 것인가에 집중한다. 그러나 transition space 자체가 어떻게 탐색되는가는 대부분 미탐구 상태로 남아 있다. 이 한계는 single-step prediction model 간에 두드러진 skill disparity가 존재한다는 우리의 관찰을 고려하면 특히 시급하다. 서로 다른 model이 molecule state에 따라 크게 다른 성능을 보이기 때문이다. 이러한 관찰에 착안하여, 우리는 각 model의 skill disparity를 명시적으로 반영하도록 retrosynthesis planning을 Partially Observable Markov Decision Process(POMDP)로 재구성한 LLM 기반 agent RetrOrchestrator를 제안한다. 각 single-step prediction model을 tool로 간주함으로써, 우리는 transition space 내 navigation policy를 최적화하는 scaffold-aware reinforcement learning algorithm도 추가로 제안한다. 그 결과 RetrOrchestrator는 현재 단계에서 어떤 molecule을 확장할지와 어떤 single-step model을 적용할지를 jointly search한다. 실험적으로 RetrOrchestrator는 Retro*-190 benchmark에서 static baseline을 크게 능가하며, state-of-the-art 94.21\\% success rate와 wallclock time 및 model query 개수 모두에서 Pareto front를 달성한다.

"},{"id":"64759","en":"Spectral-Informed Neural Networks Outperform Spectral methods in High-dimensional PDEs","ko":"Spectral-Informed 신경망이 고차원 PDE에서 Spectral 방법을 능가한다","authors":"Tianchi Yu, Ivan Oseledets","pos":"#1402","link":"https://openreview.net/forum?id=KAHCMPsPeI","abs":"For low-dimensional problems ($d\\leq3$), spectral methods can achieve exceptionally high accuracy. For middle-dimensional problems ($4 \\leq d \\lesssim 10$), spectral methods remain feasible through specific techniques such as sparse grids or hyperbolic cross. However, for high-dimensional problems ($d\\gg 10$), spectral methods suffer from the curse of dimensionality. Physics-informed neural networks (PINNs) have emerged as a promising approach to overcome this challenge, offering scalability to high dimensions, but often suffer from limited accuracy and efficiency. Recently proposed spectral-informed neural networks (SINNs) combine spectral methods with PINNs, operating directly in the spectral domain to avoid spatial derivative computations and to reduce memory consumption. In this work, we introduce Modified SINNs, which integrate coefficient decay scaling and basis embeddings motivated by harmonic analysis to enhance accuracy in high-dimensional problems and enable accurate approximation of unknown spectral coefficients. Numerical experiments on steady and time-dependent partial differential equations demonstrate that Modified SINNs outperform sparse grid spectral methods on middle-dimensional problems with incomplete spectral information and achieve superior accuracy compared to PINNs on high-dimensional problems.","absKo":"저차원 문제($d\\leq3$)에서는 spectral method가 매우 높은 정확도를 달성할 수 있다. 중차원 문제($4 \\leq d \\lesssim 10$)에서는 sparse grid나 hyperbolic cross와 같은 특정 기법을 통해 여전히 spectral method를 적용할 수 있다. 그러나 고차원 문제($d\\gg 10$)에서는 spectral method가 curse of dimensionality의 영향을 받는다. Physics-informed neural network(PINN)은 이러한 도전을 극복할 유망한 접근법으로 부상했으며, 고차원으로의 확장성을 제공하지만 종종 정확도와 효율성이 제한된다. 최근 제안된 spectral-informed neural network(SINN)는 spectral method와 PINN을 결합하여, spatial derivative 계산을 피하고 메모리 소비를 줄이기 위해 spectral domain에서 직접 작동한다. 본 연구에서는 harmonic analysis에서 동기를 얻은 coefficient decay scaling과 basis embedding을 통합하여 고차원 문제에서 정확도를 높이고, 알려지지 않은 spectral coefficient를 정확하게 근사할 수 있게 하는 Modified SINN을 소개한다. 정상 및 시간 의존 partial differential equation에 대한 수치 실험은 Modified SINN이 중차원 문제에서 불완전한 spectral information을 다루는 sparse grid spectral method보다 우수하고, 고차원 문제에서는 PINN보다 더 뛰어난 정확도를 달성함을 보여준다."},{"id":"62310","en":"Stochastic Lifting for Generating Trajectories of Stochastic Physical Systems","ko":"확률적 물리 시스템 궤적 생성을 위한 Stochastic Lifting","authors":"Jules Berman, Tobias Blickhan, Benjamin Peherstorfer","pos":"#1403","link":"https://openreview.net/forum?id=hmJdJLmwOY","abs":"

Many stochastic physical systems evolve smoothly over time in the sense that the distribution of states changes regularly with time. The precise transition from current to next state is often modeled as the interplay of a smooth map and an explicit source of randomness. Stochastic Lifting leverages this premise by attaching an independent, high-dimensional random label to each state transition in the training data and fitting a transition map from the current state and label to the next state using a standard regression loss. The labels act as auxiliary coordinates that let the model represent multiple plausible outcomes for similar current states, avoiding collapse to a mean prediction in the finite-sample size regime. At inference, drawing fresh labels and rolling the map forward generates diverse trajectories with a single network evaluation per time step, with the smoothness bias of the learned map supporting accurate sampling in practice.

","absKo":"

많은 확률적 물리계는 상태 분포가 시간에 따라 규칙적으로 변한다는 의미에서 시간에 따라 매끄럽게 진화한다. 현재 상태에서 다음 상태로의 정확한 전이는 흔히 매끄러운 map과 명시적인 무작위성의 결합으로 모델링된다. Stochastic Lifting은 이러한 전제를 활용하여, training data의 각 state transition에 독립적인 고차원 random label을 부착하고, 표준 regression loss를 사용해 현재 state와 label로부터 다음 state로의 transition map을 적합한다. 이 label은 유사한 현재 상태들에 대해 여러 가능한 결과를 모델이 표현할 수 있게 해주는 보조 좌표로 작용하여, 유한 샘플 규모 영역에서 평균 예측으로 붕괴하는 것을 방지한다. 추론 시에는 새로운 label을 샘플링하고 map을 순차적으로 전개함으로써, 시점당 한 번의 network evaluation만으로 다양한 trajectory를 생성할 수 있으며, 학습된 map의 smoothness bias가 실제 샘플링의 정확도를 뒷받침한다.

"},{"id":"62469","en":"Test-time Generalization for Physics through Neural Operator Splitting","ko":"Neural Operator Splitting을 통한 Test-time Generalization for Physics","authors":"Louis Serrano, Rudy Morel, Jiequn Han, Edouard Oyallon, Shirley Ho","pos":"#1404","link":"https://openreview.net/forum?id=gLLcjBJCXZ","abs":"

Neural operators have shown promise in learning solution maps of partial differential equations (PDEs), but they often struggle to generalize when test inputs lie outside the training distribution, such as novel initial conditions, unseen PDE coefficients or unseen physics. Prior works address this limitation with large scale multi physics pretraining followed by fine tuning, but this still requires examples from the new dynamics, falling short of true zero shot generalization. In this work, we propose a method to enhance generalization at test-time, i.e, without modifying pretrained weights.Building on DISCO, which provides a dictionary of neural operators trained across different dynamics, we introduce a neural operator splitting strategy that, at test time, searches over compositions of training operators to approximate unseen dynamics. On challenging out-of-distribution tasks including parameter extrapolation and novel combinations of physics phenomena, our approach achieves state-of-the-art zero shot generalization results, while being able to recover the underlying PDE parameters. These results underscore test-time computation as a key avenue for building flexible, compositional, and generalizable neural operators.

","absKo":"

Neural operator는 partial differential equations(PDEs)의 해 맵을 학습하는 데 유망함을 보여왔지만, test input이 training distribution 밖에 놓일 때, 예를 들어 새로운 initial condition, 보지 못한 PDE coefficient, 또는 보지 못한 physics에서 일반화에 어려움을 겪는 경우가 많다. 이전 연구들은 대규모 multi-physics pretraining 뒤 fine-tuning으로 이 한계를 다루었지만, 이는 여전히 새로운 dynamics의 예시를 필요로 하므로 진정한 zero-shot generalization에는 미치지 못한다. 본 연구에서는 pretrained weights를 수정하지 않고 test-time에서 generalization을 향상시키는 방법을 제안한다. 서로 다른 dynamics에 걸쳐 학습된 neural operator의 dictionary를 제공하는 DISCO를 기반으로, 우리는 test time에 training operator의 composition을 탐색하여 보지 못한 dynamics를 근사하는 neural operator splitting 전략을 도입한다. parameter extrapolation과 새로운 physics phenomenon의 조합을 포함한 까다로운 out-of-distribution task에서, 우리의 접근법은 underlying PDE parameter를 복원할 수 있으면서도 state-of-the-art zero-shot generalization 결과를 달성한다. 이러한 결과는 유연하고 조합 가능하며 일반화 가능한 neural operator를 구축하는 핵심 경로로서 test-time computation의 중요성을 강조한다.

"},{"id":"62217","en":"Training-Free Bayesian Filtering with Generative Emulators","ko":"생성형 Emulator를 활용한 Training-Free Bayesian Filtering","authors":"Thomas Savary, François Rozet, Gilles Louppe","pos":"#1405","link":"https://openreview.net/forum?id=ibcZNZwKfZ","abs":"

Bayesian filtering is a well-known problem that aims to estimate plausible states of a dynamical system from observations. Among existing approaches to solve this problem, particles filters are theoretically exact for non-linear dynamics and observations, but suffer from poor scalability in high dimensions. In this work, we show that diffusion-based emulators of dynamical systems can be used to implement, without additional training, an optimal variant of particle filters that has remained largely unexplored due to implementation challenges with classical numerical solvers. Experiments on nonlinear chaotic systems, including atmospheric dynamics, demonstrate that the proposed approach successfully scales particle filtering to high-dimensional settings.

","absKo":"

Bayesian filtering은 관측으로부터 동역학 시스템의 가능한 상태를 추정하는 것을 목표로 하는 잘 알려진 문제다. 이 문제를 해결하는 기존 접근법 중 particle filter는 비선형 dynamics와 observation에 대해 이론적으로 정확하지만, 고차원에서는 낮은 scalability를 보인다. 본 연구에서는 동역학 시스템의 diffusion-based emulator를 추가 학습 없이 사용하여, 고전적인 numerical solver에서의 구현 난이도 때문에 대부분 탐구되지 못했던 optimal variant의 particle filter를 구현할 수 있음을 보인다. atmospheric dynamics를 포함한 비선형 혼돈 시스템에서의 실험은, 제안한 접근법이 particle filtering을 고차원 설정으로 성공적으로 확장함을 보여준다.

"},{"id":"61167","en":"Physics-Informed Distillation of Diffusion Models for PDE-Constrained Generation","ko":"PDE 제약 생성을 위한 Diffusion Model의 Physics-Informed Distillation","authors":"Yi Zhang, Peng Wang, Difan Zou","pos":"#1501","link":"https://openreview.net/forum?id=t6dA9c3SlJ","abs":"

Diffusion models show growing promise for generative modeling of physical systems, but enforcing partial differential equation (PDE) constraints directly is infeasible during the stochastic denoising process. Current methods apply constraints to the expected clean sample, incurring a Jensen’s Gap that forces a trade-off between PDE satisfaction and generative accuracy. To bridge this gap, we propose Physics-Informed Distillation of Diffusion Models (PIDDM), a simple yet effective post-hoc distillation strategy that enforces PDE constraints after training. PIDDM enables fast single-step generation while improving both physical consistency and sample quality, supporting forward/inverse problems and reconstruction from partial observations. Extensive experiments across PDE benchmarks show PIDDM outperforms recent baselines, such as PIDM, DiffusionPDE, and ECI-sampling, in both accuracy and constraint satisfaction, with lower computation and minimal hyperparameter tuning, offering a more efficient pathway to physics-informed diffusion models.

","absKo":"

Diffusion model은 물리 시스템의 generative modeling에서 점점 더 유망한 성과를 보이고 있지만, stochastic denoising 과정에서 partial differential equation (PDE) 제약을 직접 강제하는 것은 사실상 불가능하다. 기존 방법들은 기대되는 clean sample에 제약을 적용하는데, 이 과정에서 Jensen’s Gap이 발생하여 PDE 만족도와 생성 정확도 사이의 trade-off가 생긴다. 이 간극을 메우기 위해, 우리는 학습 이후 PDE 제약을 강제하는 간단하지만 효과적인 post-hoc distillation 전략인 Physics-Informed Distillation of Diffusion Models (PIDDM)을 제안한다. PIDDM은 빠른 single-step 생성을 가능하게 하면서 물리적 일관성과 sample quality를 모두 향상시키며, forward/inverse problem과 부분 관측으로부터의 reconstruction을 지원한다. 다양한 PDE 벤치마크에서 수행한 대규모 실험은 PIDDM이 PIDM, DiffusionPDE, ECI-sampling과 같은 최근 baseline보다 정확도와 제약 만족도 모두에서 더 우수하며, 계산량은 더 적고 hyperparameter tuning도 최소화된다는 점을 보여준다. 이는 physics-informed diffusion model로 가는 더 효율적인 경로를 제시한다.

"},{"id":"60682","en":"Open Materials Generation with Inference-Time Reinforcement Learning","ko":"추론 시 Reinforcement Learning을 활용한 Open Materials Generation","authors":"Philipp Höllmer, Stefano Martiniani","pos":"#1502","link":"https://openreview.net/forum?id=xfHppnGXaH","abs":"

Continuous-time generative models for crystalline materials enable inverse materials design by learning to predict stable crystal structures, but incorporating explicit target properties into the generative process remains challenging. Policy-gradient reinforcement learning (RL) provides a principled mechanism for aligning generative models with downstream objectives but typically requires access to the score, which has prevented its application to flow-based models that learn only velocity fields. We introduce Open Materials Generation with Inference-time Reinforcement Learning (OMatG-IRL), a policy-gradient RL framework that operates directly on the learned velocity fields and eliminates the need for the explicit computation of the score. OMatG-IRL leverages stochastic perturbations of the underlying generation dynamics preserving the baseline performance of the pretrained generative model while enabling exploration and policy-gradient estimation at inference time. Using OMatG-IRL, we present the first application of RL to crystal structure prediction (CSP). Our method enables effective reinforcement of an energy-based objective while preserving diversity through composition conditioning, and it achieves performance competitive with score-based RL approaches. Finally, we show that OMatG-IRL can learn time-dependent velocity-annealing schedules, enabling accurate CSP with order-of-magnitude improvements in sampling efficiency and, correspondingly, reduction in generation time.

","absKo":"

crystalline material을 위한 continuous-time generative model은 stable crystal structure를 예측하도록 학습함으로써 inverse materials design을 가능하게 하지만, 명시적인 target property를 generative process에 통합하는 것은 여전히 어렵다. Policy-gradient reinforcement learning (RL)은 generative model을 downstream objective와 정렬하는 원칙적인 메커니즘을 제공하지만, 일반적으로 score에 대한 접근이 필요하므로 velocity field만 학습하는 flow-based model에는 적용이 어려웠다. 우리는 학습된 velocity field에 직접 작동하며 explicit score 계산의 필요성을 제거하는 policy-gradient RL framework인 Open Materials Generation with Inference-time Reinforcement Learning (OMatG-IRL)을 소개한다. OMatG-IRL은 사전학습된 generative model의 baseline 성능을 보존하면서도 inference time에서 탐색과 policy-gradient 추정을 가능하게 하도록, underlying generation dynamics의 stochastic perturbation을 활용한다. OMatG-IRL을 사용하여, 우리는 crystal structure prediction (CSP)에 대한 RL의 첫 적용을 제시한다. 우리의 방법은 composition conditioning을 통해 다양성을 보존하면서 energy-based objective를 효과적으로 강화할 수 있으며, score-based RL 접근법과 경쟁력 있는 성능을 달성한다. 마지막으로, OMatG-IRL이 time-dependent velocity-annealing schedule을 학습할 수 있음을 보이며, sampling efficiency를 한 자릿수 배수 이상 향상시키고 그에 따라 generation time을 줄이면서 정확한 CSP를 가능하게 함을 보인다.

"},{"id":"62073","en":"Meta-learning Structure-Preserving Dynamics","ko":"구조 보존 동역학을 위한 Meta-learning","authors":"Cheng Jing, Uvini Mudiyanselage, Woojin Cho, Minju Jo, Anthony Gruber, Kookjin Lee","pos":"#1503","link":"https://openreview.net/forum?id=k66TZFhUSQ","abs":"

Structure-preserving approaches to dynamics discovery have demonstrated great potential for modeling physical systems due to their use of strong inductive biases, which enforce key features such as conservation laws and dissipative behavior. However, these models are typically trained on a per-configuration basis, requiring explicit knowledge of system parameters and costly retraining when these parameters vary. While meta-learning provides a potential remedy, optimization-based approaches can suffer from limited generalizability. Motivated by recent advances in modulation-based learning aimed at mitigating these drawbacks, we systematically investigate the use of modulation techniques in learning conservative dynamical systems. We study a range of existing modulation strategies alongside newly proposed variants, integrating them into a Hamiltonian learning framework without requiring an explicit system parameterization. Through extensive experiments on benchmark problems, we demonstrate that modulation-based meta-learning enables accurate few-shot adaptation, achieving robust generalization across parameter space without compromising the conservation of key invariants responsible for the dynamics.

","absKo":"

동역학 발견을 위한 structure-preserving approach는 보존 법칙과 dissipative behavior와 같은 핵심 특징을 강제하는 강한 inductive bias를 활용하기 때문에 물리 시스템 모델링에서 큰 잠재력을 보여왔다. 그러나 이러한 모델은 대개 각 configuration마다 개별적으로 학습되므로, system parameter에 대한 명시적 지식이 필요하고, 이러한 parameter가 변할 때 비용이 큰 재학습이 요구된다. meta-learning이 잠재적 해결책을 제공하지만, optimization-based approach는 generalizability가 제한될 수 있다. 이러한 단점을 완화하기 위한 최근 modulation-based learning의 발전에 동기를 받아, 우리는 conservative dynamical system 학습에서 modulation technique의 사용을 체계적으로 조사한다. 우리는 새로 제안된 variant들과 함께 다양한 기존 modulation strategy를 연구하고, 명시적 system parameterization 없이 Hamiltonian learning framework에 통합한다. benchmark 문제에 대한 광범위한 실험을 통해, modulation-based meta-learning이 정확한 few-shot adaptation을 가능하게 하며, 동역학을 담당하는 핵심 invariant의 보존을 해치지 않으면서 parameter space 전반에 걸쳐 robust generalization을 달성함을 보인다.

"},{"id":"62418","en":"Imposing Boundary Conditions on Neural Operators via Learned Function Extensions","ko":"학습된 함수 확장을 통한 Neural Operators의 경계 조건 부여","authors":"Sepehr Mousavi, Siddhartha Mishra, Laura De Lorenzis","pos":"#1504","link":"https://openreview.net/forum?id=gnzNB7GFFo","abs":"

Neural operators have emerged as powerful surrogates for the solution of partial differential equations (PDEs), yet their ability to handle general, highly variable boundary conditions (BCs) remains limited. Existing approaches often fail when the solution operator exhibits strong sensitivity to boundary forcings. We propose a general framework for conditioning neural operators on complex non-homogeneous BCs through function extensions. Our key idea is to map boundary data to latent pseudo-extensions defined over the entire spatial domain, enabling any standard operator learning architecture to consume boundary information. The resulting operator, coupled with an arbitrary domain-to-domain neural operator, can learn rich dependencies on complex BCs and input domain functions at the same time. To benchmark this setting, we construct 18 challenging datasets spanning Poisson, linear elasticity, and hyperelasticity problems, with highly variable, mixed-type, component-wise, and multi-segment BCs on diverse geometries. Our approach achieves state-of-the-art accuracy, outperforming baselines by large margins, while requiring no hyperparameter tuning across datasets. Overall, our results demonstrate that learning boundary-to-domain extensions is an effective and practical strategy for imposing complex BCs in existing neural operator frameworks, enabling accurate and robust scientific machine learning models for a broader range of PDE-governed problems.

","absKo":"

Neural operator는 partial differential equations (PDEs)의 해를 위한 강력한 surrogate로 등장했지만, 일반적이고 매우 다양한 boundary conditions (BCs)을 처리하는 능력은 여전히 제한적이다. 기존 접근은 solution operator가 boundary forcing에 강하게 민감할 때 종종 실패한다. 우리는 function extension을 통해 복잡한 non-homogeneous BC에 neural operator를 조건화하는 일반 프레임워크를 제안한다. 핵심 아이디어는 boundary data를 전체 spatial domain에 정의된 latent pseudo-extension으로 매핑하여, 어떤 표준 operator learning architecture라도 boundary information을 소비할 수 있게 하는 것이다. 그 결과 도메인-투-도메인 neural operator와 결합된 operator는 복잡한 BC와 입력 domain function에 대한 풍부한 의존성을 동시에 학습할 수 있다. 이 설정을 평가하기 위해, 우리는 Poisson, linear elasticity, hyperelasticity 문제 전반에 걸쳐 18개의 도전적인 데이터셋을 구성했으며, 이들은 다양한 geometry에서 매우 가변적이고 mixed-type이며 component-wise이고 multi-segment인 BC를 포함한다. 우리의 접근은 데이터셋 간 hyperparameter tuning 없이도 큰 폭으로 baseline을 능가하며 최신 정확도를 달성한다. 전반적으로, 우리의 결과는 boundary-to-domain extension을 학습하는 것이 기존 neural operator framework에서 복잡한 BC를 부과하는 효과적이고 실용적인 전략이며, 더 넓은 범위의 PDE-governed problem에 대해 정확하고 견고한 scientific machine learning model을 가능하게 함을 보여준다.

"},{"id":"60975","en":"Heavy-tailed Physics-Informed Neural Networks","ko":"Heavy-tailed 물리 정보 기반 Neural Network","authors":"Jephte Abijuru, Mayank Kumar Nagda, Phil Sidney Ostheimer, Jan Tauberschmidt, Sebastian Vollmer, Stephan Mandt, Marius Kloft, Sophie Fellenz","pos":"#1505","link":"https://openreview.net/forum?id=uvB4kU3jTd","abs":"Physics-informed neural networks (PINNs) enforce physical laws by minimizing partial differential equation (PDE) residuals and auxiliary constraints. Standard training relies on a mean-squared error (MSE) objective, which implicitly assumes independent Gaussian residuals with a fixed global variance. We show theoretically and empirically that residuals encountered during PINN training are heterogeneous and heavy-tailed, revealing a systematic mismatch with this assumption. As a consequence, a small number of large residuals can disproportionately dominate both the loss and gradient, leading to poorly balanced optimization dynamics. Motivated by this mismatch, we adopt a Student-$t$ residual model to explicitly capture heavy-tailed behavior. An equivalent hierarchical representation yields an expectation–maximization (EM) algorithm that alternates between estimating residual-dependent weights and optimizing network parameters via a weighted MSE objective, allowing existing PINN solvers to be reused in the M-step. The resulting training dynamics bound the influence of extreme residuals and admit almost sure convergence guarantees under standard stochastic optimization assumptions. Experiments across a diverse suite of challenging PDE benchmarks demonstrate consistently improved solution accuracy and robustness compared to standard PINN training.","absKo":"

Physics-informed neural network(PINN)는 partial differential equation(PDE) residual과 보조 제약을 최소화함으로써 물리 법칙을 강제한다. 표준 학습은 mean-squared error(MSE) objective에 의존하는데, 이는 암묵적으로 독립적인 Gaussian residual과 고정된 전역 분산을 가정한다. 우리는 이론적으로 그리고 실증적으로, PINN 학습 동안 마주치는 residual이 heterogeneous하고 heavy-tailed임을 보이며, 이 가정과 체계적으로 어긋남을 드러낸다. 그 결과, 소수의 큰 residual이 loss와 gradient를 모두 과도하게 지배할 수 있고, 이는 균형이 좋지 않은 optimization dynamics로 이어진다. 이러한 불일치에 동기를 얻어, 우리는 heavy-tailed behavior를 명시적으로 포착하기 위해 Student-$t$ residual model을 채택한다. 이에 상응하는 hierarchical representation은 expectation–maximization(EM) algorithm을 제공하며, 이는 residual-dependent weight를 추정하는 단계와 weighted MSE objective로 network parameter를 최적화하는 단계를 번갈아 수행한다. 이를 통해 기존 PINN solver를 M-step에서 재사용할 수 있다. 이렇게 얻어진 training dynamics는 극단적인 residual의 영향을 제한하며, 표준 stochastic optimization 가정 하에서 almost sure convergence 보장을 허용한다. 다양한 challenging PDE benchmark에 대한 실험은 standard PINN training과 비교하여 일관되게 향상된 solution accuracy와 robustness를 보여준다.

"},{"id":"63611","en":"Generative Neural Operators through Diffusion Last Layer","ko":"Diffusion Last Layer를 통한 생성 신경 연산자","authors":"Sungwon Park, Anthony Zhou, Hongjoong Kim, Amir Barati Farimani","pos":"#1506","link":"https://openreview.net/forum?id=VURgLSSeQi","abs":"

Neural operators have emerged as a powerful paradigm for learning discretization-invariant function-to-function mappings in scientific computing. However, many practical systems are inherently stochastic, making principled uncertainty quantification essential for reliable deployment. To address this, we introduce a simple add-on, the diffusion last layer (DLL), a lightweight probabilistic head that can be attached to arbitrary neural operator backbones to model predictive uncertainty. Motivated by the relative smoothness and low-dimensional structure often exhibited by PDE solution distributions, DLL parameterizes the conditional output distribution directly in function space through a low-rank Karhunen-Loève expansion, enabling efficient and expressive uncertainty modeling. Across stochastic PDE operator learning benchmarks, DLL improves generalization and uncertainty-aware prediction. Moreover, even in deterministic long-horizon rollout settings, DLL enhances rollout stability and provides meaningful estimates of epistemic uncertainty for backbone neural operators.

","absKo":"

Neural operator는 scientific computing에서 discretization에 불변인 function-to-function mapping을 학습하는 강력한 패러다임으로 부상했다. 그러나 많은 실제 시스템은 본질적으로 stochastic하므로, 신뢰할 수 있는 배포를 위해 원리적인 uncertainty quantification이 필수적이다. 이를 위해 우리는 임의의 neural operator backbone에 부착할 수 있는 가벼운 probabilistic head인 diffusion last layer (DLL)라는 간단한 add-on을 도입한다. PDE 해 분포가 흔히 보이는 상대적 smoothness와 저차원 구조에 동기부여되어, DLL은 low-rank Karhunen-Loève expansion을 통해 function space에서 conditional output distribution을 직접 parameterize 하며, 이를 통해 효율적이고 표현력 있는 uncertainty modeling을 가능하게 한다. stochastic PDE operator learning benchmark 전반에서 DLL은 generalization과 uncertainty-aware prediction을 향상시킨다. 더 나아가 deterministic long-horizon rollout 설정에서도 DLL은 rollout stability를 높이고, backbone neural operator에 대해 의미 있는 epistemic uncertainty 추정치를 제공한다.

"},{"id":"61823","en":"Excited Pfaffians: Generalized Neural Wave Functions Across Structure and State","ko":"Excited Pfaffians: 구조와 상태를 아우르는 일반화된 Neural Wave Function","authors":"Nicholas Gao, Till Grutschus, Frank Noe, Stephan Günnemann","pos":"#1507","link":"https://openreview.net/forum?id=meEsugjXjv","abs":"Neural-network wave functions in Variational Monte Carlo (VMC) have achieved great success in accurately representing both ground and excited states. However, achieving sufficient numerical accuracy of state overlaps requires growing the number of Monte Carlo samples, and consequently computational cost, with the number of states. We present a nearly constant sample size approach named Multi-State Importance Sampling (MSIS), which leverages all states' samples to estimate any pairwise overlap. To efficiently evaluate all states for all samples, we introduce Excited Pfaffians. Inspired by Hartree-Fock, this architecture represents many states within a single neural network. Excited Pfaffians also serve as generalized wave functions, allowing a single model to represent multi-state potential energy surfaces. On the carbon dimer, we match the $\\mathcal{O}(N_s^4)$-scaling natural excited states while training $>200\\times$ faster and modeling 50% more states. Our favorable scaling enables us to be the first to use neural networks to find all distinct energy levels of the Beryllium atom. Finally, we demonstrate that a single wave function can represent excited states across various molecules.","absKo":"신경망 wave function을 Variational Monte Carlo (VMC)에서 사용하는 방식은 ground state와 excited state를 모두 정확하게 표현하는 데 큰 성공을 거두었다. 그러나 state overlap의 충분한 수치 정확도를 달성하려면 Monte Carlo sample 수를 state 수에 따라 늘려야 하며, 그에 따라 계산 비용도 증가한다. 우리는 모든 state의 sample을 활용하여 임의의 pairwise overlap을 추정하는, sample 크기가 거의 일정한 접근법인 Multi-State Importance Sampling (MSIS)을 제안한다. 모든 sample에 대해 모든 state를 효율적으로 평가하기 위해 우리는 Excited Pfaffians를 도입한다. Hartree-Fock에서 영감을 받은 이 아키텍처는 단일 neural network 안에 여러 state를 표현한다. Excited Pfaffians는 또한 generalized wave function으로도 작동하여, 하나의 model이 multi-state potential energy surface를 표현할 수 있게 한다. carbon dimer에서는 $\\mathcal{O}(N_s^4)$ 스케일링의 자연스러운 excited state와 일치시키면서도 학습 속도는 200배 이상 빠르고, 50% 더 많은 state를 모델링했다. 이러한 유리한 scaling 덕분에 우리는 neural network를 사용해 Beryllium 원자의 모든 서로 다른 energy level을 처음으로 찾아낼 수 있었다. 마지막으로, 하나의 wave function이 여러 molecule에 걸친 excited state를 표현할 수 있음을 보인다."},{"id":"66811","en":"Discovering Ordinary Differential Equations with LLM-Based Qualitative and Quantitative Evaluation","ko":"LLM 기반 정성적·정량적 평가를 통한 상미분방정식 발견","authors":"Sum Kyun Song, Bong Gyun Shin, JaeYong Lee","pos":"#1509","link":"https://openreview.net/forum?id=07YGl8xVZl","abs":"

Discovering governing differential equations from observational data is a fundamental challenge in scientific machine learning. Existing symbolic regression approaches rely primarily on quantitative metrics; however, real-world differential equation modeling also requires incorporating domain knowledge to ensure physical plausibility. To address this gap, we propose DoLQ, a method for discovering ordinary differential equations with LLM-based qualitative and quantitative evaluation. DoLQ employs a multi-agent architecture: a Sampler Agent proposes dynamic system candidates, a Parameter Optimizer refines equations for accuracy, and a Scientist Agent leverages an LLM to conduct both qualitative and quantitative evaluations and synthesize their results to iteratively guide the search. Experiments on multi-dimensional ordinary differential equation benchmarks demonstrate that DoLQ achieves superior performance compared to existing methods, not only attaining higher success rates but also more accurately recovering the correct symbolic terms of ground truth equations. Our code is available at https://anonymous.4open.science/r/DoLQ/README.md.

","absKo":"

관측 데이터로부터 governing differential equation을 발견하는 것은 scientific machine learning의 핵심 과제다. 기존의 symbolic regression 접근법은 주로 quantitative metric에 의존하지만, 실제 differential equation modeling은 물리적 타당성을 보장하기 위해 domain knowledge를 함께 반영해야 한다. 이 격차를 해소하기 위해, 우리는 LLM-based qualitative 및 quantitative evaluation을 사용하여 ordinary differential equation을 발견하는 방법 DoLQ를 제안한다. DoLQ는 multi-agent architecture를 사용한다. Sampler Agent는 동적 시스템 후보를 제안하고, Parameter Optimizer는 정확도를 위해 equation을 정교화하며, Scientist Agent는 LLM을 활용해 qualitative 및 quantitative evaluation을 모두 수행하고 그 결과를 종합하여 탐색을 반복적으로 안내한다. multi-dimensional ordinary differential equation benchmark에서 수행한 실험은 DoLQ가 기존 방법보다 우수한 성능을 달성함을 보여 준다. 이는 더 높은 성공률을 달성할 뿐 아니라, ground truth equation의 올바른 symbolic term도 더 정확하게 복원함을 의미한다. 우리의 code는 https://anonymous.4open.science/r/DoLQ/README.md 에서 확인할 수 있다.

"},{"id":"66309","en":"Deep Coupling Learning for Solving PDEs","ko":"PDE 풀이를 위한 심층 결합 학습","authors":"Lingshi MENG, Haosen Shi, Sinno Jialin Pan","pos":"#1510","link":"https://openreview.net/forum?id=4pfxoYPQSY","abs":"

Physics-Informed Neural Networks (PINNs) represent a significant advancement in computational methods for solving partial differential equations (PDEs). However, the adoption of deeper neural network architectures presents significant challenges, as they struggle to address differential-related complications that arise during the computation of derivatives over the input of PINNs. These complications extend beyond traditional vanishing and exploding gradients to include vanishing and exploding differentials, with both phenomena becoming more severe as networks grow deeper. By examining the computation graph of derivatives in deep neural networks, we identify key bottlenecks causing numerical instabilities in deep architectures. In response, we introduce a novel approach that utilizes Coupling Layers with carefully regulated spectral norms of Jacobian matrices to stabilize and facilitate deep PINN training, effectively addressing differential-related challenges and improving model stability. Our proposed architecture successfully mitigates the fundamental constraints of deeper PINNs while maximizing their capabilities through consistent differential propagation. Comprehensive evaluations show that our approach surpasses conventional shallow PINN methods and alternative deep PINN designs across a range of challenging problems, particularly in cases featuring high-frequency solution components.

","absKo":"

Physics-Informed Neural Networks (PINNs)는 partial differential equations (PDEs)를 푸는 계산 방법에서 중요한 진전을 대표합니다. 그러나 더 깊은 neural network architecture를 채택하는 것은 상당한 도전이 되는데, 이는 PINNs의 input에 대한 derivative 계산 과정에서 발생하는 differential 관련 복잡성을 처리하는 데 어려움을 겪기 때문입니다. 이러한 복잡성은 전통적인 vanishing 및 exploding gradient를 넘어 vanishing 및 exploding differential까지 포함하며, 네트워크가 더 깊어질수록 두 현상 모두 더 심각해집니다. 깊은 neural network의 derivative computation graph를 분석함으로써, 우리는 deep architecture에서 numerical instability를 유발하는 핵심 병목을 식별합니다. 이에 대응하여, 우리는 Jacobian matrix의 spectral norm을 세심하게 조절한 Coupling Layer를 활용해 deep PINN training을 안정화하고 촉진하는 새로운 접근법을 제안하며, 이를 통해 differential 관련 문제를 효과적으로 해결하고 model stability를 향상시킵니다. 제안하는 architecture는 일관된 differential propagation을 통해 더 깊은 PINN의 근본적 제약을 완화하는 동시에 그 잠재력을 최대한 활용합니다. 포괄적 평가 결과, 우리의 접근법은 다양한 어려운 문제들, 특히 high-frequency solution component를 포함하는 경우에 conventional shallow PINN 방법과 대안적인 deep PINN 설계를 능가합니다.

"},{"id":"64346","en":"DAISI: Data Assimilation with Inverse Sampling using Stochastic Interpolants","ko":"DAISI: 확률적 보간자를 이용한 역 샘플링 기반 데이터 동화","authors":"Martin Andrae, Erik Larsson, So Takao, Tomas Landelius, Fredrik Lindsten","pos":"#1511","link":"https://openreview.net/forum?id=OWz1n5HgcC","abs":"

Data assimilation (DA) is a cornerstone of scientific and engineering applications, combining model forecasts with sparse and noisy observations to estimate latent system states. Classical high-dimensional DA methods, such as the ensemble Kalman filter, rely on Gaussian approximations that are violated for complex dynamics or observation operators. To address this limitation, we introduce DAISI, a scalable filtering algorithm built on flow-based generative models that enables flexible probabilistic inference using data-driven priors. The core idea is to use a stationary, pre-trained generative prior that first incorporates forecast information through a novel inverse-sampling step, before assimilating observations via guidance-based conditional sampling. This allows us to leverage any forecasting model as part of the DA pipeline without having to retrain or fine-tune the generative prior at each assimilation step. Experiments on challenging nonlinear systems show that DAISI achieves accurate filtering results in regimes with sparse, noisy, and nonlinear observations where traditional methods struggle.

","absKo":"

Data assimilation (DA)은 과학 및 공학 응용의 핵심으로, model forecast와 희소하고 잡음이 섞인 observation을 결합해 latent system state를 추정한다. ensemble Kalman filter와 같은 고차원 classical DA 방법은 복잡한 dynamics나 observation operator에 대해 위배되는 Gaussian approximation에 의존한다. 이러한 한계를 해결하기 위해, 우리는 data-driven prior를 사용한 유연한 probabilistic inference를 가능하게 하는 flow-based generative model 위의 확장 가능한 filtering algorithm인 DAISI를 제안한다. 핵심 아이디어는 stationary한 pretrained generative prior에 먼저 새로운 inverse-sampling step을 통해 forecast 정보를 통합한 뒤, guidance-based conditional sampling으로 observation을 assimilate하는 것이다. 이를 통해 매 assimilation step마다 generative prior를 다시 학습하거나 fine-tune하지 않고도 어떤 forecasting model이든 DA pipeline의 일부로 활용할 수 있다. 까다로운 nonlinear system에 대한 실험은 DAISI가 희소하고, 잡음이 많고, 비선형적인 observation이 존재하는 regime에서 전통적 방법들이 어려움을 겪는 상황에서도 정확한 filtering 결과를 달성함을 보여준다.

"},{"id":"63053","en":"Cheap2Rich: A Multi-Fidelity Framework for Data Assimilation and System Identification of Multiscale Physics - Rotating Detonation Engines","ko":"Cheap2Rich: 다중 규모 물리계의 데이터 동화 및 시스템 식별을 위한 다중 충실도 프레임워크 - 회전 데토네이션 엔진","authors":"Yuxuan Bao, Jan Zajac, Megan Powers, Venkat Raman, J. Nathan Kutz","pos":"#1512","link":"https://openreview.net/forum?id=aXOchwlOHM","abs":"

Bridging the sim2real gap between computationally inexpensive models and complex physical systems remains a central challenge in machine learning applications to engineering problems, particularly in multi-scale settings where reduced-order models typically capture only dominant dynamics. In this work, we present Cheap2Rich, a multi-scale data assimilation framework that reconstructs high-fidelity state spaces from sparse sensor histories by combining a fast low-fidelity prior with learned, interpretable discrepancy corrections. We demonstrate the performance on rotating detonation engines (RDEs), a challenging class of systems that couple detonation-front propagation with injector-driven unsteadiness, mixing, and stiff chemistry across disparate scales. Our approach successfully reconstructs high-fidelity RDE states from sparse measurements while isolating physically meaningful discrepancy dynamics associated with injector-driven effects. The results highlight a general multi-fidelity framework for data assimilation and system identification in complex multi-scale systems, enabling rapid design exploration and real-time monitoring and control while providing interpretable discrepancy dynamics. Anonymous code is available at: anonymous.4open.science/r/Cheap2Rich-4C71.

","absKo":"

계산 비용이 낮은 모델과 복잡한 물리 시스템 사이의 sim2real gap을 메우는 일은 engineering 문제에 machine learning을 적용할 때 중심적인 과제로 남아 있으며, 특히 reduced-order model이 보통 지배적인 dynamics만 포착하는 multi-scale setting에서 더욱 그렇다. 본 연구에서는 빠른 low-fidelity prior와 학습된, 해석 가능한 discrepancy correction을 결합하여 sparse sensor history로부터 high-fidelity state space를 재구성하는 multi-scale data assimilation framework인 Cheap2Rich를 제시한다. 우리는 detonation front propagation과 injector-driven unsteadiness, mixing, 그리고 서로 다른 scale에 걸친 stiff chemistry를 결합하는 까다로운 시스템인 rotating detonation engines(RDEs)에서 그 성능을 입증한다. 우리의 접근법은 injector-driven effect와 관련된 물리적으로 의미 있는 discrepancy dynamics를 분리해내면서 sparse measurement로부터 high-fidelity RDE state를 성공적으로 재구성한다. 결과는 복잡한 multi-scale system에서 data assimilation과 system identification을 위한 일반적인 multi-fidelity framework를 보여주며, 해석 가능한 discrepancy dynamics를 제공하는 동시에 빠른 design exploration과 real-time monitoring 및 control을 가능하게 한다. Anonymous code는 다음에서 उपलब्ध하다: anonymous.4open.science/r/Cheap2Rich-4C71.

"},{"id":"63165","en":"BLIPs: Bayesian Learned Interatomic Potentials","ko":"BLIPs: 베이지안 학습 원자간 퍼텐셜","authors":"Dario Coscia, Pim de Haan, Max Welling","pos":"#1513","link":"https://openreview.net/forum?id=ZSZW2diTs5","abs":"

Machine Learning Interatomic Potentials (MLIPs) are becoming a central tool in simulation-based chemistry. However, like most deep learning models, MLIPs struggle to make accurate predictions on out-of-distribution data or when trained in a data-scarce regime, both common scenarios in simulation-based chemistry. Moreover, MLIPs do not provide uncertainty estimates by construction, which are fundamental to guide active learning pipelines and to ensure the accuracy of simulation results compared to quantum calculations. To address this shortcoming, we propose BLIPs: Bayesian Learned Interatomic Potentials. BLIP is a scalable, architecture-agnostic variational Bayesian framework for training or fine-tuning MLIPs, built on an adaptive version of Variational Dropout. BLIP delivers well-calibrated uncertainty estimates and minimal computational overhead for energy and forces prediction at inference time, while integrating seamlessly with (equivariant) message-passing architectures. Empirical results on simulation-based computational chemistry tasks demonstrate improved predictive accuracy with respect to standard MLIPs, and trustworthy uncertainty estimates, especially in data-scarse or heavy out-of-distribution regimes. Moreover, fine-tuning pretrained MLIPs with BLIP yields consistent performance gains and calibrated uncertainties.

","absKo":"

Machine Learning Interatomic Potential(MLIP)은 simulation-based chemistry에서 핵심 도구로 자리잡고 있다. 그러나 대부분의 deep learning model과 마찬가지로, MLIP는 out-of-distribution 데이터에서의 정확한 예측이나 데이터가 부족한 regime에서의 학습에 어려움을 겪는데, 이는 모두 simulation-based chemistry에서 흔한 상황이다. 더욱이 MLIP는 구성상 uncertainty estimate를 제공하지 않으며, 이는 active learning pipeline을 유도하고 quantum calculation과 비교했을 때 simulation 결과의 정확성을 보장하는 데 필수적이다. 이러한 한계를 해결하기 위해 우리는 BLIP: Bayesian Learned Interatomic Potential을 제안한다. BLIP은 Variational Dropout의 적응형 버전에 기반한, 확장 가능하고 architecture-agnostic한 variational Bayesian framework로서 MLIP의 학습 또는 fine-tuning을 지원한다. BLIP은 inference time에 energy 및 force prediction을 위한 잘 보정된 uncertainty estimate와 최소한의 계산 오버헤드를 제공하며, (equivariant) message-passing architecture와도 매끄럽게 통합된다. simulation-based computational chemistry task에서의 실험 결과는 표준 MLIP에 비해 예측 정확도가 향상되고, 특히 data-scarce 또는 심한 out-of-distribution regime에서 신뢰할 수 있는 uncertainty estimate가 제공됨을 보여준다. 또한 BLIP으로 pretrained MLIP를 fine-tuning하면 일관된 성능 향상과 보정된 uncertainty를 얻을 수 있다.

"},{"id":"60501","en":"AutoMS: Multi-Agent Evolutionary Search for Cross-Physics Inverse Microstructure Design","ko":"AutoMS: Cross-Physics Inverse Microstructure Design을 위한 Multi-Agent Evolutionary Search","authors":"Zhenyuan Zhao, Yu Xing, Tianyang Xue, Lingxin Cao, Xin Yan, Lin Lu","pos":"#1514","link":"https://openreview.net/forum?id=zd6V4IqBkd","abs":"

Designing microstructures that satisfy coupled cross-physics objectives is a fundamental challenge in material science. This inverse design problem involves a vast, discontinuous search space where traditional topology optimization is computationally prohibitive, and deep generative models often suffer from \"physical hallucinations,\" lacking the capability to ensure rigorous validity. To address this limitation, we introduce AutoMS, a multi-agent neuro-symbolic framework that reformulates inverse design as an LLM-driven evolutionary search. Unlike methods that treat LLMs merely as interfaces, AutoMS integrates them as \"semantic navigators\" to initialize search spaces and break local optima, while our novel Simulation-Aware Evolutionary Search (SAES) addresses the \"blindness\" of traditional evolutionary strategies. Specifically, SAES utilizes simulation feedback to perform local gradient approximation and directed parameter updates, effectively guiding the search toward physically valid Pareto frontiers. Orchestrating specialized agents (Manager, Parser, Generator, and Simulator), AutoMS achieves a state-of-the-art 83.8\\% success rate on 17 diverse cross-physics tasks, nearly doubling the performance of traditional NSGA-II (43.7\\%) and significantly outperforming ReAct-based LLM baselines (53.3\\%). Furthermore, our hierarchical architecture reduces total execution time by 23.3\\%. AutoMS demonstrates that autonomous agent systems can effectively navigate complex physical landscapes, bridging the gap between semantic design intent and rigorous physical validity.

","absKo":"

서로 결합된 cross-physics objective를 만족하는 microstructure를 설계하는 것은 material science의 근본적인 도전이다. 이 inverse design 문제는 방대하고 불연속적인 search space를 포함하며, 전통적인 topology optimization은 계산적으로 매우 비싸고, deep generative model은 종종 \"physical hallucinations\"에 시달려 엄밀한 유효성을 보장하지 못한다. 이러한 한계를 해결하기 위해 우리는 inverse design을 LLM-driven evolutionary search로 재정식화하는 multi-agent neuro-symbolic framework인 AutoMS를 제안한다. LLM을 단순한 interface로 취급하는 방법과 달리, AutoMS는 이를 search space 초기화와 local optimum 탈출을 위한 \"semantic navigator\"로 통합하며, 우리의 새로운 Simulation-Aware Evolutionary Search (SAES)는 전통적인 evolutionary strategy의 \"blindness\"를 해결한다. 구체적으로 SAES는 simulation feedback을 활용해 local gradient approximation과 directed parameter update를 수행하며, 물리적으로 유효한 Pareto frontier를 향해 search를 효과적으로 유도한다. Manager, Parser, Generator, Simulator로 구성된 specialized agent를 조율하는 AutoMS는 17개의 다양한 cross-physics task에서 83.8\\%의 state-of-the-art 성공률을 달성하여 전통적인 NSGA-II (43.7\\%)의 성능을 거의 두 배로 끌어올리고, ReAct 기반 LLM baseline (53.3\\%)을 크게 능가한다. 또한 우리의 hierarchical architecture는 총 실행 시간을 23.3\\% 줄인다. AutoMS는 autonomous agent system이 복잡한 물리적 지형을 효과적으로 탐색할 수 있음을 보여주며, semantic design intent와 엄밀한 physical validity 사이의 간극을 메운다.

"},{"id":"64523","en":"A Call to Lagrangian Action: Learning Population Mechanics from Temporal Snapshots","ko":"Lagrangian 행동에 대한 요청: 시간적 스냅샷으로부터 개체군 역학 학습","authors":"Vincent Guan, Lazar Atanackovic, Kirill Neklyudov","pos":"#1515","link":"https://openreview.net/forum?id=MjPv4kRu3H","abs":"

The population dynamics of molecules, cells, and organisms are governed by a number of unknown internal and external forces. In the last decade, population dynamics have predominately been modeled with Wasserstein gradient flows. However, since gradient flows minimize free energy, they fail to capture important dynamical properties, such as periodicity. In this work, we propose a change in perspective by considering population dynamics that minimize Wasserstein Lagrangian action, rather than free energy. As our main theoretical contributions, we derive the Hamiltonian equations of motion from the principle of least population-level action and we show that these mechanics encompass classical mechanics, quantum mechanics, and gradient flows. We further leverage the Hamiltonian perspective to propose an algorithm that learns the population mechanics from observed marginals, without specifying the Lagrangian. We demonstrate that by directly learning the population mechanics, our method forecasts and interpolates unseen marginals without a reference process, and outperforms gradient flow and flow matching methods across a wide range of real and simulated experiments.

","absKo":"

분자, 세포, 생물체의 population dynamics는 여러 가지 알려지지 않은 내부 및 외부 힘에 의해 지배된다. 지난 10년간 population dynamics는 주로 Wasserstein gradient flow로 모델링되어 왔다. 그러나 gradient flow는 free energy를 최소화하므로 periodicity와 같은 중요한 동역학적 성질을 포착하지 못한다. 본 연구에서는 free energy가 아니라 Wasserstein Lagrangian action을 최소화하는 population dynamics를 고려함으로써 관점을 전환할 것을 제안한다. 우리의 주요 이론적 기여로는 최소 population-level action의 원리로부터 Hamiltonian 운동 방정식을 유도하고, 이러한 mechanics가 classical mechanics, quantum mechanics, 그리고 gradient flow를 모두 포괄함을 보인다. 또한 Hamiltonian 관점을 활용하여 Lagrangian을 지정하지 않고도 관측된 marginals로부터 population mechanics를 학습하는 algorithm을 제안한다. 우리는 population mechanics를 직접 학습함으로써, 우리의 방법이 reference process 없이 보이지 않는 marginal을 forecast하고 interpolate할 수 있으며, 광범위한 실제 및 시뮬레이션 실험에서 gradient flow와 flow matching 방법을 능가함을 보인다.

"},{"id":"68781","en":"BM^2: Coupled Schrödinger Bridge Matching","ko":"BM^2: 결합 Schrödinger Bridge Matching","authors":"Stefano Peluchetti","pos":"#3906","abs":"A Schrödinger bridge establishes a dynamic transport map between two target distributions via a reference process, simultaneously solving an associated entropic optimal transport problem.\nWe consider the setting where samples from the target distributions are available, and the reference diffusion process admits tractable dynamics.\nWe thus introduce Coupled Bridge Matching (BM$^2$), a simple \\emph{non-iterative} approach for learning Schrödinger bridges with neural networks.\nA preliminary theoretical analysis of the convergence properties of BM$^2$ is carried out, supported by numerical experiments that demonstrate the effectiveness of our proposal.","absKo":"Schrödinger bridge는 reference process를 통해 두 target distribution 사이의 동적 transport map을 구축하며, 동시에 관련된 entropic optimal transport 문제를 푼다.\n우리는 target distribution에서 샘플을 얻을 수 있고, reference diffusion process의 dynamics를 tractable하게 다룰 수 있는 설정을 고려한다.\n이에 따라 우리는 neural network로 Schrödinger bridge를 학습하기 위한 간단한 \\emph{non-iterative} 접근법인 Coupled Bridge Matching(BM$^2$)을 제안한다.\nBM$^2$의 수렴 특성에 대한 예비 이론 분석을 수행하고, 제안 방법의 효과를 보여주는 수치 실험으로 이를 뒷받침한다."},{"id":"65327","en":"Unleashing the Representational Power of Fourier Shapes for Attacking Infrared Object Detection","ko":"적외선 객체 탐지 공격을 위한 Fourier 형상의 표현력 발휘","authors":"Yixing Yong, Jian Wang, Ming Lei, Lijun He, Fan Li","pos":"#1106","link":"https://openreview.net/forum?id=Eio2pEtbg4","abs":"

Infrared object detection is crucial for perception in autonomous driving and surveillance but remains vulnerable to physical adversarial attacks. Unlike in the RGB domain, where attacks rely on color texture, infrared attacks must manipulate thermal signatures, making the geometry shape of heat-blocking materials the primary adversarial information carrier. Current shape-based methods suffer from a fundamental trade-off between representational capability and optimization power, limiting their attack effectiveness. In this work, we overcome this dilemma by introducing learnable Fourier shapes to the infrared domain. We utilize an end-to-end differentiable framework where a compact set of Fourier coefficients, defining the shape boundary, is analytically mapped to a pixel-space mask via the winding number theorem. This enables efficient gradient-based optimization to generate potent shapes that cause human targets to evade detection. Extensive digital and physical experiments provide a comprehensive evaluation and validate our superior performance. Our resulting physical patch achieves striking robustness, successfully evading detectors across diverse distances, angles, poses, and individuals, and achieves over 88% attack success rate at distances greater than 25m (conf.=0.5).Video demo and source code are available in the supplementary material.

","absKo":"

적외선 object detection은 autonomous driving과 surveillance의 perception에 매우 중요하지만, 물리적 adversarial attack에 여전히 취약합니다. 색상 texture에 의존하는 RGB domain과 달리, 적외선 공격은 thermal signature를 조작해야 하므로, heat-blocking material의 geometry shape가 주요 adversarial information carrier가 됩니다. 현재의 shape-based 방법은 representation capability와 optimization power 사이의 근본적인 trade-off를 겪어 공격 효과가 제한됩니다. 본 연구에서는 learnable Fourier shape를 적외선 domain에 도입함으로써 이 딜레마를 극복합니다. 우리는 shape boundary를 정의하는 compact한 Fourier coefficient 집합을 winding number theorem을 통해 pixel-space mask로 해석적으로 매핑하는 end-to-end differentiable framework를 사용합니다. 이를 통해 강력한 shape를 효율적인 gradient-based optimization으로 생성하여 human target이 detection을 회피하도록 만들 수 있습니다. 광범위한 digital 및 physical 실험은 포괄적인 평가를 제공하며 우리의 우수한 성능을 검증합니다. 최종적으로 얻은 physical patch는 인상적인 robustness를 보이며, 다양한 거리, 각도, pose, 그리고 사람에 대해 detector를 성공적으로 회피하고, 25m를 초과하는 거리에서 88% 이상의 attack success rate를 달성합니다(conf.=0.5). Video demo와 source code는 supplementary material에 제공됩니다.

"},{"id":"64037","en":"Structure-Aware Consistency Priors for Shape from Polarization in Complex Media","ko":"복잡 매질에서의 편광 기반 형상 복원을 위한 구조 인식 일관성 사전","authors":"Kaimin Yu, Puyun Wang, Huayang He, Xianyu Wu","pos":"#1113","link":"https://openreview.net/forum?id=RVq8zBMmaJ","abs":"Recovering surface normals from single-view polarization images in complex media remains challenging. This paper focuses on ice as a representative complex medium, where intricate light–matter interactions lead to a nonlinear mapping between polarization observations and surface normals. To address this, a structure-aware polarization prior based on autocorrelation functions is proposed to capture the local spatial consistency of AoLP. Building on this, a dual-branch network (IceSfP) is designed to integrate raw polarization features with priors via cross-modal attention and multi-scale feature fusion, enabling accurate surface normal estimation under complex media conditions. To evaluate the method, the first real-world ice SfP dataset is constructed. Experimental results show that the method outperforms existing approaches across all metrics, achieving a MAE of 16.01$^\\circ$, which is 2.74$^\\circ$ lower than the second-best method. The framework provides a generalizable solution for high-precision geometric perception in complex media.","absKo":"복잡한 매질에서 단일 view polarization image로부터 surface normal을 복원하는 문제는 여전히 어렵다. 이 논문은 대표적인 complex medium인 ice에 초점을 맞추며, 복잡한 light-matter interaction이 polarization observation과 surface normal 사이에 비선형 mapping을 유발한다. 이를 해결하기 위해 autocorrelation function에 기반한 structure-aware polarization prior를 제안하여 AoLP의 지역적 spatial consistency를 포착한다. 이를 바탕으로, dual-branch network(IceSfP)를 설계하여 cross-modal attention과 multi-scale feature fusion을 통해 raw polarization feature와 prior를 통합함으로써 complex media 조건에서 정확한 surface normal estimation을 가능하게 한다. 방법을 평가하기 위해 최초의 real-world ice SfP dataset을 구축했다. 실험 결과는 이 방법이 모든 metric에서 기존 접근법을 능가함을 보여주며, MAE 16.01$^\\circ$를 달성해 두 번째로 좋은 방법보다 2.74$^\\circ$ 낮았다. 이 프레임워크는 complex media에서 고정밀 geometric perception을 위한 일반화 가능한 해법을 제공한다."},{"id":"66334","en":"PCGS: Deblurring 3D Gaussian Splatting with Patch Comparison","ko":"PCGS: 패치 비교를 통한 3D Gaussian Splatting 디블러링","authors":"Yilong Li, Bo Pang, Zhongtao Wang, Mai Su, Yisong Chen, Chengwei Pan, Meng GAI, Fei Zhu, Guoping Wang","pos":"#1202","link":"https://openreview.net/forum?id=4bLAhR3KK7","abs":"

Recent neural methods, such as 3D Gaussian Splatting, have achieved state-of-the-art rendering quality and speed. However, these methods frequently encounter challenges in regions with overlapping Gaussians, leading to blurring and artifacts in the rendered images. We observed that widely used view-space positional gradients are insufficient for handling such circumstances. To address this, we introduce PCGS, a Patch Comparison Gaussian Splatting method to control the densification of corresponding Gaussians adaptively. Specifically, PCGS divides the rendered image into patches and identifies those with significant errors by comparing the loss between the rendered and ground truth images. Additional densification operations are then applied to the Gaussians in these error-prone regions. Furthermore, to prevent over-densification and redundant Gaussians, we design a Gaussian control strategy to regulate the densification process. Specifically, we set a Gaussian number budget that dynamically changes according to the progress of densification, and sample the Gaussians required for each densification step based on their importance scores. Our method results in significantly fewer artifacts and less blur while maintaining a Gaussian count approximately equal to that of 3DGS. Extensive experiments on multiple standard benchmarks demonstrate the superiority of our approach.

","absKo":"

3D Gaussian Splatting과 같은 최근 neural method는 최고 수준의 rendering 품질과 속도를 달성했다. 그러나 이러한 방법들은 Gaussian이 겹치는 영역에서 자주 문제를 겪어, 렌더링된 이미지에 blur와 artifact가 발생한다. 우리는 널리 사용되는 view-space positional gradient가 이러한 상황을 처리하기에 충분하지 않음을 관찰했다. 이를 해결하기 위해, 우리는 대응하는 Gaussian의 densification을 적응적으로 제어하는 Patch Comparison Gaussian Splatting 방법인 PCGS를 도입한다. 구체적으로, PCGS는 렌더링된 이미지를 patch로 나누고, 렌더링 결과와 ground truth 이미지 간의 loss를 비교하여 오차가 큰 patch를 식별한다. 그런 다음 이러한 오류가 발생하기 쉬운 영역의 Gaussian에 추가 densification 작업을 적용한다. 더 나아가, 과도한 densification과 중복 Gaussian을 방지하기 위해, 우리는 densification 과정을 조절하는 Gaussian control strategy를 설계한다. 구체적으로, densification 진행도에 따라 동적으로 변하는 Gaussian 수 예산을 설정하고, 중요도 점수에 기반하여 각 densification step에 필요한 Gaussian을 샘플링한다. 우리의 방법은 3DGS와 거의 같은 Gaussian 수를 유지하면서도 artifact와 blur를 크게 줄인다. 여러 표준 benchmark에서의 광범위한 실험은 우리 접근법의 우수성을 보여준다.

"},{"id":"60798","en":"Rethinking Low-Confidence Pseudo Labels: Influence-Aware Semi-Supervised Fine-Tuning for Hyperspectral Change Detection","ko":"Low-Confidence Pseudo Label 재고: Hyperspectral Change Detection을 위한 Influence-Aware Semi-Supervised Fine-Tuning","authors":"Keyun Zhao, Guangchang Li, Yunpeng Bai, shao jiang, Ying Li","pos":"#1208","link":"https://openreview.net/forum?id=wbhYFWXJQq","abs":"

Hyperspectral image change detection (HSI-CD) suffers from severe annotation scarcity and complex change patterns, which fundamentally limit the effectiveness of directly fine-tuning pre-trained foundation models. Although semi-supervised learning provides a promising direction, existing approaches mainly rely on confidence-based pseudo-label selection, leading to limited data diversity or severe error propagation. In this paper, we propose Influence-Aware Semi-supervised Fine-tuning (IA-SFT), a novel framework that evaluates the influence of pseudo-labels on model decision behavior to identify truly valuable supervision signals. Instead of confidence-based selection, IA-SFT evaluates each low-confidence pseudo-label by measuring its impact on labeled data, enabling reliable filtering of high-value pseudo-labels with minimal noise. To further adapt foundation models to HSI-CD, we design an Adaptive Fusion Change Decoder (AFCD) that jointly models global semantic consistency and local change details. Extensive experiments on three benchmark datasets demonstrate that IA-SFT consistently improves pseudo-label quality and detection performance, achieving superior accuracy compared to state-of-the-art methods. Additional analyses validate the transferability of IA-SFT when integrated into different frameworks in a plug-and-play manner. Code will be released.

","absKo":"

Hyperspectral image change detection(HSI-CD)은 심각한 annotation 부족과 복잡한 변화 패턴으로 인해 pre-trained foundation model을 직접 fine-tuning하는 방식의 효과가 근본적으로 제한된다. semi-supervised learning이 유망한 방향을 제공하지만, 기존 접근은 주로 confidence 기반 pseudo-label 선택에 의존하여 데이터 다양성이 제한되거나 오류 전파가 심각해진다. 본 논문에서는 pseudo-label이 모델의 의사결정 행동에 미치는 영향을 평가하여 진정으로 가치 있는 supervision signal을 식별하는 새로운 framework인 Influence-Aware Semi-supervised Fine-tuning(IA-SFT)을 제안한다. confidence 기반 선택 대신, IA-SFT는 각 low-confidence pseudo-label이 labeled data에 미치는 영향을 측정하여 평가함으로써, 노이즈를 최소화하면서 가치가 높은 pseudo-label을 신뢰성 있게 필터링한다. foundation model을 HSI-CD에 더욱 잘 적응시키기 위해, 우리는 global semantic consistency와 local change detail을 공동으로 모델링하는 Adaptive Fusion Change Decoder(AFCD)를 설계한다. 세 개의 benchmark dataset에 대한 광범위한 실험은 IA-SFT가 pseudo-label 품질과 detection 성능을 일관되게 향상시키며, state-of-the-art 방법보다 우수한 정확도를 달성함을 보여준다. 추가 분석은 IA-SFT를 서로 다른 framework에 plug-and-play 방식으로 통합했을 때의 transferability를 검증한다. 코드는 공개될 예정이다.

"},{"id":"66519","en":"Revisiting Photometric Ambiguity for Accurate Gaussian-Splatting Surface Reconstruction","ko":"정확한 Gaussian-Splatting 표면 재구성을 위한 측광 모호성 재검토","authors":"Jiahe Li, Jiawei Zhang, Xiao Bai, Jin Zheng, Xiaohan Yu, Lin Gu, Gim Hee Lee","pos":"#1209","link":"https://openreview.net/forum?id=2zNijsn2jC","abs":"

Surface reconstruction with differentiable rendering has achieved impressive performance in recent years, yet the pervasive photometric ambiguities have strictly bottlenecked existing approaches. This paper presents AmbiSuR, a framework that explores an intrinsic solution upon Gaussian Splatting for the photometric ambiguity-robust surface reconstruction with high performance. Started by revisiting the foundation, our investigation uncovers two built-in primitive-wise ambiguities in representation, while revealing an intrinsic potential for ambiguity self-indication in Gaussian Splatting. Stemming from these, a photometric disambiguation is first introduced, constraining ill-posed geometry solution for definite surface formation. Then, we propose an ambiguity indication module that unleashes the self-indication potential to identify and further guide correcting underconstrained reconstructions. Extensive experiments demonstrate our superior performance in surface reconstruction compared to existing methods across various challenging scenarios, while excelling in broad compatibility. Our code will be made open-source upon acceptance.

","absKo":"

Differentiable rendering을 이용한 surface reconstruction은 최근 인상적인 성능을 달성했지만, 만연한 photometric ambiguity가 기존 접근법을 엄격히 병목으로 제한해 왔다. 이 논문은 Gaussian Splatting 위에서 intrinsic solution을 탐구하여, photometric ambiguity에 강인하면서도 고성능인 surface reconstruction을 위한 프레임워크 AmbiSuR를 제시한다. 기초를 다시 검토하는 것에서 시작해, 우리의 조사에서는 representation 내에 존재하는 두 가지 built-in primitive-wise ambiguity를 밝혀내는 한편, Gaussian Splatting에 ambiguity self-indication의 intrinsic potential이 존재함을 드러낸다. 이를 바탕으로 먼저 photometric disambiguation을 도입하여, definite surface formation을 위해 ill-posed geometry solution을 제약한다. سپس, ambiguity indication module을 제안하여 self-indication potential을 활용해 underconstrained reconstruction을 식별하고 이를 교정하도록 추가로 유도한다. 광범위한 실험은 다양한 challenging scenario 전반에서 기존 방법들보다 우수한 surface reconstruction 성능을 보이며, 폭넓은 compatibility에서도 뛰어남을 입증한다. 본 코드는 채택 시 open-source로 공개될 예정이다.

"},{"id":"61836","en":"Metric–-Phase Fields: Decoupling Distance and Sign for Thin-Structure Reconstruction from Unoriented Point Clouds","ko":"Metric–-Phase Fields: 방향 없는 Point Cloud에서 얇은 구조 복원을 위한 거리와 부호의 분리","authors":"Jiayi Kong, Xuhui Chen, Chen Zong, Fei Hou, Junhui Hou, Wenping Wang, Ying He","pos":"#1300","link":"https://openreview.net/forum?id=mVsb2MZ7LA","abs":"Neural Signed Distance Functions (SDFs) excel at reconstructing watertight manifolds but fail on thin structures and open boundaries due to strict inside-outside constraints. Conversely, Unsigned Distance Fields (UDFs) accommodate general geometries but suffer from gradient singularities at the zero-level set, hindering optimization and extraction. We introduce Metric–-Phase Fields (MPFs), a decoupled implicit representation that separates metric proximity from topological phase. Given an unoriented point cloud, MPFs learn (i) an unsigned metric field $r$ and (ii) a smooth phase field $\\theta$, for which we derive a bounded phase indicator $P=\\tanh(\\beta\\theta)$ that provides soft inside–outside cues where they are meaningful. We couple the two fields via a gated-metric formulation with a residual phase injection to obtain a signed implicit function with stable near-surface gradients. The phase coefficient $\\beta$ is learnable, allowing MPFs to adaptively control the sharpness of the phase transition and the degree of saturation of the soft sign indicator. Experiments on both synthetic and scanned thin-shell and thin-plate shapes demonstrate that MPFs preserve thin and layered structures more faithfully than recent SDF-based methods, while also enabling more robust training and more reliable surface extraction than UDF-based approaches. We will make the source code publicly available.","absKo":"Neural Signed Distance Functions (SDFs)는 watertight manifold를 재구성하는 데는 뛰어나지만, 엄격한 inside-outside constraint 때문에 thin structures와 open boundaries에서는 실패한다. 반대로 Unsigned Distance Fields (UDFs)는 일반적인 geometry를 수용하지만, zero-level set에서 gradient singularity를 겪어 optimization과 extraction을 어렵게 만든다. 우리는 metric proximity와 topological phase를 분리하는 decoupled implicit representation인 Metric–-Phase Fields (MPFs)를 소개한다. orientation이 없는 point cloud가 주어지면, MPFs는 (i) unsigned metric field $r$과 (ii) smooth phase field $\\theta$를 학습하며, 우리는 의미가 있는 곳에서 soft inside–outside cue를 제공하는 bounded phase indicator $P=\\tanh(\\beta\\theta)$를 유도한다. 우리는 stable near-surface gradient를 갖는 signed implicit function을 얻기 위해 gated-metric formulation과 residual phase injection을 통해 두 field를 결합한다. phase coefficient $\\beta$는 학습 가능하므로, MPFs는 phase transition의 sharpness와 soft sign indicator의 saturation 정도를 적응적으로 제어할 수 있다. synthetic 및 scanned thin-shell, thin-plate shape에 대한 실험은 MPFs가 최근 SDF-based method보다 thin 및 layered structure를 더 충실하게 보존하며, 동시에 UDF-based approach보다 더 견고한 training과 더 신뢰할 수 있는 surface extraction을 가능하게 함을 보여준다. source code는 공개할 예정이다.

"},{"id":"63065","en":"IQA-Spider: Unifying Reasoning, Grounding, and Referring for Multi-Granularity Image Quality Assessment","ko":"IQA-Spider: 다중 세분화 이미지 품질 평가를 위한 추론, 그라운딩, 참조의 통합","authors":"Xinge Peng, Yiting Lu, Xin Li, Zhibo Chen","pos":"#1304","link":"https://openreview.net/forum?id=aOxTMMLYuP","abs":"

We present IQA-Spider, the first image quality assessment (IQA) framework that unifies reasoning, grounding, and referring within a LMM-based system for multi-granularity quality understanding. Existing LMM-based IQA methods typically support only partial perception dimensions, \\egno, quality description and question answering~(\\ieno, reasoning) or pixel-level grounding, largely due to the absence of (i) a unified task-and-data formulation and (ii) effective optimization paradigms for multi-granularity learning. To address these limitations, we formulate a rigorous four-task paradigm covering global and local quality description, pixel-level grounding, and region-level referring. Based on this formulation, we construct a corresponding IQA dataset with a scalable and automatic annotation pipeline, thereby providing a solid foundation for unified multi-granularity learning. To further enable unified perception, we adopts a conflict-free two-stage design that progressively extends textual multi-granularity understanding to pixel-level grounding: (i) the first stage equips the model with fine-grained textual reasoning across multiple IQA tasks, and (ii) the second stage introduces a training-free text-to-point grounding paradigm, which bridges textual semantics and pixel-level perception by mapping token logits to spatial coordinates. Based on these efforts, we achieve IQA-Spider with unified multi-granularity explainable image quality assessment. Extensive experiments across multiple benchmarks demonstrate strong performance, validating the effectiveness and versatility of the proposed formulation and framework. Code and datasets will be released upon acceptance.

","absKo":"

우리는 IQA-Spider를 제안한다. 이는 multi-granularity quality understanding을 위해 LMM-based system 내에서 reasoning, grounding, referring을 통합한 최초의 image quality assessment (IQA) framework이다. 기존의 LMM-based IQA 방법은 일반적으로 quality description과 question answering~(\\ieno, reasoning) 또는 pixel-level grounding과 같은 부분적인 perception dimension만 지원하는데, 이는 주로 (i) task와 data를 통합적으로 정의한 formulation의 부재와 (ii) multi-granularity learning을 위한 효과적인 optimization paradigm의 부재 때문으로 볼 수 있다. 이러한 한계를 해결하기 위해 우리는 global 및 local quality description, pixel-level grounding, region-level referring을 포괄하는 엄격한 four-task paradigm을 정식화한다. 이 정식화를 바탕으로, 확장 가능하고 자동화된 annotation pipeline을 갖춘 대응하는 IQA dataset을 구축하여 통합된 multi-granularity learning의 견고한 기반을 제공한다. 더 나아가 통합된 perception을 가능하게 하기 위해, 우리는 conflict-free two-stage design을 채택하여 textual multi-granularity understanding을 점진적으로 pixel-level grounding으로 확장한다: (i) 첫 번째 stage는 여러 IQA task에 걸친 fine-grained textual reasoning을 모델에 부여하고, (ii) 두 번째 stage는 training-free text-to-point grounding paradigm을 도입하여 token logit을 spatial coordinate에 매핑함으로써 textual semantics와 pixel-level perception을 연결한다. 이러한 노력의 결과로, 우리는 통합된 multi-granularity explainable image quality assessment인 IQA-Spider를 구현한다. 여러 benchmark에 대한 광범위한 실험은 강한 성능을 보여 주며, 제안한 formulation과 framework의 효과와 다양성을 검증한다. 코드는 수락 시 공개될 예정이다.

"},{"id":"65192","en":"Hyperspectral Image Fusion with Spectral-Band and Fusion-Scale Agnosticism","ko":"스펙트럼 밴드 및 융합 스케일 비의존적 초분광 이미지 융합","authors":"Yujie Liang, ZiHan Cao, Liang-Jian Deng, Yang Yang, Malu Zhang","pos":"#1305","link":"https://openreview.net/forum?id=FvRAsWEgCS","abs":"

Current deep learning models for Multispectral and Hyperspectral Image Fusion (MS/HS fusion) are typically designed for fixed spectral bands and spatial scales, which limits their transferability across diverse sensors. To address this, we propose SSA, a universal framework for MS/HS fusion with spectral-band and fusion-scale agnosticism. Specifically, we introduce Matryoshka Kernel (MK), a novel operator that enables a single model to adapt to arbitrary numbers of spectral channels. Meanwhile, we build SSA upon an Implicit Neural Representation (INR) backbone that models the HS signal as a continuous function, enabling reconstruction at arbitrary spatial resolutions. Together, these two forms of agnosticism enable a single MS/HS fusion model that generalizes effectively to unseen sensors and spatial scales. Extensive experiments demonstrate that our single model achieves state-of-the-art performance while generalizing well to unseen sensors and scales, paving the way toward future HS foundation models.

","absKo":"

Multispectral 및 Hyperspectral Image Fusion (MS/HS fusion)을 위한 현재의 deep learning model은 일반적으로 고정된 spectral band와 spatial scale에 맞춰 설계되며, 이로 인해 다양한 sensor 간 transferability가 제한된다. 이를 해결하기 위해, 우리는 spectral-band 및 fusion-scale에 대해 agnostic한 MS/HS fusion을 위한 universal framework인 SSA를 제안한다. 구체적으로, 우리는 하나의 model이 임의 개수의 spectral channel에 적응할 수 있게 하는 새로운 operator인 Matryoshka Kernel (MK)을 도입한다. 한편, 우리는 SSA를 HS signal을 연속 함수로 모델링하는 Implicit Neural Representation (INR) backbone 위에 구축하여, 임의의 spatial resolution에서 reconstruction이 가능하도록 한다. 이 두 가지 형태의 agnosticism을 결합하면, 보지 못한 sensor와 spatial scale에 효과적으로 일반화하는 단일 MS/HS fusion model을 얻을 수 있다. 광범위한 실험은 우리의 단일 model이 state-of-the-art 성능을 달성함과 동시에 unseen sensor와 scale에 잘 일반화하여, 향후 HS foundation model로 가는 길을 연다는 점을 보여준다.

"},{"id":"63372","en":"GSFixer: Improving 3D Gaussian Splatting with Reference-Guided Video Diffusion Priors","ko":"GSFixer: 참조 유도 비디오 Diffusion 사전지식을 통한 3D Gaussian Splatting 개선","authors":"Xingyilang Yin, Qi Zhang, Jiahao Chang, Ying Feng, Qingnan Fan, Xi Yang, Chi-Man Pun, Huaqi Zhang, Xiaodong Cun","pos":"#1306","link":"https://openreview.net/forum?id=Xhctwvm3O2","abs":"

Reconstructing 3D scenes using 3D Gaussian Splatting (3DGS) from sparse views is an ill-posed problem due to insufficient information, often resulting in noticeable artifacts. While recent approaches have sought to leverage generative priors to complete information for under-constrained regions, they struggle to generate content that remains consistent with input observations. To address this challenge, we propose GSFixer, a novel framework designed to improve the quality of 3DGS representations reconstructed from sparse inputs. The core of our approach is the reference-guided video restoration model, built upon a DiT-based video diffusion model trained on paired artifact 3DGS renders and clean frames with additional reference-based conditions. Considering the input sparse views as references, our model integrates both 2D semantic and 3D geometric features of reference views extracted from the visual geometry foundation model, enhancing the semantic coherence and 3D consistency when fixing artifact novel views. Furthermore, we introduce a reference-guided trajectory sampling strategy that ensures both angular coverage and view quality, further enhancing reconstruction fidelity. Considering the lack of suitable benchmarks for 3DGS artifact restoration evaluation, we present DL3DV-Res which contains artifact frames rendered using low-quality 3DGS. Extensive experiments demonstrate our GSFixer outperforms current state-of-the-art methods in 3DGS artifact restoration and sparse-view 3D reconstruction. The Project will be made public.

","absKo":"

sparse view에서 3D Gaussian Splatting(3DGS)을 사용해 3D scene을 재구성하는 것은 정보가 부족하기 때문에 ill-posed problem이며, 종종 눈에 띄는 artifact를 초래한다. 최근 접근법들은 under-constrained region의 정보를 보완하기 위해 generative prior를 활용하려 했지만, 입력 관측과 일관된 내용을 생성하는 데에는 여전히 어려움을 겪는다. 이 문제를 해결하기 위해 우리는 sparse input으로부터 재구성된 3DGS representation의 품질을 향상시키기 위한 새로운 framework인 GSFixer를 제안한다. 우리의 접근법의 핵심은 artifact 3DGS render와 clean frame, 그리고 추가적인 reference-based condition으로 학습된 DiT 기반 video diffusion model 위에 구축된 reference-guided video restoration model이다. 입력 sparse view를 reference로 간주하여, 우리의 model은 visual geometry foundation model에서 추출한 reference view의 2D semantic feature와 3D geometric feature를 모두 통합함으로써 artifact novel view를 수정할 때 semantic coherence와 3D consistency를 향상시킨다. 또한 우리는 angular coverage와 view quality를 모두 보장하는 reference-guided trajectory sampling strategy를 도입하여 reconstruction fidelity를 더욱 높인다. 3DGS artifact restoration 평가에 적합한 benchmark가 부족한 점을 고려하여, 우리는 low-quality 3DGS로 렌더링한 artifact frame을 포함하는 DL3DV-Res를 제시한다. 광범위한 실험은 우리의 GSFixer가 3DGS artifact restoration과 sparse-view 3D reconstruction 모두에서 현재 state-of-the-art method를 능가함을 보여준다. Project는 공개될 예정이다.

"},{"id":"61447","en":"GADA: Geometry-Aware Deformable Aggregation for Image-Based Gaussian Splatting","ko":"GADA: 이미지 기반 Gaussian Splatting을 위한 기하 인식 변형 가능 Aggregation","authors":"Siwoo Lim, Sunjae Yoon, Gwanhyeong Koo, Chang D. Yoo","pos":"#1307","link":"https://openreview.net/forum?id=pyQ9gFqlec","abs":"

Gaussian Splatting has achieved significant improvements by incorporating warping-based techniques. These approaches enhance synthesis quality by warping images from source views into the target viewpoint to compensate for missing or residual pixels. However, such methods suffer from pixel-level inaccuracies due to uncertain geometry. This uncertainty leads to spatial misalignments in the warped images, which disrupt residual learning used in warping-based methods and fundamentally limit the gains of correction, particularly on thin structures and high-frequency details. Driven by our insight that useful visual cues are not lost but locally preserved under slight displacement, we propose Geometry-Aware Deformable Aggregation (GADA). This method introduces an iterative refinement module with deformable offsets to actively correct spatial misalignments and recover these displaced cues. Furthermore, to address the limitations of standard pipelines where visibility checks (i.e., thresholding) often discard valid pixels and multi-view warped image fusion relies on naive mean aggregation, our module is coupled with an implicit confidence weighting mechanism that selectively suppresses unreliable evidence. Consequently, our approach outperforms prior warping-based Gaussian Splatting, preserving high-frequency quality while achieving 2.13times faster FPS.

","absKo":"Gaussian Splatting은 warping-based 기법을 통합함으로써 큰 개선을 이뤘다. 이러한 방법은 source view의 이미지를 target viewpoint로 warp하여 누락되거나 잔여 픽셀을 보완함으로써 합성 품질을 향상시킨다. 그러나 이런 방법은 불확실한 geometry로 인해 pixel-level 부정확성에 시달린다. 이 불확실성은 warped image에서 spatial misalignment를 유발하며, 이는 warping-based method에서 사용되는 residual learning을 방해하고, 특히 가느다란 구조와 high-frequency detail에서 보정의 이득을 근본적으로 제한한다. 우리는 유용한 visual cue가 사라지는 것이 아니라 작은 displacement 하에서도 국소적으로 보존된다는 통찰에 기반해 Geometry-Aware Deformable Aggregation(GADA)을 제안한다. 이 방법은 deformable offset을 갖는 iterative refinement module을 도입하여 spatial misalignment를 적극적으로 교정하고, 이렇게 이동된 cue를 복원한다. 또한 visibility check(즉, thresholding)가 종종 유효한 픽셀을 버리고, multi-view warped image fusion이 단순한 mean aggregation에 의존하는 표준 pipeline의 한계를 해결하기 위해, 우리 모듈은 신뢰할 수 없는 evidence를 선택적으로 억제하는 implicit confidence weighting mechanism과 결합된다. 그 결과, 우리의 방법은 이전의 warping-based Gaussian Splatting을 능가하며, 2.13배 더 빠른 FPS를 달성하면서도 high-frequency quality를 보존한다."},{"id":"65518","en":"Formally Exploring Visual Anomaly Detection Evaluation Metrics","ko":"시각 이상 탐지 평가 지표의 형식적 탐구","authors":"Nasar Iqbal, Dennis Wagner, Philipp Liznerski, Nabeel Hussain Syed, Sophie Fellenz, Niki Martinel, Marius Kloft","pos":"#1309","link":"https://openreview.net/forum?id=Cotj8tWdYp","abs":"

Inaccurate Visual Anomaly Detection (VAD) can lead to critical failures in safety-sensitive domains, including autonomous navigation and industrial surveillance. With the increasing abundance and rapid proliferation of VAD algorithms, their reliable evaluation has become increasingly important and challenging. Commonly used evaluation metrics often fail to capture practically relevant aspects of model behavior, yielding inconsistent or misleading assessments by overlooking errors such as redundant detections and the spatial distribution of false positives. In this paper, we formalize the requirements for VAD evaluation by introducing a set of axiomatic, verifiable properties that an evaluation metric should satisfy. Through a systematic analysis of state-of-the-art evaluation methods, we show that none satisfies all proposed properties. To address this gap, we introduce SAAM-ALARM, a novel evaluation metric that satisfies these properties. Our results show that SAAM-ALARM provides a more nuanced and theoretically sound assessment, establishing a stronger standard for performance benchmarking in VAD.

","absKo":"

부정확한 Visual Anomaly Detection(VAD)은 자율 주행 및 산업 감시를 포함한 안전 민감 도메인에서 치명적인 실패로 이어질 수 있다. VAD 알고리즘의 수가 증가하고 빠르게 확산됨에 따라, 이들의 신뢰할 수 있는 평가는 점점 더 중요하면서도 어려운 과제가 되었다. 일반적으로 사용되는 평가 지표는 모델 행동의 실제적으로 중요한 측면을 포착하지 못하는 경우가 많아, 중복 탐지나 false positive의 공간적 분포와 같은 오류를 간과함으로써 일관성이 없거나 오해를 불러일으키는 평가를 낳는다. 본 논문에서는 평가 지표가 만족해야 할 공리적이고 검증 가능한 성질들을 도입함으로써 VAD 평가의 요구 조건을 형식화한다. 최신 평가 방법들에 대한 체계적인 분석을 통해, 제안한 모든 성질을 만족하는 방법은 없음을 보인다. 이러한 공백을 메우기 위해, 우리는 이 성질들을 만족하는 새로운 평가 지표 SAAM-ALARM을 제안한다. 우리의 결과는 SAAM-ALARM이 더 세밀하고 이론적으로 타당한 평가를 제공하며, VAD의 성능 벤치마킹을 위한 더 강한 기준을 확립함을 보여준다.

"},{"id":"65912","en":"DuRP: Dual-Stage Physics-Embedded Learning for Joint Radiance and Polarization Restoration","ko":"DuRP: 결합 복사 및 편광 복원을 위한 이중 단계 물리 임베딩 학습","authors":"Zhenshuo Yang, Qian He, Zhiyuan Liu, Baojie Fan, Jiandong Tian","pos":"#1312","link":"https://openreview.net/forum?id=8yZulpu1TF","abs":"

Polarization information is valuable for many computer vision applications. However, in hazy environments, polarization information is severely attenuated due to the degradation of captured polarized images. Existing dehazing methods struggle to effectively restore polarization information, as single-image methods are unaware of polarization, and polarization-based methods are constrained by the traditional polarization models. These deficiencies lead to inaccurate polarimetric signatures and physical inconsistencies in scattering environments. To overcome these limitations and achieve the joint restoration of scene radiance and polarization information, we propose DuRP, a dual-stage physics-embedded learning framework. Specifically, we derive generalized polarization physics models that relax the ideal assumptions of traditional theory to provide a more precise foundation for the joint restoration of polarimetric and amplitude information. We then design a dual-stage neural network to estimate latent physical parameters through differentiable operators, ensuring that both the polarimetric state and radiance are accurately recovered. Experimental results show that DuRP achieves state-of-the-art performance in joint restoration and significantly enhances polarization-based vision tasks.

","absKo":"

Polarization information은 많은 computer vision application에서 가치가 크다. 그러나 haze가 짙은 환경에서는 촬영된 polarized image의 degradation으로 인해 polarization information이 심하게 감쇠된다. 기존 dehazing method는 polarization information을 효과적으로 복원하는 데 어려움을 겪는데, single-image method는 polarization을 고려하지 못하고 polarization-based method는 전통적인 polarization model에 제약을 받기 때문이다. 이러한 결함은 scattering environment에서 부정확한 polarimetric signature와 물리적 불일치를 초래한다. 이러한 한계를 극복하고 scene radiance와 polarization information의 joint restoration을 달성하기 위해, 우리는 DuRP라는 dual-stage physics-embedded learning framework를 제안한다. 구체적으로, 우리는 전통 이론의 ideal assumption을 완화하는 generalized polarization physics model을 유도하여 polarimetric information과 amplitude information의 joint restoration을 위한 더 정밀한 기반을 제공한다. 이어서 differentiable operator를 통해 latent physical parameter를 추정하는 dual-stage neural network를 설계하여, polarimetric state와 radiance가 모두 정확하게 복원되도록 한다. 실험 결과 DuRP가 joint restoration에서 state-of-the-art performance를 달성하고 polarization-based vision task를 크게 향상시킴을 보여준다.

"},{"id":"64050","en":"Diffuse to Detect: Bi-Level Sample Rebalancing with Pseudo-Label Diffusion for Point-Supervised Infrared Small-Target Detection","ko":"탐지를 위한 확산: 점 지도 적외선 소형 표적 탐지를 위한 의사 레이블 Diffusion 기반 이중 수준 샘플 재조정","authors":"Zhu Liu, Yuanhang Yao, Ping Qian, Zihang Chen, Risheng Liu","pos":"#1313","link":"https://openreview.net/forum?id=RNn2wtNeoK","abs":"

Point supervision has become a scalable solution to address dense annotation for infrared small target detection, but its performance is limited by two coupled bottlenecks: unstable pseudo-label evolution in cluttered, low-contrast infrared imagery and severe sample-distribution imbalance. In this paper, we present a more adaptive and stable framework to address these issues. Leveraging the intrinsic consistency between thermal radiation patterns and heat diffusion, we propose a physics-induced annotation strategy that expands single-point labels into reliable pseudo-masks. To further enhance supervision and alleviate sample imbalance, we develop a bi-level dual-update framework that jointly optimizes detector weights, sample weights, and diffusion parameters. A meta-classifier dynamically predicts sample-wise loss weights, while a differentiable diffusion module refines pseudo-labels with detection feedback, enabling adaptive interaction between training and hyperparameter optimization. Extensive experiments across multiple datasets demonstrate five-fold annotation acceleration, superior detection accuracy, and comparable performance with 30\\% of the training data, validating the efficiency and practicality of our approach.

","absKo":"Point supervision은 infrared small target detection에서 dense annotation을 대체할 수 있는 확장 가능한 해결책이 되었지만, clutter가 심하고 contrast가 낮은 infrared image에서의 불안정한 pseudo-label 진화와 심각한 sample distribution imbalance라는 두 가지 결합된 병목으로 인해 성능이 제한된다. 본 논문에서는 이러한 문제를 해결하기 위한 더 적응적이고 안정적인 framework를 제시한다. thermal radiation pattern과 heat diffusion 사이의 본질적인 일관성을 활용하여, 우리는 단일 point label을 신뢰할 수 있는 pseudo-mask로 확장하는 physics-induced annotation strategy를 제안한다. supervision을 더욱 강화하고 sample imbalance를 완화하기 위해, 우리는 detector weight, sample weight, diffusion parameter를 공동으로 최적화하는 bi-level dual-update framework를 개발한다. meta-classifier는 sample-wise loss weight를 동적으로 예측하고, differentiable diffusion module은 detection feedback으로 pseudo-label을 정제하여 training과 hyperparameter optimization 사이의 적응적 상호작용을 가능하게 한다. 여러 dataset에 걸친 광범위한 실험은 annotation을 5배 더 빠르게 만들고, 더 우수한 detection accuracy를 보이며, training data의 30\\%만 사용해도 동등한 성능을 달성함을 보여주어, 제안 방법의 효율성과 실용성을 입증한다.

"},{"id":"64716","en":"3DGS$^2$-TR: A Scalable Second-Order Trust-Region Method for 3D Gaussian Splatting","ko":"3DGS$^2$-TR: 3D Gaussian Splatting을 위한 확장 가능한 2차 Trust-Region 방법","authors":"Roger Hsiao, Yuchen Fang, Xiangru Huang, Ruilong Li, Hesam Rabeti, Zan Gojcic, Javad Lavaei, James Demmel, Sophia Shao","pos":"#1406","link":"https://openreview.net/forum?id=Kaliw0BQL9","abs":"We propose 3DGS$^2$-TR, a second-order optimizer for accelerating the scene training problem in 3D Gaussian Splatting (3DGS). Unlike existing second-order approaches that rely on explicit or dense curvature representations, such as 3DGS-LM (Höllein et al., 2025) or 3DGS2 (Lan et al., 2025), our method approximates curvature using only the diagonal of the Hessian matrix, estimated efficiently via Hutchinson’s method. Our approach is fully matrix-free and has the same complexity as ADAM (Kingma, 2024), $O(n)$ in both computation and memory costs. To ensure stable optimization in the presence of strong nonlinearity in the 3DGS rasterization process, we introduce a parameter-wise trust-region technique based on the squared Hellinger distance, regularizing updates to Gaussian parameters. Under identical parameter initialization and without densification, 3DGS$^2$-TR is able to achieve better reconstruction quality on standard datasets, using 50\\% fewer training iterations compared to ADAM, while incurring less than 1GB of peak GPU memory overhead (17\\% more than ADAM and 85\\% less than 3DGS-LM), enabling scalability to very large scenes and potentially to distributed training settings.","absKo":"3D Gaussian Splatting (3DGS)에서의 scene training 문제를 가속하기 위한 second-order optimizer인 3DGS$^2$-TR을 제안한다. 기존의 3DGS-LM (Höllein et al., 2025)이나 3DGS2 (Lan et al., 2025)와 같이 명시적이거나 dense한 curvature representation에 의존하는 second-order 접근법과 달리, 우리의 방법은 Hutchinson’s method를 통해 효율적으로 추정한 Hessian matrix의 diagonal만을 사용해 curvature를 근사한다. 이 접근법은 완전히 matrix-free이며, 계산 및 메모리 비용 모두에서 ADAM (Kingma, 2024)과 동일한 복잡도인 $O(n)$을 갖는다. 3DGS rasterization 과정의 강한 비선형성 하에서 안정적인 optimization을 보장하기 위해, 우리는 squared Hellinger distance에 기반한 parameter-wise trust-region 기법을 도입하여 Gaussian parameter에 대한 update를 regularize한다. 동일한 parameter initialization과 densification 없이, 3DGS$^2$-TR은 표준 dataset에서 ADAM보다 50\\% 적은 training iteration을 사용하면서 더 나은 reconstruction quality를 달성할 수 있으며, peak GPU memory overhead는 1GB 미만이다(ADAM보다 17\\% 더 많고 3DGS-LM보다 85\\% 적음). 이는 매우 큰 scene뿐 아니라 잠재적으로 distributed training setting으로의 확장도 가능하게 한다."},{"id":"64636","en":"A Hypertoroidal Covering for Perfect Color Equivariance","ko":"완벽한 색상 등변성을 위한 Hypertoroidal Covering","authors":"Yulong Yang, Zhikun Xu, Yaojun Li, Christine Allen-Blanchette","pos":"#1407","link":"https://openreview.net/forum?id=LQNrpOXqoB","abs":"

When the color distribution of input images changes at inference, the performance of conventional neural network architectures drops considerably. A few researchers have begun to incorporate prior knowledge of color geometry in neural network design. These color equivariant architectures have modeled hue variation with 2D rotations, and saturation and luminance transformations as 1D translations. While this approach improves neural network robustness to color variations in a number of contexts, we find that approximating saturation and luminance (interval valued quantities) as 1D translations introduces appreciable artifacts. In this paper, we introduce a color equivariant architecture that is truly equivariant. Instead of approximating the interval with the real line, we lift values on the interval to values on the circle (a double-cover) and build equivariant representations there. Our approach resolves the approximation artifacts of previous methods, improves interpretability and generalizability, and achieves better predictive performance than conventional and equivariant baselines on tasks such as fine-grained classification and medical imaging tasks. Going beyond the context of color, we show that our proposed lifting can also extend to geometric transformations such as scale.

","absKo":"

입력 이미지의 색상 분포가 inference 시점에 변하면, 기존 neural network architecture의 성능은 크게 저하된다. 일부 연구자들은 neural network 설계에 color geometry에 대한 prior knowledge를 통합하기 시작했다. 이러한 color equivariant architecture는 hue variation을 2D rotation으로, saturation과 luminance transformation을 1D translation으로 모델링해 왔다. 이 접근은 여러 맥락에서 색상 변화에 대한 neural network의 robustness를 향상시키지만, saturation과 luminance(구간 값)를 1D translation으로 근사하면 상당한 artifact가 발생한다는 것을 우리는 발견했다. 본 논문에서는 진정한 의미의 equivariant한 color equivariant architecture를 제안한다. 구간을 real line으로 근사하는 대신, 우리는 구간 위의 값을 circle(double-cover)로 올려 그 위에서 equivariant representation을 구축한다. 우리의 접근은 기존 방법의 근사 artifact를 해소하고, interpretability와 generalizability를 개선하며, fine-grained classification과 medical imaging task에서 기존 및 equivariant baseline보다 더 나은 예측 성능을 달성한다. 색상 맥락을 넘어, 우리가 제안한 lifting이 scale과 같은 geometric transformation에도 확장될 수 있음을 보인다.

"},{"id":"62601","en":"Affine-Equivariant Kernel Space Encoding for NeRF Editing","ko":"NeRF Editing을 위한 Affine-Equivariant Kernel Space Encoding","authors":"Mikołaj Zieliński, Krzysztof Byrski, Tomasz Szczepanik, Dominik Belter, Przemysław Spurek","pos":"#1408","link":"https://openreview.net/forum?id=fAj3MJghc0","abs":"

Neural scene representations achieve high-fidelity rendering by encoding 3D scenes as continuous functions, but their latent spaces are typically implicit and globally entangled, making localized editing and physically grounded manipulation difficult. While several works introduce explicit control structures or point-based latent representations to improve editability, these approaches often suffer from limited locality, sensitivity to deformations, or visual artifacts. In this paper, we introduce Affine-Equivariant Kernel Space Encoding (EKS), a spatial encoding for neural radiance fields that provides localized, deformation-aware feature representations. Instead of querying latent features directly at discrete points or grid vertices, our encoding aggregates features through a field of anisotropic Gaussian kernels, each defining a localized region of influence. This kernel-based formulation enables stable feature interpolation under spatial transformations while preserving continuity and high reconstruction quality. To preserve detail without sacrificing editability, we further propose a training-time feature distillation mechanism that transfers information from multi-resolution hash grid encodings into the kernel field, yielding a compact and fully grid-free representation at inference. This enables intuitive, localized scene editing directly via Gaussian kernels without retraining, while maintaining high-quality rendering.

","absKo":"

Neural scene representations는 3D 장면을 연속 함수로 인코딩함으로써 고충실도 렌더링을 달성하지만, 잠재 공간은 보통 암묵적이고 전역적으로 얽혀 있어 국소 편집과 물리적으로 타당한 조작이 어렵다. 여러 연구가 편집성을 높이기 위해 명시적 제어 구조나 point-based latent representations를 도입했지만, 이러한 접근은 종종 제한된 국소성, 변형에 대한 민감성, 또는 시각적 아티팩트 문제를 겪는다. 본 논문에서는 neural radiance fields를 위한 공간 인코딩인 Affine-Equivariant Kernel Space Encoding (EKS)을 소개하며, 이는 국소적이고 변형을 인식하는(feature representations) 특징 표현을 제공한다. 이 인코딩은 잠재 특징을 이산 점이나 grid vertex에서 직접 질의하는 대신, 각자가 국소적인 영향 영역을 정의하는 anisotropic Gaussian kernel들의 장(field)을 통해 특징을 집계한다. 이러한 kernel 기반 정식화는 공간 변환 하에서도 안정적인 feature interpolation을 가능하게 하면서 연속성과 높은 재구성 품질을 보존한다. 세부 정보를 유지하면서도 편집성을 희생하지 않기 위해, 우리는 추가로 multi-resolution hash grid encoding의 정보를 kernel field로 전이하는 training-time feature distillation 메커니즘을 제안하며, 이를 통해 추론 시에는 작고 완전한 grid-free 표현을 얻는다. 이는 재학습 없이 Gaussian kernel을 통해 직관적이고 국소적인 장면 편집을 가능하게 하면서도 고품질 렌더링을 유지한다.

"},{"id":"63164","en":"Alleviating Sparse Rewards by Modeling Step-Wise and Long-Term Sampling Effects in Flow-Based GRPO","ko":"Flow 기반 GRPO에서 단계별 및 장기 샘플링 효과 모델링을 통한 희소 보상 완화","authors":"Yunze Tong, Mushui Liu, Canyu Zhao, Wanggui He, Shiyi Zhang, Peng Zhang, Hongwei Zhang, Jinlong Liu, Hao Jiang","pos":"#1409","link":"https://openreview.net/forum?id=ZSt77efXHI","abs":"

Deploying GRPO on Flow Matching models has proven effective for text-to-image generation. However, existing paradigms typically propagate an outcome-based reward to all preceding denoising steps without distinguishing the local effect of each step. Moreover, current group-wise ranking mainly compares trajectories at matched timesteps and ignores within-trajectory dependencies, where certain early denoising actions can affect later states via delayed, implicit interactions. We propose TurningPoint-GRPO (TP-GRPO), a GRPO framework that alleviates step-wise reward sparsity and explicitly models long-term effects within the denoising trajectory. TP-GRPO makes two key innovations: (i) it replaces outcome-based rewards with step-level incremental rewards, providing a dense, step-aware learning signal that better isolates each denoising action’s ``pure\" effect, and (ii) it identifies turning points—steps that flip the local reward trend and make subsequent reward evolution consistent with the overall trajectory trend—and assigns these actions an aggregated long-term reward to capture their delayed impact. Turning points are detected solely via sign changes in incremental rewards, making TP-GRPO efficient and hyperparameter-free. Extensive experiments also demonstrate that TP-GRPO exploits reward signals more effectively and consistently improves generation.

","absKo":"

Flow Matching model에 GRPO를 적용하는 것은 text-to-image generation에서 효과적임이 입증되었다. 그러나 기존 paradigm은 일반적으로 각 step의 국소적 효과를 구분하지 않고 outcome-based reward를 모든 이전 denoising step에 전파한다. 더욱이 현재의 group-wise ranking은 주로 일치하는 timestep에서 trajectory를 비교하며, 지연되고 암묵적인 상호작용을 통해 일부 초기 denoising action이 이후 state에 영향을 미칠 수 있는 trajectory 내부 의존성은 무시한다. 우리는 step-wise reward sparsity를 완화하고 denoising trajectory 내의 장기 효과를 명시적으로 모델링하는 GRPO framework인 TurningPoint-GRPO(TP-GRPO)를 제안한다. TP-GRPO는 두 가지 핵심 혁신을 도입한다: (i) outcome-based reward를 step-level incremental reward로 대체하여 각 denoising action의 ``순수한'' 효과를 더 잘 분리하는 촘촘하고 step-aware한 학습 신호를 제공하고, (ii) local reward trend를 뒤집어 subsequent reward evolution이 전체 trajectory trend와 일관되게 만드는 step인 turning point를 식별하여, 그 지연된 영향을 포착하기 위해 이러한 action에 누적된 장기 reward를 할당한다. Turning point는 incremental reward의 부호 변화만으로 감지되므로, TP-GRPO는 효율적이고 hyperparameter가 필요 없다. 광범위한 실험은 또한 TP-GRPO가 reward signal을 더 효과적으로 활용하며 생성 성능을 일관되게 향상시킴을 보여준다.

"},{"id":"64863","en":"Beyond Heuristics: Learnable Density Control for 3D Gaussian Splatting","ko":"휴리스틱을 넘어서: 3D Gaussian Splatting을 위한 학습 가능한 Density Control","authors":"Zhenhua Ning, Xin Li, Jun Yu, Guangming Lu, Yaowei Wang, Wenjie Pei","pos":"#1410","link":"https://openreview.net/forum?id=J3uM4v9vJv","abs":"While 3D Gaussian Splatting (3DGS) has demonstrated impressive real-time rendering performance, its efficacy remains constrained by a reliance on heuristic density control. Despite numerous refinements to these handcrafted rules, such methods inherently lack the flexibility to adapt to diverse scenes with complex geometries. In this paper, we propose a paradigm shift for density control from rigid heuristics to fully learnable policies. Specifically, we introduce LeGS, a framework that reformulates density control as a parameterized policy network optimized via Reinforcement Learning (RL). Central to our approach is the tailored effective reward function grounded in sensitivity analysis, which precisely quantifies the marginal contribution of individual Gaussians to reconstruction quality. To maintain computational tractability, we derive a closed-form solution that reduces the complexity of reward calculation from $O(N^2)$ to $O(N)$. Extensive experiments on the Mip-NeRF 360, Tanks \\& Temples, and Deep Blending datasets demonstrate that LeGS significantly outperforms state-of-the-art methods, striking a superior balance between reconstruction quality and efficiency.","absKo":"3D Gaussian Splatting(3DGS)은 인상적인 실시간 렌더링 성능을 보여주었지만, 그 효율성은 휴리스틱 density control에 의존한다는 제약을 여전히 받는다. 이러한 수작업 규칙을 수없이 개선했음에도 불구하고, 이들 방법은 본질적으로 복잡한 기하 구조를 가진 다양한 장면에 적응하는 유연성이 부족하다. 본 논문에서는 rigid heuristic에서 fully learnable policy로 density control의 패러다임 전환을 제안한다. 구체적으로, 우리는 density control을 Reinforcement Learning(RL)을 통해 최적화되는 parameterized policy network으로 재정식화한 LeGS를 소개한다. 이 접근의 핵심은 sensitivity analysis에 기반한 맞춤형 effective reward function으로, 개별 Gaussian이 reconstruction quality에 기여하는 한계 기여를 정밀하게 정량화한다. 계산 가능성을 유지하기 위해 우리는 reward 계산의 복잡도를 $O(N^2)$에서 $O(N)$으로 낮추는 closed-form 해를 유도한다. Mip-NeRF 360, Tanks \\& Temples, Deep Blending 데이터셋에서의 광범위한 실험은 LeGS가 state-of-the-art 방법을 유의미하게 능가하며, reconstruction quality와 efficiency 사이에서 더 우수한 균형을 이룸을 보여준다."},{"id":"66532","en":"Cello: A Universal Cell-wise Feature Aggregation framework for Reliable Pathology Images Analysis","ko":"Cello: 신뢰할 수 있는 병리 이미지 분석을 위한 범용 세포별 특징 집계 프레임워크","authors":"Hengrui Lou, Weihan Li, Jiazhen Yang, Lingxiang Jia, Shengxuming Zhang, Linyun Zhou, Xiuming Zhang, Zhenyang Wang, Mingli Song, Zunlei Feng","pos":"#1412","link":"https://openreview.net/forum?id=2qc3SXdpns","abs":"

Computational pathology has made progress in diagnosis and prognosis prediction from whole slide images (WSIs), yet pipelines still rely on patch-level feature extraction and aggregation, departing from the cell-centric reasoning used by pathologists. This gap limits sensitivity to micro-lesions and subtle changes, and current methods rarely provide a unified solution that supports both local and global tasks with trustworthy evidence. We propose Cello, a universal cell-wise feature aggregation framework for reliable pathology image analysis. Cello integrates cell-level representations into WSI modeling via protein-signal–supervised cell-wise learning, preserving fine-grained cellular cues under gigapixel constraints. For local tasks, Cello introduces a flexible prototype-based contrastive module for scalable, task-adaptive representation learning. For global tasks, Cello adopts a weakly supervised gated aggregation that can widely leverage WSI labels. Finally, a cell–local–global decision-route consistency objective dynamically aggregates cellular evidence and aligns local predictions with global outcomes, improving reliability and faithfulness. Trained with only hundreds to thousands of samples, Cello achieves performance gains of 3.0%~7.6% and outperforms SOTA pathology foundation models pretrained on tens of thousands of samples. Code is available at https://anonymous.4open.science/r/Cello.

","absKo":"

Computational pathology는 whole slide image(WSI)로부터의 진단 및 예후 예측에서 진전을 이루었지만, 파이프라인은 여전히 patch-level feature extraction과 aggregation에 의존하여, 병리학자들이 사용하는 cell-centric reasoning과는 거리가 있다. 이러한 간극은 micro-lesion과 미세한 변화에 대한 민감도를 제한하며, 현재 방법들은 신뢰할 수 있는 evidence를 바탕으로 local task와 global task를 모두 지원하는 통합 솔루션을 거의 제공하지 못한다. 우리는 신뢰할 수 있는 pathology image analysis를 위한 universal cell-wise feature aggregation framework인 Cello를 제안한다. Cello는 protein-signal–supervised cell-wise learning을 통해 cell-level representation을 WSI modeling에 통합하여, gigapixel 제약 하에서도 세밀한 cellular cue를 보존한다. local task를 위해 Cello는 확장 가능하고 task-adaptive한 representation learning을 위한 유연한 prototype-based contrastive module을 도입한다. global task를 위해 Cello는 WSI label을 폭넓게 활용할 수 있는 weakly supervised gated aggregation을 채택한다. 마지막으로, cell–local–global decision-route consistency objective는 cellular evidence를 동적으로 aggregate하고 local prediction을 global outcome과 정렬하여 신뢰성과 faithful성을 향상시킨다. 수백에서 수천 개의 샘플만으로 학습한 Cello는 3.0%~7.6%의 성능 향상을 달성하며, 수만 개의 샘플로 pretrain된 SOTA pathology foundation model을 능가한다. code는 https://anonymous.4open.science/r/Cello 에서 उपलब्ध하다.

"},{"id":"65016","en":"Counterfactual Occlusion-Aware Learning via Visibility Intervention for LiDAR Anomaly Detection","ko":"LiDAR 이상 탐지를 위한 Visibility Intervention을 통한 Counterfactual Occlusion-Aware Learning","authors":"Longyu Yang, Jun Liu, Yap-Peng Tan, Fumin Shen, Heng Tao Shen, Xiaofeng Zhu, Ping Hu","pos":"#1413","link":"https://openreview.net/forum?id=Hly2QYcuLW","abs":"

LiDAR point cloud anomaly detection is critical for autonomous system safety, yet most existing methods rely only on visible measurements, overlooking occlusion as a structured consequence of the LiDAR sensing process. We argue that anomalies are characterized not only by what is observed, but also by the spatial voids they create, which alter occlusion patterns and volumetric visibility. We propose Counterfactual Occlusion-Visibility Anomaly Learning (COVAL), a framework that intervenes on volumetric visibility during training. Using physics-conformed synthetic anomaly construction, COVAL generates paired factual and counterfactual observations with identical scene geometry but different occlusion patterns. Then, we introduce two complementary objectives: Visibility-Variant Counterfactual Reconstruction, which models occlusion-induced missing regions, and Visibility-Invariant Counterfactual Consistency, which enforces stable representations across visibility changes. Together, these objectives isolate anomaly-induced structural missingness and in turn refine representation of normal scenes, thus improving anomaly sensitivity at test time. Experiments on standard LiDAR anomaly segmentation benchmarks show that COVAL achieves state-of-the-art performance.

","absKo":"

LiDAR point cloud anomaly detection은 자율 시스템 안전에 매우 중요하지만, 기존 방법의 대부분은 보이는 measurement에만 의존하여 LiDAR sensing process의 구조적 결과로서의 occlusion을 간과한다. 우리는 anomaly가 단지 관측되는 것에 의해서만이 아니라, occlusion pattern과 volumetric visibility를 바꾸는 spatial void에 의해서도 특징지어진다고 주장한다. 우리는 학습 동안 volumetric visibility에 개입하는 Counterfactual Occlusion-Visibility Anomaly Learning(COVAL)을 제안한다. physics-conformed synthetic anomaly construction을 사용하여, COVAL은 동일한 scene geometry를 가지지만 서로 다른 occlusion pattern을 갖는 factual 및 counterfactual observation 쌍을 생성한다. 그런 다음, 우리는 두 개의 상보적 objective를 도입한다. Visibility-Variant Counterfactual Reconstruction은 occlusion으로 인해 유발되는 missing region을 모델링하고, Visibility-Invariant Counterfactual Consistency는 visibility 변화 전반에 걸쳐 안정적인 representation을 강제한다. 이 목적함수들은 함께 anomaly가 유발한 structural missingness를 분리하고, 그 결과 normal scene의 representation을 정제하여 test time에 anomaly sensitivity를 향상시킨다. 표준 LiDAR anomaly segmentation benchmark에서의 실험은 COVAL이 state-of-the-art 성능을 달성함을 보여 준다.

"},{"id":"65348","en":"CoverPruneGS: Coverage-Preserving Structured Pruning for Hierarchical 3D Gaussian Splatting from Sparse-View Monocular Videos","ko":"CoverPruneGS: Sparse-View 단안 영상으로부터 계층적 3D Gaussian Splatting을 위한 커버리지 보존 구조적 가지치기","authors":"Yang Xiao, Guoan Xu, Guxue Gao, Qiang Wu, Wenjing Jia","pos":"#1414","link":"https://openreview.net/forum?id=ETnzn3onnZ","abs":"

Reconstructing a complete yet compact 3DGS from sparse-view monocular long videos is challenging: hierarchical training with VFI can improve coverage, yet correlated pseudo views and repeated merging tend to accumulate near-duplicate Gaussians and exacerbate co-adaptation. To address this, we propose CoverPruneGS, a coverage-preserving structured pruning framework tailored for hierarchical merging with VFI-augmented supervision, which performs coarse-to-fine pruning via voxel-based local diversity selection and GT-guided lazy refinement with randomized dropout rendering. To make refinement reliable, we introduce a footprint-aware CUDA attribution that aggregates GT-aligned error degradation over Gaussian-influenced pixels in a manner consistent with alpha compositing, yielding faithful per-Gaussian scores for quantile-based rescue. Experiments on multiple datasets demonstrate that CoverPruneGS substantially reduces the Gaussian count by 56.8\\% and accelerates inference while maintaining or improving novel view synthesis quality.

","absKo":"

sparse-view monocular long video로부터 완전하면서도 compact한 3DGS를 재구성하는 것은 어렵다. VFI를 포함한 hierarchical training은 coverage를 개선할 수 있지만, 상관된 pseudo view와 반복적인 merging은 near-duplicate Gaussian을 축적하고 co-adaptation을 악화시키는 경향이 있다. 이를 해결하기 위해 우리는 VFI-augmented supervision을 사용하는 hierarchical merging에 맞춘 coverage-preserving structured pruning framework인 CoverPruneGS를 제안한다. 이 방법은 voxel-based local diversity selection과 GT-guided lazy refinement, randomized dropout rendering을 활용해 coarse-to-fine pruning을 수행한다. refinement를 신뢰할 수 있게 만들기 위해, 우리는 alpha compositing과 일관되게 Gaussian의 영향을 받는 pixel 전반의 GT-aligned error degradation을 집계하는 footprint-aware CUDA attribution을 도입하여, quantile-based rescue를 위한 신뢰할 수 있는 per-Gaussian score를 산출한다. 여러 dataset에서의 실험은 CoverPruneGS가 Gaussian 수를 56.8\\% 크게 줄이면서도 novel view synthesis 품질을 유지하거나 향상시키고, 추론 속도를 높인다는 것을 보여준다.

"},{"id":"61853","en":"Spatial Priors via Space Filling Curves for Small and Limited Data Vision Transformers","ko":"작고 제한된 데이터 Vision Transformer를 위한 Space Filling Curve 기반 공간 Prior","authors":"Leyla Candogan, Arshia Afzal, Pol Puigdemont, Volkan Cevher","pos":"#2914","link":"https://openreview.net/forum?id=mLVSIefyWj","abs":"

Though Vision Transformers (ViTs) have become the dominant backbone in many computer vision tasks, due to permutation invariance, their attention mechanism lacks explicit spatial inductive biases. This become particularly important in two common settings: when model capacity is small or training data is limited. Inspired by the attention masking strategies in Linear Transformers and the scanning patterns of Vision SSMs, we introduce VIOLIN, a lightweight masked attention mechanism that encodes spatial structure within attention via Space Filling Curves (SFCs) with less than 0.0002% extra parameters and negligible computational overhead. VIOLIN scans the image using multiple SFCs to construct curve-specific decay masks, which are then combined and multiplied with the attention matrix. Across a wide range of evaluations, VIOLIN consistently improves performance. In limited data regimes such as fine-tuning on VTAB-1K, it boosts accuracy across all task groups and by up to 8.7% on the Structured tasks, where spatial information is essential. It can be combined with parameter-efficient fine-tuning methods such as LoRA to further increase the performance. Beyond fine-tuning, VIOLIN improves various small scale ViT architectures (e.g., DeiT, DINO) during pretraining on ImageNet-1K, achieving gains of up to 0.9%. Additionally, on pixel-level CIFAR-100 training, a task that is highly dependent on location information, VIOLIN increases accuracy by up to 7.2%. Overall, VIOLIN provides a computationally efficient yet effective way to inject spatial inductive bias into ViTs, especially benefiting small models and limited data settings.

","absKo":"

Vision Transformer(ViT)가 많은 computer vision task에서 지배적인 backbone이 되었지만, permutation invariance 때문에 그 attention mechanism은 명시적인 spatial inductive bias가 부족하다. 이는 특히 두 가지 흔한 설정에서 중요해진다: 모델 capacity가 작거나 training data가 제한적인 경우이다. Linear Transformer의 attention masking strategy와 Vision SSM의 scanning pattern에서 영감을 받아, 우리는 Space Filling Curve(SFC)를 통해 attention 내부에 spatial structure를 인코딩하는 lightweight masked attention mechanism인 VIOLIN을 제안한다. 이는 0.0002% 미만의 추가 parameter와 무시할 수 있는 computational overhead만을 필요로 한다. VIOLIN은 여러 SFC를 사용해 이미지를 스캔하여 curve-specific decay mask를 구성하고, 이를 결합한 뒤 attention matrix와 곱한다. 광범위한 평가에서 VIOLIN은 일관되게 성능을 향상시킨다. VTAB-1K에서의 fine-tuning과 같은 제한된 data regime에서는 모든 task group에서 accuracy를 높이며, spatial information이 필수적인 Structured task에서는 최대 8.7%까지 향상시킨다. 또한 LoRA와 같은 parameter-efficient fine-tuning method와 결합하여 성능을 더 끌어올릴 수 있다. fine-tuning을 넘어, VIOLIN은 ImageNet-1K에서의 pretraining 동안 DeiT, DINO와 같은 다양한 small-scale ViT architecture를 개선하여 최대 0.9%의 향상을 달성한다. 추가로, location information에 크게 의존하는 pixel-level CIFAR-100 training에서 VIOLIN은 accuracy를 최대 7.2%까지 높인다. 종합하면, VIOLIN은 ViT에 spatial inductive bias를 주입하는 계산적으로 효율적이면서도 효과적인 방법을 제공하며, 특히 작은 모델과 제한된 data setting에서 이점을 보인다.

"},{"id":"65972","en":"Required Spine Optional Limbs: Heterogeneous Federated Learning via Backbone-sharing and Activation-guided Selection","ko":"필수 척추, 선택적 사지: 백본 공유와 활성화 유도 선택을 통한 이종 연합 학습","authors":"Mingsheng Cao, Hongliang Chen, Ming Hu, Fei Gao, Qiaolong Ding, Wenke Huang, Xiaofei Xie, Junlong Zhou","pos":"#3008","link":"https://openreview.net/forum?id=8LZfyxIQdO","abs":"

Although Federated Learning (FL) offers advantages in privacy-preserving for cross-device collaborative learning, its practical deployment remains severely constrained by heterogeneous hardware resources and non-IID (non-independent and identically distributed) data across devices. Sub-model extraction has emerged as a widely adopted strategy for enabling collaborative training among devices with heterogeneous models. However, existing sub-model extraction methods in FL typically rely on coarse-grained stochastic selection or rigid rule-based neuron selection, which severely limits training performance. Specifically, stochastic strategies lead to severe parameter conflicts under non-IID data distributions, while rule-based approaches lack diversity in neuron selection per device, preventing comprehensive parameter optimization. To address this problem, this paper presents a novel sub-model extraction-based FL framework, named SpineFL, which adopts a backbone-sharing mechanism and an activation-guided pruning strategy for sub-model extraction. Specifically, SpineFL decomposes each global model layer into two portions: i) a mandatory backbone shared by all the sub-models to maintain model generalization, and ii) a dynamic portion for sub-model extraction. SpineFL adopts the activation-guided selection strategy to probabilistically select neurons according to their activation frequency from the dynamic portion to generate sub-model, where neurons exhibiting higher historical activation are more likely to be included, thereby simultaneously addressing parameter conflicts while preserving selection diversity. Experimental results demonstrate that compared with state-of-the-art heterogeneous FL methods, SpineFL can achieve up to 3.28% accuracy improvement.

","absKo":"

Federated Learning (FL)은 device 간 협업 학습에서 privacy-preserving 측면의 장점을 제공하지만, 실제 배포는 이질적인 hardware resource와 device 간 non-IID (non-independent and identically distributed) data로 인해 여전히 심각한 제약을 받는다. Sub-model extraction은 이질적인 model을 가진 device 간 협업 학습을 가능하게 하는 널리 채택된 전략으로 부상했다. 그러나 FL에서의 기존 sub-model extraction 방법은 대체로 거친 수준의 stochastic selection 또는 경직된 rule-based neuron selection에 의존하며, 이는 training performance를 심각하게 제한한다. 구체적으로, stochastic strategy는 non-IID data distribution 하에서 심각한 parameter conflict를 유발하는 반면, rule-based approach는 device별 neuron selection의 다양성이 부족하여 포괄적인 parameter optimization을 방해한다. 이 문제를 해결하기 위해, 본 논문은 backbone-sharing mechanism과 activation-guided pruning strategy를 sub-model extraction에 적용한 새로운 sub-model extraction-based FL framework인 SpineFL을 제안한다. 구체적으로 SpineFL은 각 global model layer를 두 부분으로 분해한다. i) model generalization을 유지하기 위해 모든 sub-model이 공유하는 필수 backbone, ii) sub-model extraction을 위한 dynamic portion. SpineFL은 activation-guided selection strategy를 채택하여 dynamic portion에서 activation frequency에 따라 neuron을 확률적으로 선택해 sub-model을 생성하는데, 역사적으로 더 높은 activation을 보이는 neuron일수록 포함될 가능성이 높아져 parameter conflict를 해소함과 동시에 selection diversity를 보존한다. 실험 결과는 state-of-the-art heterogeneous FL method와 비교했을 때 SpineFL이 최대 3.28%의 accuracy 향상을 달성할 수 있음을 보여준다.

"},{"id":"60762","en":"PointCHR: Point Cloud Analysis via Curvature-Aware Hyperbolic Rectification","ko":"PointCHR: Curvature-Aware Hyperbolic Rectification을 통한 Point Cloud 분석","authors":"Xinxing Yu, Liying Yang, Hao Mo, Hui Ma, Fang Kai, Ajian Liu, Yanyan Liang","pos":"#3013","link":"https://openreview.net/forum?id=wtmbT1fKnZ","abs":"

High-curvature regions in 3D point clouds encapsulate critical fine-grained geometric semantics yet exhibit a distinct long-tail sparsity in their spatial distribution. The inherent limitations of polynomial volume growth in Euclidean space frequently render these intricate geometric features challenging to adequately resolve within a uniform-scale feature space. Consequently, these regions are frequently overshadowed by smooth global features dominated by low-curvature regions, thereby limiting the discriminative capacity of the network. To address this issue, we propose PointCHR, a curvature-aware hyperbolic rectification (CHR) for point cloud analysis. Utilising the property of exponential volume expansion in the vicinity of hyperbolic manifolds, CHR presents a learnable curvature-guided radial rectification mechanism. By adaptively projecting high-curvature points towards boundary regions endowed with larger effective embedding capacities, PointCHR effectively mitigates the representation crowding problem inherent in Euclidean settings. Extensive experimentation has demonstrated that PointCHR significantly enhances the ability of backbone to capture fine-grained geometric details, achieving state-of-the-art performance across multiple benchmarks.

","absKo":"

3D point cloud의 고곡률 영역은 중요한 세밀한 geometric semantic을 담고 있지만, 공간 분포에서는 뚜렷한 long-tail sparsity를 보인다. Euclidean space에서 polynomial volume growth가 갖는 본질적 한계 때문에 이러한 정교한 geometric feature는 uniform-scale feature space 안에서 충분히 분해해내기 어렵다. 그 결과, 이 영역들은 종종 저곡률 영역이 지배하는 매끄러운 global feature에 가려지며, 이로 인해 network의 discriminative capacity가 제한된다. 이 문제를 해결하기 위해, 우리는 point cloud analysis를 위한 curvature-aware hyperbolic rectification (CHR)인 PointCHR을 제안한다. Hyperbolic manifold 근처에서의 exponential volume expansion 특성을 활용하여, CHR은 학습 가능한 curvature-guided radial rectification mechanism을 제시한다. 고곡률 point를 더 큰 effective embedding capacity를 갖는 boundary region 쪽으로 적응적으로 투영함으로써, PointCHR은 Euclidean setting에 내재된 representation crowding 문제를 효과적으로 완화한다. 광범위한 실험은 PointCHR이 backbone이 세밀한 geometric detail을 포착하는 능력을 크게 향상시키며, 여러 benchmark에서 state-of-the-art 성능을 달성함을 보여준다.

"},{"id":"66225","en":"Transform Trained Transformer for Accelerating Native 4K Video Generation","ko":"네이티브 4K 비디오 생성 가속을 위한 훈련된 Transformer 변환","authors":"Jiangning Zhang, Junwei Zhu, Teng Hu, Yabiao Wang, Donghao Luo, Weijian Cao, Zhenye Gan, Xiaobin Hu, Zhucun Xue, Xiangtai Li, Chengjie Wang, Yong Liu","pos":"#3803","link":"https://openreview.net/forum?id=5dabjiOBpS","abs":"Native 4K (2176$\\times$3840) video generation remains a critical challenge due to the quadratic computational explosion of full-attention as spatiotemporal resolution increases, making it difficult for models to strike a balance between efficiency and quality. This paper proposes a novel Transformer retrofit strategy termed T3 (**T**ransform **T**rained **T**ransformer) that, without altering the core architecture of full-attention pretrained models, significantly reduces compute requirements by optimizing their forward logic. Specifically, T3-Video introduces a multi-scale weight-sharing window attention mechanism and, via hierarchical blocking together with an axis-preserving full-attention design, can effect an “attention pattern” transformation of a pretrained model using only modest compute and data. Results on 4K-VBench show that T3-Video substantially outperforms existing approaches: while delivering performance improvements (+4.29$\\uparrow$ VQA and +0.08$\\uparrow$ VTC), it accelerates native 4K video generation by more than 10$\\times$. Demo and source code are available in \\#Supp.","absKo":"Native 4K (2176$\\times$3840) video generation은 spatiotemporal resolution이 증가함에 따라 full-attention의 quadratic computational explosion 때문에 여전히 중요한 도전 과제이며, model이 efficiency와 quality 사이의 균형을 맞추기 어렵게 만든다. 본 논문은 T3 (**T**ransform **T**rained **T**ransformer)라는 새로운 Transformer retrofit strategy를 제안한다. 이는 full-attention pretrained model의 core architecture를 바꾸지 않고도 forward logic을 최적화하여 compute 요구량을 크게 줄인다. 구체적으로, T3-Video는 multi-scale weight-sharing window attention mechanism을 도입하고, hierarchical blocking과 axis-preserving full-attention design을 통해 modest compute와 data만으로 pretrained model의 “attention pattern” transformation을 수행할 수 있다. 4K-VBench 결과에서 T3-Video는 기존 방법을 크게 능가했다. 성능 향상(+4.29$\\uparrow$ VQA 및 +0.08$\\uparrow$ VTC)을 제공하는 동시에 native 4K video generation을 10$\\times$ 이상 가속한다. Demo와 source code는 \\#Supp에서 확인할 수 있다."},{"id":"60671","en":"ScaleErasure: Inference-Time Minimal Intervention for Precise Concept Erasure in Next-Scale Autoregressive Image Generation","ko":"ScaleErasure: Next-Scale Autoregressive Image Generation에서 정밀한 Concept Erasure를 위한 추론 시 최소 개입","authors":"Cong Wang, Haiyu Wu, Zhiwei Jiang, Zifeng Cheng, SHEN FEI, Yafeng Yin, Qing Gu","pos":"#1211","link":"https://openreview.net/forum?id=xnap60lY3b","abs":"

Concept erasure aims to prevent image generative models from producing unsafe content while preserving their general generative capability. Meanwhile, next-scale autoregressive (AR) image generation has recently emerged as a new generative paradigm characterized by next-scale prediction, for which concept erasure remains largely unexplored. In this paradigm, semantic information is highly compressed at early scales, leading to severe entanglement between unsafe and unrelated semantics. In this paper, we propose ScaleErasure, an inference-time concept erasure method that performs minimal intervention. ScaleErasure precisely selects and guides predicted logits that are most relevant to the unsafe concept, thereby enabling effective erasure under severe semantic entanglement. Specifically, ScaleErasure performs two additional forward passes conditioned on the unsafe concept and the corresponding safe concept, and leverages their outputs to guide the target logits away from unsafe concepts toward safe concepts. To enable precise and minimal intervention, logits selection and guidance are conducted across three dimensions: scales, tokens, and bit channels. Experiments demonstrate that ScaleErasure outperforms adapted baselines in the next-scale AR paradigm, achieving more precise concept erasure while largely preserving general generative capability.

","absKo":"

Concept erasure는 이미지 생성 모델이 일반적인 생성 능력은 유지하면서도 unsafe content를 생성하지 않도록 하는 것을 목표로 한다. 한편, next-scale autoregressive (AR) image generation은 최근 next-scale prediction을 특징으로 하는 새로운 생성 패러다임으로 등장했지만, 이 설정에서 concept erasure는 아직 거의 탐구되지 않았다. 이 패러다임에서는 초기 scale에서 semantic information이 매우 압축되므로, unsafe semantics와 unrelated semantics 사이의 얽힘이 심해진다. 본 논문에서는 최소 개입을 수행하는 inference-time concept erasure 방법인 ScaleErasure를 제안한다. ScaleErasure는 unsafe concept와 가장 관련 있는 predicted logits를 정밀하게 선택하고 이를 guidance함으로써, 심각한 semantic entanglement 하에서도 효과적인 erasure를 가능하게 한다. 구체적으로, ScaleErasure는 unsafe concept와 대응되는 safe concept를 조건으로 두 번의 추가 forward pass를 수행하고, 그 출력을 활용해 target logits이 unsafe concepts에서 safe concepts 쪽으로 벗어나도록 유도한다. 정밀하고 최소한의 개입을 가능하게 하기 위해, logits selection과 guidance는 scale, token, bit channel의 세 차원에서 수행된다. 실험은 ScaleErasure가 next-scale AR 패러다임에서 adapted baselines를 능가하며, general generative capability를 대체로 보존하면서도 더 정밀한 concept erasure를 달성함을 보인다.

"},{"id":"62442","en":"SpaCeFormer: Space-Curve Transformer for Open-Vocabulary 3D Instance Segmentation without Proposals","ko":"SpaCeFormer: Proposal 없이 Open-Vocabulary 3D Instance Segmentation을 위한 Space-Curve Transformer","authors":"Christopher Choy, Junha Lee, Chunghyun Park, Minsu Cho, Jan Kautz","pos":"#1213","link":"https://openreview.net/forum?id=gYBeaguAN5","abs":"Open-vocabulary 3D segmentation is crucial for real-world applications, yet existing methods are constrained by fragmented masks and inconsistent captions in dataset generation, and by multi-stage pipelines prone to error propagation. We present SpaCeFormer-3M, the largest open-vocabulary 3D instance segmentation dataset with 846K instances from 15K scenes, and SpaCeFormer (Space-Curve Transformer), a proposal-free segmentation architecture. Our data pipeline leverages multi-view mask clustering to produce geometry-consistent 3D instances and employs multi-view VLM prompting for view-consistent captions. On the modeling side, SpaCeFormer combines spatial window attention with Morton curve serialization for spatially coherent features, and a RoPE-enhanced decoder to predict instance masks directly from learned queries without external proposals. On ScanNet200, our approach achieves 11.1 zero-shot mAP, a 2.8$\\times$ improvement over prior proposal-free methods while requiring only 0.21 seconds per scene.","absKo":"Open-vocabulary 3D segmentation은 실제 응용에 중요하지만, 기존 방법은 dataset 생성에서 fragmented mask와 일관성 없는 caption, 그리고 오류 전파에 취약한 multi-stage pipeline의 제약을 받는다. 우리는 15K scenes에서 846K instances를 포함하는 최대 규모의 open-vocabulary 3D instance segmentation dataset인 SpaCeFormer-3M과, proposal-free segmentation architecture인 SpaCeFormer(Space-Curve Transformer)를 제시한다. 우리의 data pipeline은 multi-view mask clustering을 활용해 geometry-consistent 3D instances를 생성하고, multi-view VLM prompting을 사용해 view-consistent caption을 만든다. 모델 측면에서 SpaCeFormer는 spatial window attention과 Morton curve serialization을 결합해 spatially coherent feature를 형성하고, RoPE가 강화된 decoder를 통해 외부 proposal 없이 학습된 query로부터 instance mask를 직접 예측한다. ScanNet200에서 우리의 방법은 11.1 zero-shot mAP을 달성했으며, 이는 기존 proposal-free 방법 대비 2.8$\\times$ 향상된 수치이고 scene당 0.21 seconds만 필요로 한다."},{"id":"65315","en":"Uni-DocRobust: Universal Plug-and-Play Robustness Enhancement for Multi-modal LLMs via Feature Restoration","ko":"Uni-DocRobust: 특징 복원을 통한 Multi-modal LLM의 범용 Plug-and-Play 강건성 향상","authors":"Yuxuan Zhou, Baole Wei, Xingjian Hu, Haowei Chen, Yu Li, Xingyue Lin, Liangcai Gao, Zhi Tang","pos":"#1107","link":"https://openreview.net/forum?id=ErZRlWv8kZ","abs":"

Real-world degradations, such as noise, blur, and low resolution, significantly impair the performance of Multi-modal Large Language Models (MLLMs) in document understanding tasks. Despite recent advancements, progress in this field remains stifled by two critical bottlenecks: the scarcity of large-scale, aligned training data necessary for learning robustness, and the lack of transferable restoration solutions across diverse MLLM architectures. To bridge the data gap, we first present DocRobust-VQA, a large-scale dataset explicitly constructed to support robustness training. Comprising 189K aligned low/high-quality document image pairs and 417K QA pairs, it provides the first substantial corpus for fine-tuning MLLMs to handle varying degradation conditions. Leveraging this data, we propose Uni-DocRobust, a universal plug-and-play framework that decouples restoration capabilities from specific visual encoders. Our method employs a frozen Universal Restoration Core pre-trained in a canonical feature space via multi-teacher distillation, which can be seamlessly integrated into target MLLMs (e.g., Qwen-VL, InternVL) through lightweight Feature Adapters. Extensive experiments demonstrate that Uni-DocRobust significantly enhances robust performance on MLLMs and enables a cost-effective ``pre-train once, deploy everywhere'' paradigm for robust MLLM deployment.

","absKo":"

노이즈, blur, low resolution과 같은 실제 환경의 degradation은 document understanding task에서 Multi-modal Large Language Models (MLLMs)의 성능을 크게 저해한다. 최근의 진전에도 불구하고, 이 분야의 발전은 두 가지 핵심 병목에 의해 여전히 제약된다. 하나는 robustness 학습에 필요한 대규모 정렬된 training data의 부족이고, 다른 하나는 다양한 MLLM architecture 전반에 걸쳐 transferable restoration solution이 없다는 점이다. 데이터 부족을 해소하기 위해, 우리는 먼저 robustness training을 명시적으로 지원하도록 구성된 대규모 데이터셋 DocRobust-VQA를 제시한다. 이 데이터셋은 189K개의 정렬된 low/high-quality document image pair와 417K개의 QA pair로 이루어져 있으며, 다양한 degradation condition을 다룰 수 있도록 MLLM을 fine-tuning하기 위한 첫 번째로 충분히 큰 corpus를 제공한다. 이 데이터를 활용하여, 우리는 특정 visual encoder로부터 restoration capability를 분리한 universal plug-and-play framework인 Uni-DocRobust를 제안한다. 우리의 방법은 multi-teacher distillation을 통해 canonical feature space에서 pre-training된 frozen Universal Restoration Core를 사용하며, 이를 lightweight Feature Adapter를 통해 target MLLM(예: Qwen-VL, InternVL)에 매끄럽게 통합할 수 있다. 광범위한 실험은 Uni-DocRobust가 MLLM의 robust performance를 크게 향상시키고, robust MLLM 배포를 위한 비용 효율적인 ``한 번 pre-train, 어디서나 deploy'' 패러다임을 가능하게 함을 보여준다.

"},{"id":"60956","en":"Towards Trustworthy and Identifiable Virtual Face Generation","ko":"신뢰할 수 있고 식별 가능한 Virtual Face Generation을 향하여","authors":"Chunyang Li, Haoyue Wang, Zhenxing Qian, Sheng Li, Xinpeng Zhang, Jian liu, Yixuan Pei, Weiqiang Wang","pos":"#1108","link":"https://openreview.net/forum?id=v3L3mIvUDc","abs":"

Identifiable virtual face (IVF) generation aims to transform a user's original face into a virtual face for high utility privacy protection. The IVF is visually and statistically different from the original face, which can still be used for recognizing the user's identity. Despite the advantage, these schemes are unable to verify the trustworthiness of the IVF, the quality and controllability of which is often limited. To address these issues, we propose TIVDiff, a diffusion-based framework for trustworthy and identifiable virtual face generation. TIVDiff learns a virtual identity (VID) space via Virtual Identity Projection (VIP) and synthesizes high-quality virtual faces conditioned on VID and 3D facial geometry for pose and expression preservation. To enable the trustworthiness of IVF, we introduce Identity-Guarded Generative Watermarking (IGGW) to bind the diffusion initial noise with VID through a reversible mapping. This enables the embedding of an imperceptible cue into IVF for legitimacy verification. Experiments demonstrate the advantage of our TIVDiff over the state-of-the-art IVF generation schemes in terms of image quality, identifiability and trustworthiness.

","absKo":"

Identifiable virtual face(IVF) generation은 사용자의 원본 얼굴을 가상 얼굴로 변환하여 높은 유용성을 유지하는 privacy protection을 목표로 한다. IVF는 시각적으로나 통계적으로 원본 얼굴과 다르지만, 여전히 사용자의 신원을 인식하는 데 활용될 수 있다. 이러한 장점에도 불구하고, 이러한 scheme은 IVF의 trustworthiness를 검증할 수 없으며, 그 quality와 controllability도 종종 제한적이다. 이러한 문제를 해결하기 위해, 우리는 trustworthy하고 identifiable한 virtual face generation을 위한 diffusion-based framework인 TIVDiff를 제안한다. TIVDiff는 Virtual Identity Projection(VIP)을 통해 virtual identity(VID) space를 학습하고, 자세와 표정 보존을 위해 VID와 3D facial geometry를 조건으로 고품질 virtual face를 합성한다. IVF의 trustworthiness를 가능하게 하기 위해, 우리는 reversible mapping을 통해 diffusion initial noise와 VID를 결합하는 Identity-Guarded Generative Watermarking(IGGW)을 도입한다. 이를 통해 IVF에 보이지 않는 cue를 삽입하여 legitimacy verification이 가능해진다. 실험은 image quality, identifiability, trustworthiness 측면에서 우리의 TIVDiff가 state-of-the-art IVF generation scheme보다 우수함을 보여준다.

"},{"id":"64091","en":"Towards Realistic Lifelong Re-identification: Identity Recurrence with Changing Clothes","ko":"현실적인 평생 재식별을 향하여: 옷 변경을 동반한 신원 재출현","authors":"Wuxuan Shi, Zhijie Lu, He Li, Mang Ye","pos":"#1109","link":"https://openreview.net/forum?id=R16bvfNfld","abs":"

Existing lifelong person re-identification (Re-ID) methods assume that each identity maintains a relatively stable appearance distribution over time. However, in real-world scenarios, identities often reappear asynchronously with substantial clothing changes, which is not modeled in existing lifelong Re-ID formulations. We therefore study a realistic lifelong cloth-changing Re-ID (LCC) setting, in which identities reappear asynchronously under substantial clothing changes. This setting leads to two core difficulties: the model must acquire new identities while adapting representations of recurring ones, and at the same time remain robust to the substantial representation drift induced by clothing changes over time, which undermines cross-temporal identity consistency. To address these, we develop a framework that disentangles identity-intrinsic representations from clothing-induced biases, enabling identity modeling beyond appearance changes. We further introduce a Dynamic Identity-Anchor Alignment to maintain stable identity anchors under stage-wise distribution shifts. Experiments on the LTCC and PRCC benchmarks demonstrate superior performance and representational stability across multiple learning stages.

","absKo":"

기존 lifelong person re-identification(Re-ID) 방법은 각 identity가 시간에 따라 비교적 안정적인 appearance distribution을 유지한다고 가정한다. 그러나 실제 환경에서는 identity가 substantial clothing change와 함께 비동기적으로 다시 나타나는 경우가 많으며, 이는 기존 lifelong Re-ID formulation에서 모델링되지 않았다. 따라서 우리는 실제적인 lifelong cloth-changing Re-ID(LCC) setting을 연구한다. 이 setting에서는 identity가 substantial clothing change 하에서 비동기적으로 재등장한다. 이러한 setting은 두 가지 핵심 어려움을 낳는다. 모델은 새로운 identity를 습득하는 동시에 반복적으로 등장하는 identity의 representation을 적응시켜야 하며, 동시에 시간이 지남에 따라 clothing change가 유발하는 substantial representation drift에도 견고해야 하는데, 이는 cross-temporal identity consistency를 저해한다. 이를 해결하기 위해, 우리는 identity-intrinsic representation과 clothing-induced bias를 분리하는 framework를 개발하여 appearance 변화 너머의 identity modeling을 가능하게 한다. 또한 stage-wise distribution shift 하에서 안정적인 identity anchor를 유지하기 위해 Dynamic Identity-Anchor Alignment를 도입한다. LTCC와 PRCC benchmark에서의 실험은 여러 learning stage에 걸쳐 우수한 성능과 representation stability를 보여준다.

"},{"id":"61618","en":"Towards Effective Waste Segmentation for Automated Waste Recycling in Cluttered Background","ko":"복잡한 배경에서 자동 폐기물 재활용을 위한 효과적인 폐기물 분할을 향하여","authors":"Mamoona Javaid, Mubashir Noman, Abdul Hannan, Shah Nawaz, Mustansar Fiaz, Sajid Ghuffar","pos":"#1111","link":"https://openreview.net/forum?id=oSuKfnbPFB","abs":"

Rapid expansion of urban areas and population growth is causing an immense increase in waste production, which demands the need for efficient and automated waste management. In this scenario, automated waste recycling (AWR) that utilizes deep learning methods to separate the recyclable waste objects may emerge as a savior to humanity. Recent deep learning approaches for AWR provide promising waste segmentation performance, however, these methods rely on large backbone networks that are inefficient for AWR systems and suffer from performance deterioration in cluttered scenes. To this end, an optimal waste segmentation network is introduced which effectively utilizes the spatial domain to capture localized structural dependencies and the spectral domain to efficiently extract global contextual relationships. This cascaded design allows the network to progressively leverage both local and global representations across complementary domains to highlight the semantic information necessary for effective segmentation of various waste objects. Furthermore, auxiliary feature enhancement module (AFEM) is introduced to enhance the target objects' boundaries and blob amplification for better segmentation in cluttered scenarios. Extensive experimentation on ZeroWaste-aug, ZeroWaste-f and SpectralWaste reveals the merits of the proposed method.

","absKo":"

도시 지역의 급격한 확장과 인구 증가는 폐기물 발생을 엄청나게 늘리고 있으며, 이는 효율적이고 자동화된 폐기물 관리의 필요성을 요구한다. 이러한 상황에서, 재활용 가능한 폐기물 객체를 분리하기 위해 deep learning 방법을 활용하는 automated waste recycling(AWR)은 인류를 구원할 수 있는 해법으로 떠오를 수 있다. 최근 AWR을 위한 deep learning 접근법들은 유망한 폐기물 segmentation 성능을 제공하지만, 이러한 방법들은 AWR 시스템에 비효율적인 대규모 backbone network에 의존하며 복잡하게 뒤엉킨 장면에서는 성능 저하를 겪는다. 이를 위해, 공간 도메인을 효과적으로 활용해 국소적인 구조적 의존성을 포착하고 spectral 도메인을 통해 전역 문맥 관계를 효율적으로 추출하는 optimal waste segmentation network를 제안한다. 이러한 cascaded 설계는 네트워크가 상호보완적인 도메인 전반에서 local 및 global representation을 점진적으로 활용하여 다양한 폐기물 객체의 효과적인 segmentation에 필요한 semantic 정보를 부각하도록 한다. 또한, 복잡한 상황에서 더 나은 segmentation을 위해 target object의 경계를 강화하고 blob amplification을 수행하는 auxiliary feature enhancement module(AFEM)을 도입한다. ZeroWaste-aug, ZeroWaste-f, SpectralWaste에 대한 광범위한 실험은 제안 방법의 장점을 보여준다.

"},{"id":"61405","en":"Supervise Less, See More: Training-free Nuclear Instance Segmentation with Prototype-Guided Prompting","ko":"Supervise Less, See More: Prototype-Guided Prompting을 활용한 Training-free Nuclear Instance Segmentation","authors":"Wen Zhang, Qin Ren, Wenjing Liu, Haibin Ling, Chenyu You","pos":"#1112","link":"https://openreview.net/forum?id=qYfNhYenuu","abs":"

Accurate nuclear instance segmentation is a pivotal task in computational pathology, supporting data-driven clinical insights and facilitating downstream translational applications. While large vision foundation models have shown promise for zero-shot biomedical segmentation, most existing approaches still depend on dense supervision and computationally expensive fine-tuning. Consequently, training-free methods present a compelling research direction, yet remain largely unexplored. In this work, we introduce SPROUT, a fully training- and annotation-free prompting framework for nuclear instance segmentation. SPROUT leverages histology-informed priors to construct slide-specific reference prototypes that mitigate domain gaps. These prototypes progressively guide feature alignment through a partial optimal transport scheme. The resulting foreground and background features are transformed into positive and negative point prompts, enabling the Segment Anything Model (SAM) to produce precise nuclear delineations without any parameter updates. Extensive experiments across multiple histopathology benchmarks demonstrate that SPROUT achieves competitive performance without supervision or retraining, establishing a novel paradigm for scalable, training-free nuclear instance segmentation in pathology.

","absKo":"

정확한 nuclear instance segmentation은 computational pathology에서 핵심적인 과제로, 데이터 기반 임상 통찰을 지원하고 downstream translational application을 촉진한다. 대규모 vision foundation model이 zero-shot biomedical segmentation에서 가능성을 보여 주었지만, 기존 접근법의 대부분은 여전히 dense supervision과 계산 비용이 큰 fine-tuning에 의존한다. 따라서 training-free method는 매우 유망한 연구 방향이지만, 아직 충분히 탐구되지 않았다. 본 연구에서는 nuclear instance segmentation을 위한 완전한 training- 및 annotation-free prompting framework인 SPROUT를 소개한다. SPROUT는 histology-informed prior를 활용하여 slide-specific reference prototype을 구성하고, 이를 통해 domain gap을 완화한다. 이러한 prototype은 partial optimal transport scheme을 통해 feature alignment를 점진적으로 유도한다. 그 결과 얻어진 foreground 및 background feature는 positive 및 negative point prompt로 변환되며, 이를 통해 Segment Anything Model (SAM)이 어떠한 parameter update 없이도 정밀한 nuclear delineation을 생성할 수 있다. 여러 histopathology benchmark에 걸친 광범위한 실험은 SPROUT가 supervision이나 retraining 없이도 경쟁력 있는 성능을 달성함을 보여 주며, pathology에서 확장 가능하고 training-free한 nuclear instance segmentation의 새로운 패러다임을 확립한다.

"},{"id":"62651","en":"Spectral–Spatial Mixing with Morphology-Aware Adaptive Loss for Medical Image Segmentation.","ko":"의료 영상 분할을 위한 Morphology-Aware Adaptive Loss 기반 Spectral–Spatial Mixing","authors":"Sanaullah Chowdhury, Lameya Sabrin","pos":"#1114","link":"https://openreview.net/forum?id=eh48NIgu9z","abs":"Medical image segmentation requires balancing global context with computational efficiency, where self-attention mechanisms suffer from quadratic $\\mathcal{O}((HW)^2 C)$ complexity. We propose S2M-Net, a parameter-efficient architecture (4.7M parameters) that achieves computational savings through Spectral--Spatial Token Mixing (SSTM). SSTM achieves $\\mathcal{O}(HWC^2)$ complexity through efficient combination of $\\mathcal{O}(HWC \\log(HW))$ frequency-domain processing and $\\mathcal{O}(HWCd)$ bottlenecked spatial gating ($d{=}16$), exploiting spectral concentration where $>93\\%$ of energy is captured by $K{=}32$ low-frequency components ($\\sim$0.8\\% of the spectrum at $352{\\times}352$ resolution). This design avoids self-attention's prohibitive $\\mathcal{O}((HW)^2C)$ attention map computations while preserving global receptive fields. To handle geometric diversity, we introduce Morphology-Aware Adaptive Segmentation Loss (MASL), which automatically modulates five loss objectives based on per-sample morphological descriptors (tubularity, compactness, irregularity, and scale). Evaluation across 15 datasets spanning 8 modalities demonstrates competitive performance, obtaining the best performance on 14 of 15 datasets, with statistically significant improvements ($p < 0.0033$, Bonferroni-corrected) on 7 challenging tasks (complex morphology, class imbalance, and multi-class segmentation), and clinically meaningful gains ($0.5$--$1.6\\%$ Dice) on 8 mature benchmarks. Notably, S2M-Net achieves $83.43\\%$ Dice on EndoVis17 multiclass instrument segmentation ($+8.69\\%$ over TransUNet and $+9.14\\%$ over the best baseline UMamba at $74.29\\%$), while using $12.8{\\times}$ fewer parameters (4.7M vs.\\ 60M).","absKo":"의료 이미지 분할은 전역 문맥과 계산 효율성의 균형이 요구되며, self-attention mechanism은 이차 복잡도 $\\mathcal{O}((HW)^2 C)$로 인해 병목이 된다. 우리는 Spectral--Spatial Token Mixing (SSTM)을 통해 계산 절감을 달성하는 parameter-efficient architecture인 S2M-Net(4.7M parameters)을 제안한다. SSTM은 효율적인 $\\mathcal{O}(HWC \\log(HW))$ frequency-domain processing과 $\\mathcal{O}(HWCd)$ bottlenecked spatial gating ($d{=}16$)을 결합하여 $\\mathcal{O}(HWC^2)$ 복잡도를 달성한다. 이는 $K{=}32$개의 저주파 성분이 에너지의 $>93\\%$를 포착하는 spectral concentration을 활용한 것으로, $352{\\times}352$ 해상도에서 전체 spectrum의 약 0.8%에 해당한다. 이 설계는 전역 receptive field를 유지하면서 self-attention의 막대한 $\\mathcal{O}((HW)^2C)$ attention map 계산을 피한다. 기하학적 다양성을 처리하기 위해, 우리는 샘플별 morphology descriptor(tubularity, compactness, irregularity, scale)에 따라 다섯 개의 loss objective를 자동으로 조절하는 Morphology-Aware Adaptive Segmentation Loss (MASL)를 도입한다. 8개 modality에 걸친 15개 dataset에서의 평가는 경쟁력 있는 성능을 보이며, 15개 데이터셋 중 14개에서 최고 성능을 달성했다. 또한 7개의 어려운 task(complex morphology, class imbalance, multi-class segmentation)에서는 통계적으로 유의미한 향상($p < 0.0033$, Bonferroni-corrected)을, 8개의 성숙한 benchmark에서는 임상적으로 의미 있는 향상($0.5$--$1.6\\%$ Dice)을 보였다. 특히 S2M-Net은 EndoVis17 multiclass instrument segmentation에서 $83.43\\%$ Dice를 달성했으며($+8.69\\%$ over TransUNet, $+9.14\\%$ over the best baseline UMamba at $74.29\\%$), 이는 12.8배 적은 parameter(4.7M vs. 60M)를 사용하면서 이룬 성과다."},{"id":"62039","en":"Navigating the Flatlands: Dual Adaptive Sharpness-Aware Minimization for Domain Generalization","ko":"Flatlands 항해: 도메인 일반화를 위한 Dual Adaptive Sharpness-Aware Minimization","authors":"Junwen He, Yang He, Lebing Zheng, Zirui Yin, Hong-Yu Zhang, Yulong Wang","pos":"#1200","link":"https://openreview.net/forum?id=kR5qsKt8gz","abs":"

Finding flat minima in the loss landscape is a key strategy for Domain Generalization (DG). However, its effectiveness is often limited by two crucial challenges. 1) Domain Shift: Existing methods like Sharpness-Aware Minimization (SAM) apply a uniform optimization strategy across all domains, overlooking the differences of the learning difficulties among multiple domains and thus performing poorly on challenging domains. 2) Anisotropic Sharpness: By perturbing parameters along a single gradient direction, SAM and its variants ignore multi-directional flatness, making the model converge to minima that remain sharp in other directions. The combined challenges make it more difficult for the model to find truly robust solutions in multi-domain scenarios. To overcome these limitations, we propose the Dual Adaptive Sharpness-Aware Minimization (DA-SAM), which comprises two key modules: Dynamic Adaptive Scaling (DAS) module and Adaptive Multi-Directional Flattening (AMDF) module. First, to tackle the domain shift problem, the DAS module computes the real-time loss on each domain to adaptively generate domain-specific scaling factors that guide the generation of perturbation directions. Second, the AMDF module calculates local flatness by generating multiple directions to simulate perturbations in the parameter space. Based on the learned local flatness metric, it dynamically adjusts the perturbation step size to guide the model parameters to be away from anisotropic sharp regions. Crucially, DAS provides domain-level guidance that makes AMDF’s multi-directional geometric exploration more targeted and effective. Extensive experiments on five DG benchmarks demonstrate the effectiveness of our DA-SAM algorithm.

","absKo":"

loss landscape에서 flat minima를 찾는 것은 Domain Generalization (DG)의 핵심 전략이다. 그러나 그 효과는 두 가지 중요한 도전으로 인해 종종 제한된다. 1) Domain Shift: Sharpness-Aware Minimization (SAM) 같은 기존 방법은 모든 domain에 동일한 optimization strategy를 적용하여, 여러 domain 간 학습 난이도의 차이를 간과하고, 그 결과 어려운 domain에서 성능이 저조하다. 2) Anisotropic Sharpness: SAM과 그 변형은 단일 gradient 방향을 따라 parameter를 perturb하기 때문에, 다방향 flatness를 무시한다. 이로 인해 model은 다른 방향에서는 여전히 sharp한 minima로 수렴할 수 있다. 이러한 결합된 문제는 multi-domain 시나리오에서 model이 진정으로 robust한 solution을 찾는 것을 더 어렵게 만든다. 이러한 한계를 극복하기 위해 우리는 두 개의 핵심 module로 구성된 Dual Adaptive Sharpness-Aware Minimization (DA-SAM)을 제안한다: Dynamic Adaptive Scaling (DAS) module과 Adaptive Multi-Directional Flattening (AMDF) module이다. 먼저 domain shift 문제를 해결하기 위해 DAS module은 각 domain의 실시간 loss를 계산하여 perturbation direction 생성을 안내하는 domain-specific scaling factor를 적응적으로 생성한다. 둘째, AMDF module은 parameter space에서의 perturbation을 모사하기 위해 여러 방향을 생성하여 local flatness를 계산한다. 학습된 local flatness metric을 바탕으로 perturbation step size를 동적으로 조정하여, model parameter가 anisotropic sharp region에서 멀어지도록 유도한다. 특히 DAS는 domain-level guidance를 제공하여 AMDF의 multi-directional geometric exploration을 더 목표 지향적이고 효과적으로 만든다. 5개의 DG benchmark에서 수행한 광범위한 실험은 DA-SAM algorithm의 효과를 입증한다.

"},{"id":"61917","en":"Open-Text Aerial Detection: A Unified Framework For Aerial Visual Grounding And Detection","ko":"Open-Text Aerial Detection: 항공 Visual Grounding 및 Detection을 위한 통합 프레임워크","authors":"Guoting Wei, Xia Yuan, Yu Liu, Yangzhou, Haizhao Jing, Xianbiao Qi, Chunxia Zhao, Haokui Zhang, Rong Xiao","pos":"#1201","link":"https://openreview.net/forum?id=lgzkDjvVwP","abs":"

Open-Vocabulary Aerial Detection (OVAD) and Remote Sensing Visual Grounding (RSVG) have emerged as two key paradigms for aerial scene understanding. However, each paradigm suffers from inherent limitations when operating in isolation: OVAD is restricted to coarse category-level semantics, while RSVG is structurally limited to single-target localization. These limitations prevent existing methods from simultaneously supporting rich semantic understanding and multi-target detection. To address this, we propose OTA-Det, the first unified framework that bridges both paradigms into a cohesive architecture. Specifically, we introduce a task reformulation strategy that unifies task objectives and supervision mechanisms, enabling joint training across datasets from both paradigms with dense supervision signals. Furthermore, we propose a dense semantic alignment strategy that establishes explicit correspondence at multiple granularities, from holistic expressions to individual attributes, enabling fine-grained semantic understanding. To ensure real-time efficiency, OTA-Det builds upon the RT-DETR architecture, extending it from closed-set detection to open-text detection by introducing several high efficient modules, achieving state-of-the-art performance on six benchmarks spanning both OVAD and RSVG tasks while maintaining real-time inference at 34 FPS.

","absKo":"

Open-Vocabulary Aerial Detection (OVAD)와 Remote Sensing Visual Grounding (RSVG)은 aerial scene understanding을 위한 두 가지 핵심 paradigm으로 부상했다. 그러나 각 paradigm은 단독으로 작동할 때 본질적인 한계를 가진다. OVAD는 거친 category-level semantics에 제한되고, RSVG는 구조적으로 single-target localization에 국한된다. 이러한 한계 때문에 기존 방법들은 풍부한 semantic understanding과 multi-target detection을 동시에 지원하지 못한다. 이를 해결하기 위해, 우리는 두 paradigm을 하나의 응집된 architecture로 연결하는 최초의 unified framework인 OTA-Det을 제안한다. 구체적으로, task objective와 supervision mechanism을 통합하는 task reformulation strategy를 도입하여 두 paradigm의 dataset 전반에서 dense supervision signal과 함께 joint training이 가능하도록 한다. 또한 holistic expression부터 개별 attribute까지 여러 granularity에서 explicit correspondence를 구축하는 dense semantic alignment strategy를 제안하여 세밀한 semantic understanding을 가능하게 한다. 실시간 효율성을 보장하기 위해 OTA-Det은 RT-DETR architecture 위에 구축되며, 여러 high efficient module을 도입해 closed-set detection을 open-text detection으로 확장한다. 그 결과 OVAD와 RSVG task를 아우르는 여섯 개 benchmark에서 state-of-the-art 성능을 달성하면서도 34 FPS의 real-time inference를 유지한다.

"},{"id":"60716","en":"Pix2Key: Controllable Open-Vocabulary Retrieval with Semantic Decomposition and Self-Supervised Visual Dictionary Learning","ko":"Pix2Key: 의미 분해와 자기지도 시각 사전 학습을 통한 제어 가능한 Open-Vocabulary 검색","authors":"Guoyizhe Wei, Yang Jiao, Nan Xi, Zhishen Huang, Jingjing Meng, Rama Chellappa, Yan Gao","pos":"#1203","link":"https://openreview.net/forum?id=xFX7jtHi8H","abs":"

Composed image retrieval uses a reference image plus a natural-language edit to retrieve images that apply the requested change while preserving other relevant visual content. Classic fusion pipelines typically rely on supervised triplets and can lose fine-grained cues, while recent zero-shot approaches often caption the reference image and merge the caption with the edit, which may miss implicit user intent and return repetitive results. We present Pix2Key, which represents both queries and candidates as open-vocabulary visual dictionaries, enabling intent-aware constraint matching and diversity-aware reranking in a unified embedding space. A self-supervised pretraining component, V-Dict-AE, further improves the dictionary representation using only images, strengthening fine-grained attribute understanding without CIR-specific supervision. On the extended DeepFashion-MM benchmark, Pix2Key improves Recall@10 up to 3.2 points, and adding V-Dict-AE yields an additional 2.3-point gain while improving intent consistency and maintaining high list diversity.

","absKo":"

Composed image retrieval은 reference image와 자연어 edit를 함께 사용하여, 요청된 변경을 적용하면서도 다른 관련 visual content는 보존한 이미지를 검색한다. 기존 fusion pipeline은 대개 supervised triplet에 의존하며 세밀한 단서를 잃을 수 있는 반면, 최근 zero-shot 접근은 흔히 reference image를 caption으로 변환한 뒤 그 caption을 edit와 결합하는데, 이 과정에서 암묵적인 user intent를 놓치고 반복적인 결과를 반환할 수 있다. 우리는 query와 candidate를 모두 open-vocabulary visual dictionary로 표현하는 Pix2Key를 제안하며, 이를 통해 unified embedding space에서 intent-aware constraint matching과 diversity-aware reranking을 가능하게 한다. self-supervised pretraining component인 V-Dict-AE는 이미지만을 사용해 dictionary representation을 추가로 개선하여, CIR-specific supervision 없이도 fine-grained attribute understanding을 강화한다. extended DeepFashion-MM benchmark에서 Pix2Key는 Recall@10을 최대 3.2포인트 향상시키며, V-Dict-AE를 추가하면 intent consistency를 개선하고 높은 list diversity를 유지하면서 추가로 2.3포인트 향상된다.

"},{"id":"62071","en":"QuantWear: Quantum-scale Wear Particle Detection for Jet Engine Diagnosis","ko":"QuantWear: 제트 엔진 진단을 위한 Quantum-scale Wear Particle 탐지","authors":"Zheng Wang, Yanwen Wang, Tianyu Fang, Jiaxing Shen, Yisen Kang, Di Wu, Yuanqing Zheng","pos":"#1205","link":"https://openreview.net/forum?id=k7SRuKcbD3","abs":"

The quantity and 3-D shape of wear particles are essential indicators for assessing the health of jet engines, enabling early detection of potential damage and preventing accidents caused by catastrophic failures. However, capturing wear particles is difficult due to their minute sizes and ultra high-speed movement within intense jet flows. Existing technologies struggle with the extreme background noise and low resolution in such harsh environments. In this paper, we propose QuantWear, the first quantum sensing system designed to directly quantify and profile wear particles on the sub-millimeter scale. QuantWear innovatively tracks wear particles by monitoring the spectral signatures of Sodium (Na) and Potassium (K) atoms within jet flow, which naturally adhere to particle surfaces due to electrochemical reactions in high-temperature combustion. We construct a custom atomic detector that leverages quantum jump and Faraday rotation effects to isolate these specific atomic signals, effectively suppressing the broad-spectrum flame noise. Next, we apply a deep learning framework to effectively measure the quantity of wear particles in dynamic vaporous backgrounds. Finally, we generate a fully reconstructed 3-D model of the wear particles from multiple 2-D images. Extensive field tests and high-fidelity simulations demonstrate that QuantWear achieves an imaging Signal-to-Noise Ratio (SNR) of 22.5 dB and a 3-D reconstruction similarity exceeding 95%, significantly outperforming state-of-the-art technologies.

","absKo":"

마모 입자의 양과 3-D 형상은 jet engine의 건강 상태를 평가하는 데 필수적인 지표로, 잠재적 손상을 조기에 탐지하고 치명적 고장으로 인한 사고를 예방할 수 있게 한다. 그러나 마모 입자는 매우 작은 크기와 강한 jet flow 내에서의 초고속 이동 때문에 포집하기 어렵다. 기존 기술은 이러한 가혹한 환경에서 극심한 background noise와 낮은 해상도 문제를 극복하는 데 어려움을 겪는다. 본 논문에서는 sub-millimeter scale에서 마모 입자를 직접 정량화하고 프로파일링하도록 설계된 최초의 quantum sensing system인 QuantWear를 제안한다. QuantWear는 고온 연소에서의 electrochemical reaction 때문에 입자 표면에 자연스럽게 부착되는 jet flow 내 Sodium (Na)과 Potassium (K) atom의 spectral signature를 모니터링함으로써 마모 입자를 혁신적으로 추적한다. 우리는 quantum jump와 Faraday rotation effect를 활용하는 맞춤형 atomic detector를 구축하여 이러한 특정 atomic signal을 분리하고, broad-spectrum flame noise를 효과적으로 억제한다. 다음으로, dynamic vaporous background에서 마모 입자의 양을 효과적으로 측정하기 위해 deep learning framework를 적용한다. 마지막으로, 여러 2-D image로부터 마모 입자의 fully reconstructed 3-D model을 생성한다. 광범위한 현장 시험과 고정밀 시뮬레이션은 QuantWear가 22.5 dB의 imaging Signal-to-Noise Ratio (SNR)과 95%를 초과하는 3-D reconstruction similarity를 달성하여, 최신 기술을 크게 능가함을 보여준다.

"},{"id":"66578","en":"Refining Context-Entangled Content Segmentation via Curriculum Selection and Anti-Curriculum Promotion","ko":"커리큘럼 선택과 반커리큘럼 촉진을 통한 맥락 얽힘 콘텐츠 분할 정제","authors":"Chunming He, Rihan Zhang, Fengyang Xiao, Dingming Zhang, Zhiwen Cao, Sina Farsiu","pos":"#1207","link":"https://openreview.net/forum?id=2TCOninOND","abs":"

Biological learning proceeds from easy to difficult tasks, gradually reinforcing perception and robustness. Inspired by this principle, we address Context‑Entangled Content Segmentation (CECS)—a challenging setting where objects share intrinsic visual patterns with their surroundings, as in camouflaged object detection. Conventional segmentation networks predominantly rely on architectural enhancements but often ignore the learning dynamics that govern robustness under entangled data distributions. We introduce CurriSeg, a dual‑phase learning framework that unifies curriculum and anti‑curriculum principles to improve representation reliability. In the Curriculum Selection phase, CurriSeg dynamically selects training data based on the temporal statistics of sample losses, distinguishing hard‑but‑informative samples from noisy or ambiguous ones, thus enabling stable capability enhancement. In the Anti‑Curriculum Promotion phase, we design Spectral‑Blindness Fine‑Tuning, which suppresses high‑frequency components to enforce dependence on low‑frequency structural and contextual cues. Combined with Sharpness‑Aware Minimization, this promotes convergence to flatter minima and strengthens generalization. Extensive experiments demonstrate that CurriSeg achieves consistent improvements across diverse CECS benchmarks without adding parameters or increasing total training time, offering a principled view of how progression and challenge interplay to foster robust and context‑aware segmentation. Code will be released.

","absKo":"

생물학적 학습은 쉬운 task에서 어려운 task로 진행되며, 점차 perception과 robustness를 강화한다. 이러한 원리에 착안하여, 우리는 Context‑Entangled Content Segmentation (CECS)에 대응한다. 이는 camouflaged object detection처럼 객체가 주변 환경과 본질적인 visual pattern을 공유하는 까다로운 설정이다. 기존 segmentation network는 주로 architecture enhancement에 의존하지만, 얽힌 data distribution 하에서 robustness를 좌우하는 learning dynamics는 종종 간과한다. 우리는 representation reliability를 향상시키기 위해 curriculum 원리와 anti-curriculum 원리를 통합한 dual-phase learning framework인 CurriSeg를 도입한다. Curriculum Selection phase에서 CurriSeg는 sample loss의 temporal statistics를 기반으로 training data를 동적으로 선택하여, hard-but-informative sample과 noisy하거나 ambiguous한 sample을 구분함으로써 안정적인 capability 향상을 가능하게 한다. Anti-Curriculum Promotion phase에서는 고주파 성분을 억제하여 low-frequency structural 및 contextual cue에 의존하도록 만드는 Spectral-Blindness Fine-Tuning을 설계한다. Sharpness-Aware Minimization과 결합하면, 이는 더 flat한 minimum으로의 수렴을 촉진하고 generalization을 강화한다. 광범위한 실험은 CurriSeg가 parameter를 추가하거나 total training time을 늘리지 않으면서 다양한 CECS benchmark 전반에서 일관된 성능 향상을 달성함을 보여주며, progression과 challenge가 어떻게 상호작용하여 robust하고 context-aware한 segmentation을 이끄는지에 대한 원리적 관점을 제시한다. 코드는 공개될 예정이다.

"},{"id":"64003","en":"SHERPA: Fine-tuning Segment Anything Models with Task-relevant Guidance","ko":"SHERPA: 작업 관련 가이던스로 Segment Anything Models 미세조정하기","authors":"Jingcheng Xie, Yinda Chen, Xiaoyu Liu, Haoyuan Shi, Zhiwei Xiong","pos":"#1210","link":"https://openreview.net/forum?id=RrZAajAWHr","abs":"

Segment Anything Models (SAMs) often struggle with certain specialized tasks. A common approach is to fine-tune models with specific task labels, but this often leads to overfitting, introduces model bias and significantly degrades their generalization ability. To overcome these challenges, we propose SHERPA, a novel framework that leverages a smaller SAM to guide the fine-tuning of a larger SAM via task-relevant features. Specifically, we first leverage the Fisher Ratio Separation (FRS) module to separate high task-relevant features and preserve the ability of the large SAM to perform other general tasks. Then, the Guiding Feature Extraction (GFE) module is used to extract representative guiding features from the fine-tuned small SAMs. We leverage small SAMs tailored for specific tasks (including natural image segmentation, biomedical image segmentation, and video object segmentation) as guidance and then evaluate the SHERPA scheme to fine-tune larger SAM series models. Our experiments demonstrate that SHERPA enhances the retention of generalization ability across those diverse tasks, by up to 11.1\\%, and improves specific task performance by up to 2.2\\%.

","absKo":"

Segment Anything Model(SAM)은 종종 특정 specialized task에서 어려움을 겪는다. 일반적인 접근법은 특정 task label로 model을 fine-tuning하는 것이지만, 이는 과적합을 초래하고 model bias를 유발하며 generalization 능력을 크게 저하시킨다. 이러한 문제를 극복하기 위해 우리는 task-relevant feature를 통해 작은 SAM이 더 큰 SAM의 fine-tuning을 guide하도록 하는 새로운 framework SHERPA를 제안한다. 구체적으로, 우리는 먼저 Fisher Ratio Separation(FRS) module을 활용해 task 관련성이 높은 feature를 분리하고, 큰 SAM이 다른 general task를 수행할 수 있는 능력을 보존한다. 그런 다음 Guiding Feature Extraction(GFE) module을 사용해 fine-tuned small SAM으로부터 대표적인 guiding feature를 추출한다. 우리는 특정 task에 맞게 조정된 small SAM(자연 이미지 segmentation, biomedical image segmentation, video object segmentation 포함)을 guidance로 활용한 뒤, SHERPA scheme을 사용해 더 큰 SAM series model을 fine-tuning하고 평가한다. 실험 결과, SHERPA는 다양한 task 전반에서 generalization 능력의 보존을 최대 11.1\\%까지 향상시키고, specific task 성능을 최대 2.2\\%까지 개선함을 보여준다.

"},{"id":"64967","en":"Spatially-Regularized Entropy for Discriminative Token Merging in Fine-Grained Re-Identification","ko":"세밀한 Re-Identification에서 Discriminative Token Merging을 위한 Spatially-Regularized Entropy","authors":"Shangze Li, Yifan Xu, Jingmiao Liang, Yongfei Zhang, Yuzhuo Ma, Yingbo Qu","pos":"#1214","link":"https://openreview.net/forum?id=IJ2C56ra1a","abs":"

While Vision Transformers (ViTs) offer strong global modeling, their quadratic computational cost limits utility in latency-sensitive applications like person re-identification (ReID). Existing compression strategies, such as token pruning or generic merging, typically rely on coarse-grained criteria tailored for image classification. In fine-grained retrieval, these approaches often discard or smooth out subtle but discriminative local details. To resolve this, we propose SRE-Merge, a training-free framework designed for discriminative token compression. SRE-Merge injects spatial priors into the merging process through three mechanisms: (i) Spatial-Entropy Saliency Assessment (SES-Assess), which quantifies token importance as Spatial-Entropic Mass (SE-Mass) by coupling spatial structure with local attention entropy; (ii) Hybrid Context-Affinity Matching (HCA-Match), which guides precise pair selection by combining feature similarity with mass-derived context; and (iii) Energy-Preserving Weighted Fusion (EPW-Fuse), which incorporates SE-Mass weighting to counteract feature variance reduction. Extensive experiments on standard benchmarks show that SRE-Merge reduces GFLOPs of the base ViT model by about 24\\% while maintaining state-of-the-art accuracy, establishing a superior accuracy-efficiency trade-off.

","absKo":"

Vision Transformer (ViT)는 강력한 global modeling을 제공하지만, quadratic computational cost 때문에 person re-identification (ReID) 같은 latency-sensitive application에서 활용성이 제한된다. token pruning이나 generic merging과 같은 기존 compression 전략은 대체로 image classification에 맞춘 coarse-grained criteria에 의존한다. fine-grained retrieval에서는 이러한 접근이 미세하지만 구별력 있는 local detail을 종종 버리거나 흐리게 만든다. 이를 해결하기 위해, 우리는 discriminative token compression을 위해 설계된 training-free framework인 SRE-Merge를 제안한다. SRE-Merge는 세 가지 메커니즘을 통해 merging 과정에 spatial prior를 주입한다: (i) Spatial-Entropy Saliency Assessment (SES-Assess)는 spatial structure와 local attention entropy를 결합하여 token importance를 Spatial-Entropic Mass (SE-Mass)로 정량화하고, (ii) Hybrid Context-Affinity Matching (HCA-Match)는 feature similarity와 mass-derived context를 결합해 정확한 pair selection을 유도하며, (iii) Energy-Preserving Weighted Fusion (EPW-Fuse)는 feature variance reduction을 상쇄하기 위해 SE-Mass weighting을 도입한다. 표준 benchmark 전반의 광범위한 실험 결과, SRE-Merge는 base ViT model의 GFLOPs를 약 24\\% 줄이면서도 state-of-the-art accuracy를 유지하여, 우수한 accuracy-efficiency trade-off를 입증한다.

"},{"id":"64264","en":"Light Up Your Face: A Physically Consistent Dataset and Diffusion Model for Face Fill-Light Enhancement","ko":"당신의 얼굴을 밝히다: 얼굴 필 라이트 개선을 위한 물리적으로 일관된 데이터셋과 Diffusion 모델","authors":"Jue Gong, Zihan Zhou, Jingkai Wang, Xiaohong Liu, Yulun Zhang, Xiaokang Yang","pos":"#1301","link":"https://openreview.net/forum?id=PLT2FKIs2c","abs":"

Face fill-light enhancement (FFE) brightens underexposed faces by adding virtual fill light while keeping the original scene illumination and background unchanged. Most face relighting methods aim to reshape overall lighting, which can suppress the input illumination or modify the entire scene, leading to foreground–background inconsistency and mismatching practical FFE needs. To support scalable learning, we introduce LightYourFace-160K (LYF-160K), a large-scale paired dataset built with a physically consistent renderer that injects a disk-shaped area fill light controlled by six disentangled factors, producing 160K before-and-after pairs. We first pretrain a physics-aware lighting prompt (PALP) that embeds the 6D parameters into diffusion-compatible conditioning tokens, using an auxiliary planar-light reconstruction objective. Building on a pretrained diffusion backbone, we then train FiLitDiff, an efficient one-step model conditioned on these physically grounded lighting codes, enabling fast, controllable, and high-fidelity fill lighting at low computational cost. Experiments on held-out paired sets demonstrate strong perceptual quality and competitive full-reference metrics, while better preserving background illumination. The dataset and model will be released.

","absKo":"

Face fill-light enhancement (FFE)는 원래 장면의 조명과 배경을 그대로 유지하면서 가상 fill light를 추가해 노출 부족 얼굴을 밝게 만든다. 대부분의 face relighting 방법은 전체 조명을 재구성하는 데 초점을 맞추기 때문에, 입력 조명을 억제하거나 장면 전체를 수정하여 foreground–background 불일치를 초래하고 실제 FFE 요구와 맞지 않는다. 확장 가능한 learning을 지원하기 위해, 우리는 물리적으로 일관된 renderer로 구축한 대규모 paired dataset LightYourFace-160K (LYF-160K)를 소개한다. 이 데이터셋은 여섯 개의 disentangled factor로 제어되는 disk-shaped area fill light를 주입하여 16만 개의 before-and-after pair를 생성한다. 먼저, 우리는 6D parameter를 diffusion-compatible conditioning token에 임베딩하는 physics-aware lighting prompt (PALP)를 사전학습하며, 보조 planar-light reconstruction objective를 사용한다. 이어서 사전학습된 diffusion backbone을 기반으로, 물리적으로 정립된 lighting code에 조건화된 효율적인 one-step model FiLitDiff를 학습한다. 이를 통해 낮은 계산 비용으로 빠르고 제어 가능하며 고충실도의 fill lighting을 구현한다. 보류된 paired set에서의 실험은 배경 조명을 더 잘 보존하면서도 강한 perceptual quality와 경쟁력 있는 full-reference metric을 보여준다. 데이터셋과 모델은 공개될 예정이다.

"},{"id":"64779","en":"Large-capacity and Receiver Authenticable Generative Image Steganography","ko":"대용량 및 수신자 인증 가능한 Generative Image Steganography","authors":"Jiannian Wang, Yao Lu, Guangming Lu","pos":"#1302","link":"https://openreview.net/forum?id=Jui4miI76g","abs":"

Diffusion-based generative image steganography converts the input single secret image into noise, and generates the stego image with it serves as the initial noise. Nevertheless, existing methods exhibit three severe limitations: (1) the fixed hiding space constrains their capacity to one secret image; (2) severe inter-secret interference arising from substantial information divergence among multiple secret images while concealing them within a shared hiding space; (3) security risks owing to the absence of the receiver-side verification mechanism. To systematically address these issues, this paper proposes a novel Receiver Authenticable Generative Image Steganography framework based on diffusion models. We introduce a Dynamic Cover Selection and Optimization Engine to adaptively allocate suitable hiding spaces for different secret images. This design permits the concealment of disparate secret images (or fragments of a single image) into separate spaces, enabling dynamic multi-image concealment while effectively preventing inter-secret interference and expanding capacity through the enlarged hiding spaces. Furthermore, a Signature Authentication Controller cryptographically signs the secret container after concealing and verifies it before extraction, ensuring secure receiver isolation and precise localization of the secret data container. Experiments demonstrate that the proposed framework achieves superior secure multi-receiver isolation and high-performance generative image steganography with large capacity.

","absKo":"

diffusion-based generative image steganography는 입력된 단일 secret image를 noise로 변환하고, 이를 initial noise로 사용하여 stego image를 생성한다. 그럼에도 기존 방법은 다음 세 가지 심각한 한계를 보인다: (1) 고정된 hiding space는 capacity를 하나의 secret image로 제한한다; (2) 여러 secret image를 shared hiding space에 숨길 때 secret들 간의 substantial information divergence로 인해 심각한 inter-secret interference가 발생한다; (3) receiver-side verification mechanism의 부재로 인한 security risk가 존재한다. 이러한 문제를 체계적으로 해결하기 위해, 본 논문은 diffusion models에 기반한 새로운 Receiver Authenticable Generative Image Steganography framework를 제안한다. 우리는 서로 다른 secret image에 적합한 hiding space를 적응적으로 할당하는 Dynamic Cover Selection and Optimization Engine를 도입한다. 이 설계는 서로 다른 secret image(또는 하나의 image의 fragments)를 별도의 space에 은닉할 수 있게 하여, inter-secret interference를 효과적으로 방지하면서 enlarged hiding space를 통해 capacity를 확장하는 dynamic multi-image concealment를 가능하게 한다. 또한 Signature Authentication Controller는 은닉 후 secret container에 cryptographic signing을 수행하고 extraction 전에 이를 검증함으로써, secure receiver isolation과 secret data container의 precise localization을 보장한다. 실험은 제안한 framework가 large capacity를 갖는 high-performance generative image steganography와 superior secure multi-receiver isolation을 달성함을 보여준다.

"},{"id":"66776","en":"Is Training Necessary for Anomaly Detection?","ko":"이상 탐지에 학습이 필요한가?","authors":"Xingwu Zhang, Guanxuan Li, Paul Henderson, Gerardo Aragon-Camarasa, ZIJUN LONG","pos":"#1303","link":"https://openreview.net/forum?id=0Qudez0gLp","abs":"

Current state-of-the-art multi-class unsupervised anomaly detection (MUAD) methods rely on training encoder–decoder models to reconstruct anomaly-free features. We first show these approaches have an inherent fidelity–stability dilemma in how they detect anomalies via reconstruction residuals. We then abandon the reconstruction paradigm entirely and propose Retrieval-based Anomaly Detection (RAD). RAD is a training-free approach that stores anomaly-free features in a memory and detects anomalies through multi-level retrieval, matching test patches against the memory. Experiments demonstrate that RAD achieves state-of-the-art performance across four established benchmarks (MVTec-AD, VisA, Real-IAD, 3D-ADAM) under both standard and few-shot settings. On MVTec-AD, RAD reaches 96.7% Pixel AUROC with just a single anomaly-free image compared to 98.5% of RAD's full-data performance. We further prove that retrieval-based scores theoretically upper-bound reconstruction-residual scores. Collectively, these findings overturn the assumption that MUAD requires task-specific training, showing that state-of-the-art anomaly detection is feasible with memory-based retrieval. Our code is available at https://anonymous.4open.science/r/RAD-3A6C.

","absKo":"

현재 state-of-the-art multi-class unsupervised anomaly detection(MUAD) 방법은 encoder–decoder model을 학습시켜 anomaly-free feature를 재구성하는 데 의존한다. 우리는 먼저 이러한 접근법이 reconstruction residual을 통해 anomaly를 검출하는 방식에서 본질적인 fidelity–stability 딜레마를 지닌다는 점을 보인다. 그런 다음 reconstruction paradigm 자체를 완전히 버리고 Retrieval-based Anomaly Detection(RAD)을 제안한다. RAD는 anomaly-free feature를 memory에 저장하고, test patch를 memory와 매칭하는 multi-level retrieval을 통해 anomaly를 검출하는 training-free 접근법이다. 실험 결과, RAD는 standard 및 few-shot 설정 모두에서 네 개의 확립된 benchmark(MVTec-AD, VisA, Real-IAD, 3D-ADAM) 전반에 걸쳐 state-of-the-art 성능을 달성한다. MVTec-AD에서 RAD는 anomaly-free image가 단 1장뿐인 경우에도 96.7% Pixel AUROC에 도달하며, 이는 RAD의 full-data 성능인 98.5%와 비교할 만한 수준이다. 우리는 또한 retrieval-based score가 이론적으로 reconstruction-residual score의 상한을 이룬다는 점을 증명한다. 종합하면, 이러한 결과는 MUAD가 task-specific training을 필요로 한다는 가정을 뒤집으며, memory-based retrieval만으로도 state-of-the-art anomaly detection이 가능함을 보여준다. 코드는 https://anonymous.4open.science/r/RAD-3A6C 에서 제공된다.

"},{"id":"64697","en":"Fully Zero-Shot Image Dehazing","ko":"완전 Zero-Shot 이미지 Dehazing","authors":"Shuocheng Wang, Ruoxi Zhu, Jiaming Liu, Zhengyang Cao, Kun Wang, Chengkang Huang, Lizhuo Liu, Shichen Peng, Minge Jing, Yibo Fan","pos":"#1308","link":"https://openreview.net/forum?id=KjRmj39WfF","abs":"

Image dehazing, an important image restoration problem, aims to recover clear scene content from images degraded by atmospheric haze. Existing dehazing methods rely on observing the distribution of hazy images during training: supervised approaches typically depend on synthetic datasets, leading to poor generalization in real-world scenarios; unsupervised methods are constrained by the limited diversity of observed haze conditions due to the difficulty of collecting real hazy images, and fail to generalize to unseen haze types. To address these challenges, we propose the first fully zero-shot dehazing framework that is trained without any hazy images. The framework is built upon a set of representations that remain invariant across clean and hazy images to bridge the two domains, which is both theoretically derived and experimentally validated. Consequently, we formulate dehazing as a conditional generative modeling problem and train a diffusion model solely with the invariant representations of the abundant and readily available clean images. During testing, the same representations extracted from hazy images serve as the conditional input to guide the diffusion process toward the clean image distribution. Quantitative analyses verify the effectiveness of the proposed representations, and extensive experiments across various real-world hazy datasets demonstrate our framework’s remarkable generalization ability, significantly outperforming existing methods. Our code will be available after the review process.

","absKo":"

Image dehazing은 대기 중 haze로 인해 열화된 이미지에서 선명한 장면 내용을 복원하는 것을 목표로 하는 중요한 image restoration 문제이다. 기존 dehazing method는 학습 시 hazy image의 분포를 관찰하는 데 의존한다. supervised approach는 일반적으로 synthetic dataset에 의존하므로 실제 환경에서의 generalization이 좋지 않고, unsupervised method는 실제 hazy image를 수집하기 어렵기 때문에 관측된 haze condition의 다양성이 제한되며, 보지 못한 haze type으로 일반화하는 데 실패한다. 이러한 문제를 해결하기 위해, 우리는 어떠한 hazy image도 사용하지 않고 학습하는 최초의 fully zero-shot dehazing framework를 제안한다. 이 framework는 clean image와 hazy image 사이에서 불변으로 유지되는 일련의 representation 위에 구축되어 두 domain을 연결하며, 이는 이론적으로 도출되고 실험적으로 검증되었다. 그 결과 dehazing을 conditional generative modeling 문제로 정식화하고, 풍부하고 쉽게 확보할 수 있는 clean image의 invariant representation만으로 diffusion model을 학습한다. 테스트 시에는 hazy image에서 추출한 동일한 representation이 conditional input으로 사용되어 diffusion process를 clean image distribution 쪽으로 유도한다. 정량 분석은 제안한 representation의 효과를 검증하며, 다양한 실제 haze dataset에 대한 광범위한 실험은 우리의 framework가 기존 method를 크게 능가하는 뛰어난 generalization 능력을 보임을 입증한다. 우리의 코드는 review process 이후 공개될 예정이다.

"},{"id":"60657","en":"Expandable, Compressible, Mineable: Open-World Thermal Infrared Image Restoration","ko":"확장 가능하고, 압축 가능하고, 마이닝 가능한 Open-World 열적외선 이미지 복원","authors":"Pu Li, Huafeng Li, Yafei Zhang, Wen Wang, Neng Dong, Jie Wen","pos":"#1311","link":"https://openreview.net/forum?id=xyKr7qopq6","abs":"

In open-world settings, thermal infrared (TIR) image degradations continuously emerge and evolve, while most existing all-in-one restoration methods are built on a closed-set assumption and struggle to continually adapt to novel degradations. To address this, we propose ECMRNet, an Expandable, Compressible, and Mineable Restoration Network for open-world TIR restoration from the perspective of continual degradation learning. Conceptually, ECMRNet unifies continual degradation learning as an ``expand–compress–mine'' closed-loop process, enabling sustained adaptation to new degradations with controllable evolution. Structurally, ECMRNet decomposes intermediate representations into group-isolated subspaces, and achieves strict parameter isolation and fast adaptation to new degradations by freezing historical groups and isomorphically expanding new ones. To curb model growth as tasks accumulate, we present Structural Entropy Pruning, which identifies and removes redundant channel groups via two-dimensional structural entropy minimization, achieving information contribution–driven adaptive compression. Moreover, we design a Sub-degradation Knowledge Mining module that dynamically retrieves and recombines transferable components from historical representations to improve restoration under compound degradations. Experimental results demonstrate that ECMRNet achieves superior overall performance across diverse single and compound degradations while using fewer parameters and lower computational cost, highlighting its scalability and efficiency in open-world TIR restoration.

","absKo":"

open-world 설정에서는 thermal infrared(TIR) 이미지 열화가 지속적으로 발생하고 진화하는 반면, 기존의 대부분의 all-in-one restoration 방법은 closed-set 가정에 기반하고 있어 새로운 열화에 지속적으로 적응하는 데 어려움을 겪는다. 이를 해결하기 위해 우리는 continual degradation learning 관점에서 open-world TIR restoration을 위한 Expandable, Compressible, and Mineable Restoration Network인 ECMRNet을 제안한다. 개념적으로 ECMRNet은 continual degradation learning을 ``expand–compress–mine'' closed-loop process로 통합하여, 제어 가능한 진화와 함께 새로운 열화에 지속적으로 적응할 수 있게 한다. 구조적으로 ECMRNet은 intermediate representation을 group-isolated subspace로 분해하고, 역사적 그룹을 고정한 채 새로운 그룹을 isomorphic하게 확장함으로써 새로운 열화에 대한 엄격한 parameter isolation과 빠른 적응을 달성한다. 작업이 누적되면서 model growth를 억제하기 위해, 우리는 2차원 structural entropy minimization을 통해 중복 channel group을 식별하고 제거하는 Structural Entropy Pruning을 제시하며, 정보 기여도 기반의 적응적 압축을 달성한다. 더 나아가, 우리는 과거 representation에서 transfer 가능한 구성요소를 동적으로 검색하고 재조합하여 compound degradation 하에서의 restoration을 향상시키는 Sub-degradation Knowledge Mining 모듈을 설계한다. 실험 결과는 ECMRNet이 다양한 단일 및 compound degradation 전반에서 더 적은 parameter와 더 낮은 계산 비용을 사용하면서도 우수한 전체 성능을 달성함을 보여주며, open-world TIR restoration에서의 확장성과 효율성을 강조한다.

"},{"id":"63980","en":"Deep Residual Injection for Full-Spectrum Forensic Signal Perception in Multimodal Large Language Models","ko":"멀티모달 Large Language Models에서 전 스펙트럼 포렌식 신호 인식을 위한 깊은 잔차 주입","authors":"Kaiqing Lin, Zhiyuan Yan, Ruoxin Chen, Ke-Yue Zhang, Yue Zhou, Caiyong Piao, Bin Li, Taiping Yao, Bo Wang, Youchang xiao, Shouhong Ding","pos":"#1314","link":"https://openreview.net/forum?id=S2ilvpV9du","abs":"

Multimodal large language models (MLLMs) have been increasingly adopted in forensics for their robust semantic understanding. As AI-generated images become realistic, semantic-level inconsistencies alone are often insufficient for reliable detection. This motivates a critical question: whether MLLMs can achieve full-spectrum forensic signal perception, i.e., capturing low-level generator artifacts without sacrificing pre-trained semantic knowledge. We then conduct a layer-wise analysis of forensic signal perception in MLLMs and find that semantic information is mainly encoded in the early-to-middle layers, and directly fine-tuning MLLMs for artifact learning causes rapid semantic forgetting. Based on this insight, we propose Deep Visual Residual MLLM (Deep-VRM) to \\textit{preserve early semantic processing while injecting artifact-specific visual signals as a residual path into an intermediate layer}, where they are fused with semantic token representations and propagated through subsequent trainable layers. This enables later layers to jointly model semantic reasoning and signal-level forensic cues, and surprisingly, the model learns to adaptively leverage different levels of forensic signals depending on the input, achieving robust and generalizable detection performance. Extensive experiments show that our method achieves state-of-the-art across all benchmarks.

","absKo":"Multimodal large language models (MLLMs)는 강력한 semantic understanding 때문에 forensic 분야에서 점점 더 많이 채택되고 있다. AI-generated image가 현실적으로 변함에 따라, semantic-level inconsistency만으로는 신뢰할 수 있는 탐지에 충분하지 않은 경우가 많다. 이는 중요한 질문을 제기한다. MLLM이 full-spectrum forensic signal perception, 즉 사전 학습된 semantic knowledge를 희생하지 않으면서 low-level generator artifact를 포착할 수 있는가? 우리는 이어서 MLLM의 forensic signal perception을 layer-wise로 분석하고, semantic information이 주로 early-to-middle layer에 인코딩되며, artifact learning을 위해 MLLM을 직접 fine-tuning하면 semantic forgetting이 빠르게 발생한다는 사실을 발견한다. 이러한 통찰에 기반하여 우리는 Deep Visual Residual MLLM (Deep-VRM)을 제안한다. 이는 early semantic processing을 보존하면서 artifact-specific visual signal을 intermediate layer로 residual path 형태로 주입하고, 그 지점에서 semantic token representation과 fusion된 뒤 이후의 trainable layer를 통해 전파되도록 한다. 이를 통해 이후 layer는 semantic reasoning과 signal-level forensic cue를 함께 모델링할 수 있으며, 놀랍게도 모델은 입력에 따라 서로 다른 수준의 forensic signal을 적응적으로 활용하는 법을 학습하여 robust하고 generalizable한 detection 성능을 달성한다. 광범위한 실험은 우리의 방법이 모든 benchmark에서 state-of-the-art를 달성함을 보여 준다.

"},{"id":"65143","en":"Butterworth as Attention: Anisotropic Spectral Gating for Pansharpening","ko":"Attention으로서의 Butterworth: Pansharpening을 위한 이방성 스펙트럼 게이팅","authors":"Zhenggang Wang, Wang Wu, Lianghuazhe, Tai-Xiang Jiang","pos":"#1411","link":"https://openreview.net/forum?id=GNrBEnHPPv","abs":"

Pansharpening fuses high-resolution panchromatic (PAN) images with low-resolution multispectral (LMS) images. For spatial-spectral fusion, Fast Fourier Transform (FFT)-based methods provide a global receptive field to capture long-range dependencies and naturally separate frequency components. However, most existing approaches directly transplant spatial operators like convolution or self-attention, while disregarding the fundamental structure of the spectrum: a strict spatial correspondence where each coordinate represents a specific frequency component, and a highly non-uniform, radially decaying energy distribution. To address this, we revisit the classical Butterworth filter, a frequencydomain operator defined directly on spectral coordinates that is inherently suited for processing such structured representations. We generalize the standard isotropic Butterworth filter into an anisotropic, learnable frequency-domain gating mechanism, establishing an efficient alternative to self-attention, and propose the Anisotropic Butterworth Fusion Network (ABFNet). Its core is a novel dual-branch gating module that employs learnable anisotropic Butterworth filters to perform adaptive direction-aware feature selection, integrating global context and local details with linear complexity. Extensive experiments show that ABFNet achieves state-of-the-art (SOTA) performance on pansharpening benchmarks with low computational overhead. Furthermore, its superior accuracy on CIFAR-100 classification validates the broader applicability of this frequency-domain learning paradigm.

","absKo":"

Pansharpening은 고해상도 panchromatic (PAN) 이미지와 저해상도 multispectral (LMS) 이미지를 결합합니다. 공간-스펙트럼 fusion을 위해 Fast Fourier Transform (FFT) 기반 방법은 장거리 의존성을 포착할 수 있는 global receptive field를 제공하고 주파수 성분을 자연스럽게 분리합니다. 그러나 기존의 대부분 방법은 convolution이나 self-attention 같은 spatial operator를 그대로 이식하면서, spectrum의 근본적인 구조를 간과합니다. 즉, 각 좌표가 특정 frequency component를 나타내는 엄격한 spatial correspondence와, 에너지가 반경 방향으로 감소하는 매우 비균일한 분포가 그것입니다. 이를 해결하기 위해, 우리는 이러한 구조화된 표현 처리에 본질적으로 적합하도록 spectral coordinates에 직접 정의된 frequency-domain operator인 고전적 Butterworth filter를 다시 살펴봅니다. 우리는 표준 isotropic Butterworth filter를 anisotropic하고 learnable한 frequency-domain gating mechanism으로 일반화하여 self-attention의 효율적인 대안임을 확립하고, Anisotropic Butterworth Fusion Network (ABFNet)를 제안합니다. 그 핵심은 learnable anisotropic Butterworth filter를 사용해 적응적인 direction-aware feature selection을 수행하는 새로운 dual-branch gating module로, linear complexity로 global context와 local details를 통합합니다. 광범위한 실험은 ABFNet이 낮은 계산 오버헤드로 pansharpening benchmark에서 state-of-the-art (SOTA) 성능을 달성함을 보여줍니다. 또한 CIFAR-100 분류에서의 우수한 정확도는 이 frequency-domain learning paradigm의 더 넓은 적용 가능성을 검증합니다.

"},{"id":"65002","en":"Explainable Forensics of Manipulated Segments in Untrimmed Long Videos","ko":"미편집 긴 비디오에서 조작된 구간의 설명 가능한 포렌식","authors":"Yue Feng, Wei Ji, Qijia Lu, Jingrou Zhang, Fei Shen, Jingjing Li, Xiao Li, Yizhen Jia, Qiang Chen, Limin Wang, Wentong Li, Jie Qin","pos":"#2411","link":"https://openreview.net/forum?id=Hu8tqsCnEn","abs":"

The rapid advancement of AI-driven video generation has transformed content creation, while simultaneously increasing the risk of misinformation through localized manipulations in long-form videos. Existing video forensic methods predominantly operate on short, independent clips, and thus fail to capture realistic scenarios where AI-generated content is sparsely embedded within otherwise authentic footage. To bridge this gap, we formulate the task of Temporal AI-Generated Segment Localization and Explanation, which targets authenticity detection, temporal localization, and interpretable analysis of manipulated segments in untrimmed long videos. We further introduce TASLE, a large-scale benchmark comprising 12,472 untrimmed videos with diverse manipulation patterns and rich annotation signals, including temporal boundaries, authenticity labels, and segment-level rationales. In addition, we propose MSLoc, a coarse-to-fine forensic baseline that combines a boundary-sensitive proposal generation module for efficient long-video scanning with an MLLM-based refinement module for precise boundary localization and interpretable reasoning. Experiments validate the effectiveness of the proposed baseline, highlighting the importance of segment-level explainable forensics for long-form AI-generated video analysis. Dataset and code will be made publicly available.

","absKo":"

AI 기반 video generation의 급속한 발전은 콘텐츠 제작을 변화시켰지만, 동시에 장편 비디오 내의 국소적 조작을 통한 허위정보 위험도 증가시켰다. 기존 video forensic method는 주로 짧고 독립적인 clip에서 동작하므로, AI 생성 콘텐츠가 진짜 영상 속에 드물게 삽입된 현실적인 상황을 포착하지 못한다. 이러한 간극을 메우기 위해 우리는 Temporal AI-Generated Segment Localization and Explanation task를 정식화한다. 이는 편집되지 않은 긴 비디오에서 authenticity detection, temporal localization, 그리고 조작된 segment의 해석 가능한 분석을 목표로 한다. 또한 우리는 다양한 조작 패턴과 풍부한 annotation signal을 포함하는 12,472개의 untrimmed video로 구성된 대규모 benchmark인 TASLE을 소개한다. 여기에는 temporal boundary, authenticity label, segment-level rationale이 포함된다. 더 나아가 우리는 boundary-sensitive proposal generation module을 이용해 효율적인 장편 비디오 스캔을 수행하고, MLLM-based refinement module을 통해 정확한 boundary localization과 해석 가능한 추론을 수행하는 coarse-to-fine forensic baseline인 MSLoc을 제안한다. 실험 결과는 제안한 baseline의 효과를 검증하며, 장편 AI 생성 비디오 분석에서 segment-level explainable forensics의 중요성을 강조한다. Dataset과 code는 공개될 예정이다.

"},{"id":"63320","en":"Unifying Deep Stochastic Processes for Image Enhancement","ko":"이미지 향상을 위한 심층 확률 과정 통합","authors":"Wojciech Kozłowski, Radosław Kuczbański, Kamil Adamczewski, Karol Szczypkowski, Maciej Zieba","pos":"#2709","link":"https://openreview.net/forum?id=Y7szlTydrp","abs":"

Deep stochastic processes have recently become a central paradigm for image enhancement, with many methods explicitly conditioning the stochastic trajectory on the degraded input. However, the relationship between these conditional processes and standard diffusion models remains unclear. In this work, we introduce a unified perspective on stochastic image enhancement by classifying recent methods into three families of continuous-time processes: unconditional diffusion models, Ornstein–Uhlenbeck (OU) processes, and diffusion bridges. We show that all of these approaches arise from a common stochastic differential equation (SDE) formulation. This framework makes explicit that seemingly disparate methods differ primarily in their drift and diffusion terms, terminal distributions, and boundary conditions, while schedulers and samplers constitute orthogonal design choices. Leveraging this unification, we conduct a controlled empirical study across multiple image enhancement tasks using identical architectures and training protocols. Our results reveal no consistently dominant method; instead, we identify and disentangle the specific design choices that most strongly influence performance. Finally, we release ItoVision, a modular PyTorch library that implements the unified framework and enables rapid prototyping and fair comparison of stochastic image enhancement methods.

","absKo":"

Deep stochastic process는 최근 이미지 향상(image enhancement)의 중심 패러다임으로 떠올랐으며, 많은 방법이 stochastic trajectory를 열화된 입력에 명시적으로 조건화한다. 그러나 이러한 conditional process와 표준 diffusion model 사이의 관계는 아직 명확하지 않다. 본 연구에서는 최근 방법들을 세 가지 continuous-time process 계열, 즉 unconditional diffusion model, Ornstein–Uhlenbeck (OU) process, diffusion bridge로 분류함으로써 stochastic image enhancement에 대한 통합적 관점을 제시한다. 우리는 이 모든 접근이 공통의 stochastic differential equation (SDE) formulation에서 비롯됨을 보인다. 이 프레임워크는 겉보기에는 서로 다른 방법들이 주로 drift와 diffusion term, terminal distribution, boundary condition에서 차이를 보일 뿐이며, scheduler와 sampler는 그와 독립적인 설계 선택임을 명시해 준다. 이러한 통합을 활용해, 우리는 동일한 architecture와 training protocol을 사용한 채 여러 image enhancement task에 걸쳐 통제된 실증 연구를 수행한다. 그 결과, 일관되게 우세한 방법은 없으며, 대신 성능에 가장 강하게 영향을 미치는 구체적 설계 선택들을 식별하고 분리해 낸다. 마지막으로, 이 통합 프레임워크를 구현하고 stochastic image enhancement method의 빠른 프로토타이핑과 공정한 비교를 가능하게 하는 모듈형 PyTorch 라이브러리 ItoVision을 공개한다.

"},{"id":"66389","en":"Elastic Diffusion Transformer","ko":"탄력적 Diffusion Transformer","authors":"Jiangshan Wang, Zeqiang Lai, Jiarui Chen, Jiayi Guo, Hang Guo, Xiu Li, Xiangyu Yue, Chunchao Guo","pos":"#2901","link":"https://openreview.net/forum?id=493YXJPdeq","abs":"Diffusion Transformers (DiT) have demonstrated remarkable generative capabilities but remain highly computationally expensive. Previous acceleration methods, such as pruning and distillation, typically rely on a fixed computational capacity, leading to insufficient acceleration and degraded generation quality. To address this limitation, we propose \\textbf{Elastic Diffusion Transformer (E-DiT)}, an adaptive acceleration framework for DiT that effectively improves efficiency while maintaining generation quality. Specifically, we observe that the generative process of DiT exhibits substantial sparsity (i.e., some computations can be skipped with minimal impact on quality), and this sparsity varies significantly across samples. Motivated by this observation, E-DiT equips each DiT block with a lightweight router that dynamically identifies sample-dependent sparsity from the input latent. Each router adaptively determines whether the corresponding block can be skipped. If the block is not skipped, the router then predicts the optimal MLP width reduction ratio within the block. During inference, we further introduce a block-level feature caching mechanism that leverages router predictions to eliminate redundant computations in a training-free manner. Extensive experiments across 2D image (Qwen-Image and FLUX) and 3D asset (Hunyuan3D-3.0) demonstrate the effectiveness of E-DiT, achieving up to $\\sim$2$\\times$ speedup with negligible loss in generation quality.","absKo":"Diffusion Transformers (DiT)는 놀라운 generative capability를 보여주었지만, 여전히 계산 비용이 매우 크다. pruning과 distillation 같은 기존 acceleration method는 대개 고정된 computational capacity에 의존하므로, acceleration이 충분하지 않고 generation quality도 저하된다. 이러한 한계를 해결하기 위해, 우리는 generation quality를 유지하면서 효율을 효과적으로 향상시키는 DiT용 adaptive acceleration framework인 \\textbf{Elastic Diffusion Transformer (E-DiT)}를 제안한다. 구체적으로, 우리는 DiT의 generative process가 상당한 sparsity를 보이며(즉, 일부 computation은 quality에 미치는 영향이 거의 없이 생략 가능함), 이 sparsity가 sample마다 크게 달라진다는 점을 관찰한다. 이러한 관찰에 기반하여, E-DiT는 각 DiT block에 입력 latent로부터 sample-dependent sparsity를 동적으로 식별하는 lightweight router를 장착한다. 각 router는 대응되는 block을 skip할 수 있는지 여부를 적응적으로 결정한다. block이 skip되지 않는 경우, router는 그 block 내에서 최적의 MLP width reduction ratio를 추가로 예측한다. inference 동안에는 router prediction을 활용해 redundant computation을 training 없이 제거하는 block-level feature caching mechanism도 도입한다. 2D image(Qwen-Image와 FLUX) 및 3D asset(Hunyuan3D-3.0)에 대한 광범위한 실험은 E-DiT의 효과를 보여주며, generation quality의 손실은 거의 없이 최대 $\\sim$2$\\times$의 speedup을 달성한다."},{"id":"66330","en":"EvReflection: Event-Driven Micro-Dynamics for Reflection Removal","ko":"EvReflection: 반사 제거를 위한 이벤트 구동 미세 동역학","authors":"Jiaxiao Wang, Dachun Kai, Huyue Zhu, Quanquan Hu, Zhenyang Xu, Xiaoyan Sun","pos":"#3007","link":"https://openreview.net/forum?id=4cjHlaUmu0","abs":"Reflection removal is a highly challenging problem. Though remarkable progress has been made, current methods primarily exploit static image priors from a single frame. Due to the inherent ambiguity between layers, existing methods still suffer from severe residual artifacts. In this paper, we propose leveraging event signals to break this ambiguity. By employing event cameras to capture micro-dynamics, we reveal the differential motion between the reflection and background layers. We thereby present a novel event-driven reflection removal network, EvReflection, that utilizes these dynamic cues for layer separation. Specifically, we design a Micro-Dynamics Decoupler to disentangle layer-specific motions from event streams as priors, which then guide a Parallax-Attention Rectifier to cleanly remove artifacts from the RGB image. Furthermore, to address the data shortage, we develop a physics-based simulation pipeline and construct the EVR$^2$ benchmark, the first real-world dataset for this task. Extensive experiments demonstrate that EvReflection significantly outperforms existing methods, recovering clean images in challenging real-world scenarios.","absKo":"reflection removal은 매우 어려운 문제이다. 비록 괄목할 만한 진전이 있었지만, 현재 방법들은 주로 단일 프레임에서 얻는 정적 image prior를 활용한다. layer 간 본질적인 모호성 때문에 기존 방법들은 여전히 심각한 residual artifact 문제를 겪는다. 본 논문에서는 이 모호성을 깨기 위해 event signal을 활용하는 방법을 제안한다. event camera를 사용해 micro-dynamics를 포착함으로써, 우리는 reflection layer와 background layer 사이의 differential motion을 드러낸다. 이를 바탕으로, 이러한 동적 cue를 layer separation에 사용하는 새로운 event-driven reflection removal network인 EvReflection을 제시한다. 구체적으로, 우리는 event stream에서 layer-specific motion을 prior로 분리하는 Micro-Dynamics Decoupler를 설계하고, 이는 이후 Parallax-Attention Rectifier가 RGB image의 artifact를 깨끗하게 제거하도록 안내한다. 또한 데이터 부족 문제를 해결하기 위해 physics-based simulation pipeline을 개발하고, 이 task를 위한 최초의 real-world dataset인 EVR$^2$ benchmark를 구축한다. 광범위한 실험은 EvReflection이 기존 방법보다 현저히 우수하며, 어려운 실제 세계 시나리오에서 깨끗한 이미지를 복원함을 보여준다."},{"id":"66704","en":"Order within Chaos: Capturing Intrinsic Energy Anomalies for AI-Manipulated Image Forgery Localization","ko":"혼돈 속의 질서: AI 조작 이미지 위조 위치 추정을 위한 내재적 에너지 이상 포착","authors":"Yiming Wang, Baiqi Wu, Qingming Li, Jiahao Chen, Leqi Zheng, Shouling Ji","pos":"#3100","link":"https://openreview.net/forum?id=1BLni5Uxsd","abs":"

Recent advancements in generative AI have led to image editing models capable of producing realistic forgeries that evade traditional image forgery localization methods, as these approaches depend on physical noise absent in synthetic data. To address this challenge, we theoretically demonstrate that the diffusion process inherently suppresses local high-frequency variance, creating a statistical energy gap that is distinguishable from the natural entropy of optical imaging. Guided by this insight, we propose FLAME, a unified framework that utilizes a LAD map to capture these intrinsic anomalies, coupled with a parameter-efficient adapter for the SAM 3 to achieve precise, pixel-level forgery localization. Furthermore, to bridge the lag between forensic benchmarks and evolving generative models, we introduce EditStream, an automated pipeline for continuous, instruction-based training data synthesis. Extensive experiments demonstrate that FLAME establishes a new state-of-the-art, significantly outperforming previous methods on AI-generated forgery datasets while effectively generalizing to unseen generative architectures.

","absKo":"

생성형 AI의 최근 발전으로 인해 이미지 편집 model이 기존의 이미지 forgery localization 방법을 회피하는 현실적인 위조물을 생성할 수 있게 되었다. 이러한 접근법은 synthetic data에는 존재하지 않는 물리적 noise에 의존하기 때문이다. 이 문제를 해결하기 위해, 우리는 diffusion process가 본질적으로 local high-frequency variance를 억제하여 optical imaging의 자연 entropy와 구별 가능한 statistical energy gap을 만든다는 점을 이론적으로 보인다. 이러한 통찰에 따라, 우리는 이러한 intrinsic anomaly를 포착하기 위한 LAD map을 활용하고, SAM 3에 대한 parameter-efficient adapter를 결합하여 정밀한 pixel-level forgery localization을 달성하는 통합 프레임워크 FLAME을 제안한다. 또한 forensic benchmark와 진화하는 generative model 사이의 격차를 줄이기 위해, 우리는 instruction-based training data를 지속적으로 합성하는 자동화 파이프라인 EditStream을 도입한다. 광범위한 실험은 FLAME이 새로운 state-of-the-art를 수립하며, AI-generated forgery dataset에서 기존 방법을 크게 능가하는 동시에 보지 못한 generative architecture에도 효과적으로 일반화함을 보여준다.

"},{"id":"65652","en":"VFMF: Dense Forecasting by Generating Foundation Model Features","ko":"VFMF: 기반 모델 특징 생성을 통한 조밀 예측","authors":"Gabrijel Boduljak, Yushi Lan, Christian Rupprecht, Andrea Vedaldi","pos":"#3810","link":"https://openreview.net/forum?id=BWAr39vOQY","abs":"

Forecasting by generating RGB videos is computationally expensive, often physically implausible, and not directly actionable, since it requires translation into decision-making signals. Direct modality forecasting (e.g., predicting future segmentation) produces directly actionable outputs but fails to scale due to the need for labels. Vision Foundation Model (VFM) features offer the best of both worlds: they contain actionable semantic and geometric information that can be easily decoded from the predicted features, while requiring no labels on the downstream task for training. However, almost all existing VFM feature forecasting methods regress future features from fixed number of input frames, with evaluation predominantly on short horizons matching the training setup. We firstly show that existing regression methods struggle with forecasting from partial observations because they average over multiple plausible futures, failing to capture uncertainty in the future given the past. Interestingly, naively replacing deterministic forecasting with generative flow matching does not match the sample quality of the regression model, despite being a mathematically appropriate formulation of the forecasting task. In this work, we explain why this is the case, and we show how to optimally generate foundation model features. Our key insight is that generative modeling of VFM features requires (auto)encoding into a compact latent space suitable for diffusion. We show that this latent space preserves information more effectively than previously used alternatives, such as uncompressed feature diffusion or PCA-based compression, both for forecasting and other applications, such as image generation. Our results suggest that conditional generation of (compressed) VFM features offers a promising and scalable foundation for future scene forecasters.

","absKo":"

RGB video를 생성하여 forecasting하는 것은 계산 비용이 크고, 물리적으로도 종종 그럴듯하지 않으며, decision-making signal로 변환해야 하므로 직접적으로 실행 가능하지도 않다. 직접 modality forecasting(예: 미래 segmentation 예측)은 곧바로 실행 가능한 출력을 제공하지만, label이 필요하기 때문에 확장성이 떨어진다. Vision Foundation Model (VFM) feature는 두 세계의 장점을 모두 제공한다. 즉, 예측된 feature로부터 쉽게 디코딩할 수 있는 실행 가능한 semantic 및 geometric 정보를 담고 있으면서도, 학습 시 downstream task에 대한 label이 필요하지 않다. 그러나 기존 VFM feature forecasting 방법의 거의 전부는 고정된 수의 입력 frame으로부터 미래 feature를 회귀하며, 평가는 주로 training setup과 맞는 short horizon에 집중되어 있다. 우리는 먼저 기존 regression 방법이 부분 관측으로부터의 forecasting에 취약함을 보인다. 이는 이 방법들이 여러 가능한 미래를 평균내어 과거가 주어진 미래의 불확실성을 포착하지 못하기 때문이다. 흥미롭게도, deterministic forecasting을 generative flow matching으로 단순히 대체하는 것은, forecasting task에 수학적으로 적절한 formulation임에도 불구하고 regression model의 sample quality와 맞지 않는다. 이 연구에서 우리는 그 이유를 설명하고, foundation model feature를 최적으로 생성하는 방법을 보인다. 우리의 핵심 통찰은 VFM feature의 generative modeling이 diffusion에 적합한 compact latent space로의 (auto)encoding을 필요로 한다는 점이다. 우리는 이 latent space가 이전에 사용된 uncompressed feature diffusion이나 PCA 기반 compression과 같은 대안들보다 forecasting은 물론 image generation 같은 다른 응용에서도 정보를 더 효과적으로 보존함을 보인다. 우리의 결과는 (compressed) VFM feature의 conditional generation이 미래 scene forecaster를 위한 유망하고 확장 가능한 foundation을 제공함을 시사한다.

"},{"id":"65337","en":"From Coarse to Fine: Deep Prototype Refinement Network for Few-Shot Point Cloud Semantic Segmentation","ko":"거친 것에서 세밀한 것으로: Few-Shot 포인트 클라우드 의미 분할을 위한 심층 프로토타입 정교화 네트워크","authors":"Changshuo Wang, Shuting He, Xiang Fang, Weijun Li, Xingyu Gao, Zhonghang Liu, Prayag Tiwari, Dimitrios Kanoulas","pos":"#3813","link":"https://openreview.net/forum?id=EdzZDEjpQe","abs":"

Few-shot point cloud semantic segmentation (FS-PCSS) aims to achieve precise segmentation of novel categories using only limited labeled samples. Existing prototype-based methods typically rely on shallow feature fusion strategies, failing to adequately model the feature distribution shift between support and query sets, resulting in insufficient prototype adaptation. To address this, we propose the Deep Prototype Refinement Network (DPR-Net), which systematically achieves progressive adaptation by constructing a coarse-to-fine prototype evolution trajectory. Our core Dynamic Prototype Refinement (DPR) module explicitly decomposes features into common and distinctive subspaces based on channel activation, enabling targeted adjustment of domain-sensitive features while preserving class-shared semantics. By cascading multiple refinement modules, we construct a prototype trajectory transitioning from support-biased to query-adapted representations, mitigating both under- and over-adaptation. Furthermore, our Mixture of Prototype Experts (MoPE) mechanism treats prototypes at different stages as complementary experts and adaptively ensembles their predictions through confidence-driven weighting. Extensive experiments demonstrate that DPR-Net achieves state-of-the-art performance with high efficiency. Notably, with only 0.28M parameters, DPR-Net achieves 80.76% mIoU on S3DIS (2-way 1-shot), representing a 15.92% improvement over the baseline.

","absKo":"

Few-shot point cloud semantic segmentation(FS-PCSS)은 제한된 labeled sample만으로 novel category를 정밀하게 segmentation하는 것을 목표로 한다. 기존 prototype-based 방법은 대개 얕은 feature fusion 전략에 의존하여 support set과 query set 사이의 feature distribution shift를 충분히 모델링하지 못하고, 그 결과 prototype adaptation이 불충분하다. 이를 해결하기 위해 우리는 coarse-to-fine prototype evolution trajectory를 구축함으로써 점진적 adaptation을 체계적으로 달성하는 Deep Prototype Refinement Network(DPR-Net)를 제안한다. 우리의 핵심 Dynamic Prototype Refinement(DPR) 모듈은 channel activation을 기반으로 feature를 common subspace와 distinctive subspace로 명시적으로 분해하여, class-shared semantic을 보존하면서 domain-sensitive feature를 목표 지향적으로 조정할 수 있게 한다. 여러 refinement module을 연쇄적으로 적용함으로써, 우리는 support-biased representation에서 query-adapted representation으로 전이하는 prototype trajectory를 구성하여 과소 적응과 과대 적응을 모두 완화한다. 더 나아가, Mixture of Prototype Experts(MoPE) 메커니즘은 서로 다른 단계의 prototype을 상호 보완적인 expert로 간주하고 confidence-driven weighting을 통해 예측을 적응적으로 ensemble한다. 광범위한 실험은 DPR-Net이 높은 효율성과 함께 state-of-the-art 성능을 달성함을 보여준다. 특히 0.28M개의 파라미터만으로 DPR-Net은 S3DIS(2-way 1-shot)에서 80.76% mIoU를 달성하여 baseline 대비 15.92% 향상을 보였다.

"},{"id":"63865","en":"Doppler Prompting for Stable mmWave-based Human Pose Estimation","ko":"안정적인 mmWave 기반 사람 자세 추정을 위한 Doppler prompting","authors":"Shuntian Zheng, Jiaqi Li, Xiaoman Lu, Shuai He, Yu Guan","pos":"#4600","link":"https://openreview.net/forum?id=T4JzJ0V69W","abs":"

Millimeter-wave (mmWave) enables privacy-preserving, illumination-robust human pose estimation (HPE), with each mmWave frame represented as a range--angle--Doppler tensor, providing spatial magnitude for localization and Doppler signatures for motion-related cues. However, existing mmWave-based HPE methods either underutilize or naïvely fuse Doppler signatures with spatial magnitude, disregarding their distinct physical semantics. As a result, non-human Doppler signatures can be misinterpreted as human motion cues, leading to jittery trajectories. We propose \\textbf{PULSE}, which converts Doppler signatures into confidence-aware motion prompts and injects them into spatial magnitude reasoning through constrained interactions. By screening Doppler prompts before they influence prediction, PULSE first suppresses spurious spectral motion cues and then uses the screened prompts to stabilize prediction. Across three datasets spanning single- and multi-person settings, PULSE consistently improves pose accuracy and temporal stability, indicating that controlled Doppler prompting is a practical direction for stable mmWave HPE. Codes are available in supplementary materials.

","absKo":"

Millimeter-wave(mmWave)는 privacy-preserving하고 illumination-robust한 human pose estimation(HPE)을 가능하게 하며, 각 mmWave frame은 range--angle--Doppler tensor로 표현되어 위치 추정을 위한 공간적 크기 정보와 motion 관련 단서를 위한 Doppler signature를 제공한다. 그러나 기존의 mmWave 기반 HPE 방법은 Doppler signature를 공간적 크기 정보와 충분히 활용하지 못하거나 단순히 naïve하게 결합하며, 이들의 서로 다른 물리적 의미를 무시한다. 그 결과, 인간이 아닌 Doppler signature가 사람의 motion cue로 오인되어 trajectory가 흔들릴 수 있다. 우리는 Doppler signature를 confidence-aware motion prompt로 변환하고, 이를 constrained interaction을 통해 spatial magnitude reasoning에 주입하는 \\textbf{PULSE}를 제안한다. PULSE는 Doppler prompt가 prediction에 영향을 미치기 전에 이를 선별하여, 먼저 가짜 spectral motion cue를 억제한 뒤, 선별된 prompt를 사용해 prediction을 안정화한다. 단일 인물과 다중 인물 setting을 포괄하는 세 개의 dataset에서, PULSE는 일관되게 pose accuracy와 temporal stability를 향상시켜, controlled Doppler prompting이 안정적인 mmWave HPE를 위한 실용적인 방향임을 보여 준다. 코드는 supplementary materials에서 확인할 수 있다.

"},{"id":"65807","en":"SleepLM: Natural-Language Intelligence for Human Sleep","ko":"SleepLM: 인간 수면을 위한 자연어 지능","authors":"Zongzhe Xu, Zitao Shuai, Eideen Mozaffari, Ravi Aysola, Rajesh Kumar, Yuzhe Yang","pos":"#4601","link":"https://openreview.net/forum?id=9wpwfSJCp9","abs":"

We present SleepLM, a family of sleep-language foundation models that enable human sleep alignment, interpretation, and interaction with natural language. Despite the critical role of sleep, learning-based sleep analysis systems operate in closed label spaces (e.g., predefined stages or events) and fail to describe, query, or generalize to novel sleep phenomena. SleepLM bridges natural language and multimodal polysomnography, enabling language-grounded representations of sleep physiology. To support this alignment, we introduce a multilevel sleep caption generation pipeline that enables the curation of the first large-scale sleep-text dataset, comprising over 100K hours of data from more than 10,000 individuals. Furthermore, we present a unified pretraining objective that combines contrastive alignment, caption generation, and signal reconstruction to better capture physiological fidelity and cross-modal interactions. Extensive experiments on real-world sleep understanding tasks verify that SleepLM outperforms state-of-the-art in zero-shot and few-shot learning, cross-modal retrieval, and sleep captioning. Importantly, SleepLM also exhibits intriguing capabilities including language-guided event localization, targeted insight generation, and zero-shot generalization to unseen tasks. All code and data will be open-sourced.

","absKo":"

우리는 인간의 수면 정렬, 해석, 그리고 natural language와의 상호작용을 가능하게 하는 sleep-language foundation model 계열인 SleepLM을 제시한다. 수면은 매우 중요하지만, learning-based sleep analysis system은 닫힌 label space(예: 미리 정의된 stage나 event)에서 동작하며 새로운 수면 현상을 설명하거나 질의하거나 generalize하지 못한다. SleepLM은 natural language와 multimodal polysomnography를 연결하여, 수면 생리의 language-grounded representation을 가능하게 한다. 이러한 정렬을 지원하기 위해, 우리는 10,000명 이상의 개인으로부터 수집한 10만 시간 이상의 데이터를 포함하는 최초의 대규모 sleep-text dataset을 구축할 수 있게 하는 multilevel sleep caption generation pipeline을 도입한다. 나아가, 생리적 fidelity와 cross-modal interaction을 더 잘 포착하기 위해 contrastive alignment, caption generation, signal reconstruction을 결합한 통합 pretraining objective를 제안한다. 실제 수면 이해 과제에 대한 광범위한 실험은 SleepLM이 zero-shot 및 few-shot learning, cross-modal retrieval, 그리고 sleep captioning에서 state-of-the-art를 능가함을 검증한다. 특히 SleepLM은 language-guided event localization, targeted insight generation, 보지 못한 task에 대한 zero-shot generalization을 포함한 흥미로운 능력도 보인다. 모든 코드와 데이터는 오픈소스로 공개될 것이다.

"},{"id":"64583","en":"R$^3$L: Reasoning 3D Layouts from Relative Spatial Relations","ko":"R$^3$L: 상대적 공간 관계로부터의 3D 레이아웃 추론","authors":"Zhifeng Gu, Yuqi Wang, Bing WANG","pos":"#1206","link":"https://openreview.net/forum?id=LxqIJW626c","abs":"Relative spatial relations provide a compact description of spatial structure, serving as a key component of relative spatial reasoning in 3D layout generation. Recent works leverage Multimodal Large Language Models (MLLMs) to infer these relations, but the inferred relations are often unreliable and are resolved by post-hoc heuristics at the cost of semantic fidelity. In this paper, we propose R$^3$L, a general framework that improves the reliability and consistency of relative spatial reasoning for 3D layout generation. Our key motivation is that multi-hop reasoning requires repeated reference-frame shifts, which accumulate errors and lead to semantic and metric drift. To mitigate this, we propose invariant spatial decomposition to shorten relations chains, and consistent spatial imagination that uses an imagine-and-revise loop to encourage self-consistency during relation inference. We further design supportive spatial optimization that eases pose optimization by global-to-local coordinate re-parameterization. Extensive experiments across diverse scene types and instructions demonstrate that R$^3$L improves layout feasibility and semantic consistency. Notably, our analysis shows that resolving frame-induced inconsistencies during reasoning is crucial for reliable multi-hop relative spatial reasoning. Code will be released upon acceptance.","absKo":"상대적 spatial relation은 공간 구조를 간결하게 기술하며, 3D layout generation에서 relative spatial reasoning의 핵심 구성 요소로 작용한다. 최근 연구들은 이러한 relation을 추론하기 위해 Multimodal Large Language Models (MLLMs)를 활용하지만, 추론된 relation은 종종 신뢰성이 낮고, semantic fidelity를 희생하는 후처리 heuristic으로 해결된다. 본 논문에서는 3D layout generation에서 relative spatial reasoning의 신뢰성과 일관성을 향상시키는 일반적 framework인 R$^3$L을 제안한다. 우리의 핵심 동기는 multi-hop reasoning이 반복적인 reference-frame shift를 필요로 하며, 이로 인해 error가 누적되고 semantic 및 metric drift가 발생한다는 점이다. 이를 완화하기 위해, 우리는 relation chain을 짧게 만드는 invariant spatial decomposition과, relation 추론 동안 self-consistency를 촉진하기 위해 imagine-and-revise loop를 사용하는 consistent spatial imagination을 제안한다. 또한 pose optimization을 global-to-local coordinate re-parameterization으로 완화하는 supportive spatial optimization도 설계한다. 다양한 scene type과 instruction에 대한 광범위한 실험은 R$^3$L이 layout feasibility와 semantic consistency를 향상시킴을 보여준다. 특히 우리의 분석은 reasoning 과정에서 frame-induced inconsistency를 해결하는 것이 신뢰할 수 있는 multi-hop relative spatial reasoning에 중요함을 보여준다. 코드는 acceptance 후 공개될 예정이다."},{"id":"65662","en":"DiLA: Disentangled Latent Action World Models","ko":"DiLA: 분리된 잠재 행동 World Model","authors":"Tianqiu Zhang, Muyang Lyu, Yufan Zhang, Fang Fang, Si Wu","pos":"#1601","link":"https://openreview.net/forum?id=BRBHruBDkb","abs":"

Latent Action Models (LAMs) enable the learning of world models from unlabeled video by inferring abstract actions between consecutive frames. However, LAMs face a fundamental trade-off between action abstraction and generation fidelity. Existing methods typically circumvent this issue by using two-stage training with pre-trained world models or by limiting predictions to optical flow. In this paper, we introduce DiLA, a novel Disentangled Latent Action world model that aims to resolve this trade-off via content-structure disentanglement. Our key insight is that disentanglement and latent action learning are co-evolving: the predictive bottleneck inherent in latent action learning serves as a driving force for disentanglement, compelling the model to distill spatial layouts into the structure pathway while offloading visual details to a separate content pathway for generation. This synergy yields a continuous, semantically structured latent action space without compromising generative quality. DiLA achieves superior results in video generation quality, action transfer, visual planning, and manifold interpretability. These findings establish DiLA as a unified framework that simultaneously achieves high-level action abstraction and high-fidelity generation, advancing the frontier of self-supervised world model learning.

","absKo":"

Latent Action Model (LAM)은 연속된 frame 사이의 abstract action을 추론함으로써 unlabeled video로부터 world model을 학습할 수 있게 해준다. 그러나 LAM은 action abstraction과 generation fidelity 사이에서 근본적인 trade-off에 직면한다. 기존 방법들은 보통 pre-trained world model을 사용하는 two-stage training으로 이 문제를 우회하거나, 예측을 optical flow로 제한한다. 본 논문에서는 content-structure disentanglement를 통해 이 trade-off를 해소하고자 하는 새로운 DiLA, 즉 Disentangled Latent Action world model을 제안한다. 우리의 핵심 통찰은 disentanglement와 latent action learning이 co-evolving한다는 점이다. latent action learning에 내재한 predictive bottleneck은 disentanglement를 이끄는 원동력으로 작용하여, 모델이 spatial layout을 structure pathway로 증류하도록 강제하는 한편 visual detail은 별도의 content pathway로 넘겨 생성하게 만든다. 이러한 시너지는 생성 품질을 손상시키지 않으면서 연속적이고 semantic하게 구조화된 latent action space를 형성한다. DiLA는 video generation quality, action transfer, visual planning, manifold interpretability에서 우수한 결과를 달성한다. 이러한 결과는 DiLA를 고수준 action abstraction과 고충실도 generation을 동시에 달성하는 통합 프레임워크로 확립하며, self-supervised world model learning의 최전선을 진전시킨다.

"},{"id":"62192","en":"RoboTwin 2.0: A Scalable Data Generator and Benchmark with Strong Domain Randomization for Robust Bimanual Robotic Manipulation","ko":"RoboTwin 2.0: 강건한 양손 로봇 조작을 위한 강한 Domain Randomization 기반 확장형 데이터 생성기 및 Benchmark","authors":"Tianxing Chen, Zanxin Chen, Baijun Chen, Zijian Cai, Yibin Liu, Zixuan Li, Qiwei Liang, Xianliang Lin, Yiheng Ge, Zhenyu Gu, Weiliang Deng, Yubin Guo, Tian Nian, Xuanbing Xie, Qiangyu Chen, KailunSu, Tianling Xu, Guodong Liu, Mengkang Hu, Huan-ang Gao, Kaixuan Wang, Zhixuan Liang, Yusen Qin, Xiaokang Yang, Ping Luo, Yao Mu","pos":"#310","link":"https://openreview.net/forum?id=itonej9GIV","abs":"

Simulation-based data synthesis has emerged as a powerful paradigm for enhancing real-world robotic manipulation. However, existing synthetic datasets remain insufficient for robust bimanual manipulation due to two key challenges: (1) the lack of an autonomous self-correcting mechanism to resolve execution failures in complex coordination tasks, and (2) the scarcity of diverse visual and spatial variations required to bridge the sim-to-real gap. To this end, we present RoboTwin 2.0, a scalable simulation framework that enables closed-loop, automated, large-scale generation of diverse and realistic data, along with unified evaluation protocols for dual-arm manipulation. Built upon RoboTwin-OD, a foundational library of 731 instances across 147 categories with rich semantic annotations, our framework integrates Multimodal Large Language Models (MLLMs) with simulation-in-the-loop verification. This integration forms an automated feedback mechanism that significantly boosts the success rate of expert task program generation. To enhance robust sim-to-real transfer, RoboTwin 2.0 incorporates structured domain randomization along five axes: clutter, lighting, background, tabletop height and language instructions, thereby maximizing data diversity. We instantiate this framework across 50 dual-arm tasks spanning five robot embodiments. Empirical evaluations demonstrate that Vision-Language-Action (VLA) models pre-trained on our synthetic data achieve a 3.6x improvement in few-shot real-world transfer (over a 10-demo baseline) and a 2.2x gain in zero-shot generalization. We release the data generator, benchmark, pre-collected dataset, and code to support scalable research in robust bimanual manipulation.

","absKo":"

Simulation-based data synthesis는 real-world robotic manipulation을 향상시키는 강력한 패러다임으로 부상했다. 그러나 기존 synthetic dataset은 다음 두 가지 핵심 과제로 인해 robust bimanual manipulation에 여전히 충분하지 않다. (1) 복잡한 coordination task에서 실행 실패를 해결할 autonomous self-correcting mechanism의 부재, (2) sim-to-real gap을 메우는 데 필요한 다양한 visual 및 spatial variation의 부족. 이를 위해 우리는 closed-loop의 자동 대규모 생성이 가능한 다양한 현실적 데이터를 생성하고, dual-arm manipulation을 위한 통합 evaluation protocol을 제공하는 확장 가능한 simulation framework인 RoboTwin 2.0을 제시한다. 풍부한 semantic annotation을 갖춘 147개 category의 731개 instance로 이루어진 기반 라이브러리 RoboTwin-OD 위에 구축된 우리의 framework는 Multimodal Large Language Models (MLLMs)와 simulation-in-the-loop verification을 통합한다. 이 통합은 expert task program 생성의 성공률을 크게 높이는 자동 feedback mechanism을 형성한다. 강인한 sim-to-real transfer를 강화하기 위해 RoboTwin 2.0은 clutter, lighting, background, tabletop height, language instruction의 다섯 축에 걸친 structured domain randomization을 도입하여 data diversity를 극대화한다. 우리는 이 framework를 5개의 robot embodiment에 걸친 50개의 dual-arm task에 적용했다. 실증적 평가는 우리의 synthetic data로 pre-training된 Vision-Language-Action (VLA) model이 few-shot real-world transfer(10-demo baseline 대비)에서 3.6배 향상되고, zero-shot generalization에서 2.2배의 성능 향상을 보임을 입증한다. 우리는 robust bimanual manipulation 연구의 확장성을 지원하기 위해 data generator, benchmark, 사전 수집된 dataset, code를 공개한다.

"},{"id":"64080","en":"Dismantling the Illusion of Vision-Language-Action Models Competence via Explicit Distributional Shifts","ko":"명시적 분포 변화를 통한 Vision-Language-Action 모델 역량의 착시 해체","authors":"Xueyang Zhou, Yangming Xu, Guiyao Tie, Yongchao Chen, Chaoran Hu, Bo Tao, xingwei zhao, Xiang Xiang, Pan Zhou, Lichao Sun","pos":"#3703","link":"https://openreview.net/forum?id=R7vl47raSn","abs":"

Given that simulation can never exhaustively enumerate reality, generalization is the determining factor for whether Vision-Language-Action (VLA) models can translate benchmark success into real-world functionality. However, current evaluation protocols often incentivize mechanical memorization rather than robust policy learning, leading to a paradoxical duality of failure: high-scoring models exhibit spurious invariance to semantic changes while simultaneously displaying extreme brittleness to trivial environmental perturbations. To address this, we introduce LIBERO-Gen, a diagnostic benchmark systematically designed to shift evaluation from intuition-driven heuristics to explicit distributional assumptions. Through a hierarchical protocol spanning In-distribution, Compositional, and Domain Generalization, LIBERO-Gen reveals performance stratifications previously masked by standard metrics. Our analysis identifies Pi0.5 as the top performer (64.0% in Spatial-CG; 21.2% in Task-CG). By identifying perceptual instability and action binding collapse as primary failure modes while validating the efficacy of structured ``Stair” sampling, LIBERO-Gen establishes a rigorous baseline for deployment reliability.

","absKo":"

simulation이 현실을 완전하게 열거할 수 없다는 점을 고려하면, generalization은 Vision-Language-Action (VLA) model이 benchmark 성공을 실제 세계 기능으로 전환할 수 있는지를 결정하는 요인이다. 그러나 현재의 평가 프로토콜은 견고한 policy learning보다 기계적 암기만을 장려하는 경우가 많으며, 그 결과 고득점 model은 semantic 변화에 대해 spurious invariance를 보이는 동시에, 사소한 환경 perturbation에는 extreme brittleness를 드러내는 역설적 이중 실패를 보인다. 이를 해결하기 위해 우리는 LIBERO-Gen을 제안한다. 이는 직관에 의존한 heuristic에서 명시적 distributional assumption으로 평가를 전환하도록 체계적으로 설계된 diagnostic benchmark이다. In-distribution, Compositional, Domain Generalization을 아우르는 계층적 프로토콜을 통해, LIBERO-Gen은 표준 metric이 가려온 성능의 층위를 드러낸다. 우리의 분석은 Pi0.5를 최고 성능 모델로 식별했으며 (Spatial-CG에서 64.0%, Task-CG에서 21.2%), perceptual instability와 action binding collapse를 주요 실패 모드로 규명하는 한편 구조화된 ``Stair” sampling의 효과를 검증함으로써, LIBERO-Gen이 배포 신뢰성을 위한 엄밀한 baseline을 확립함을 보인다.

"},{"id":"65193","en":"DreamDojo: A Real-Time Robot World Model from Large-Scale Human Videos","ko":"DreamDojo: 대규모 인간 영상으로부터의 실시간 로봇 World Model","authors":"Shenyuan Gao, William Liang, Kaiyuan Zheng, Ayaan Malik, Seonghyeon Ye, Sihyun Yu, Wei-Cheng Tseng, Yuzhu Dong, Kaichun Mo, Chen-Hsuan Lin, Jiannan Xiang, Yuqi Xie, Ruijie Zheng, Dantong Niu, Pooya Jannaty, Jinwei Gu, Jun Zhang, Jitendra Malik, Pieter Abbeel, Ming-Yu Liu, Yuke Zhu, Joel Jang, Jim Fan","pos":"#3815","link":"https://openreview.net/forum?id=FuvU7PTyED","abs":"

Being able to simulate the outcomes of actions in varied environments will revolutionize the development of generalist agents at scale. However, modeling these world dynamics, especially for dexterous robotics tasks, poses significant challenges due to limited data coverage and scarce action labels. As an endeavor towards this end, we introduce DreamDojo, a foundation world model that learns diverse interactions and dexterous controls from 44k hours of egocentric human videos. Our data mixture represents the largest video dataset to date for world model pretraining, spanning a wide range of daily scenarios with diverse objects and skills. To address the scarcity of action labels, we introduce continuous latent actions as unified proxy actions, enhancing interaction knowledge transfer from unlabeled videos. After post-training on small-scale target robot data, DreamDojo demonstrates a strong understanding of physics and precise action controllability. We also devise a distillation pipeline that accelerates DreamDojo to a real-time speed of 10.93 FPS and further improves consistency to the context. Our work enables several important applications based on generative world models, including live teleoperation, policy evaluation, and model-based planning. Systematic evaluation on multiple challenging out-of-distribution (OOD) benchmarks verifies the significance of our method for simulating open-world, contact-rich tasks, paving the way for general-purpose robot world models.

","absKo":"

다양한 환경에서 행동의 결과를 시뮬레이션할 수 있는 능력은 대규모 generalist agent 개발을 혁신할 것이다. 그러나 이러한 world dynamics, 특히 dexterous robotics task를 모델링하는 일은 제한적인 data coverage와 부족한 action label 때문에 큰 도전을 수반한다. 이러한 목표를 향한 노력으로, 우리는 44k시간의 egocentric human video로부터 다양한 interaction과 dexterous control을 학습하는 foundation world model인 DreamDojo를 소개한다. 우리의 data mixture는 world model pretraining을 위한 현재까지 가장 큰 video dataset으로, 다양한 object와 skill을 포괄하는 폭넓은 일상 시나리오를 담고 있다. action label의 희소성을 해결하기 위해, 우리는 unlabeled video로부터의 interaction 지식 전달을 향상시키는 통합 proxy action으로서 continuous latent action을 도입한다. 소규모 target robot data에 대한 post-training 후, DreamDojo는 물리 현상에 대한 강한 이해와 정밀한 action controllability를 보여준다. 또한 우리는 DreamDojo를 real-time speed인 10.93 FPS로 가속하고 context에 대한 consistency를 더 향상시키는 distillation pipeline도 설계한다. 우리의 작업은 live teleoperation, policy evaluation, model-based planning을 포함한 generative world model 기반의 여러 중요한 응용을 가능하게 한다. 여러 까다로운 out-of-distribution (OOD) benchmark에 대한 체계적 평가를 통해 open-world의 contact-rich task를 시뮬레이션하는 데 있어 우리의 방법의 중요성이 검증되며, 범용 robot world model로 가는 길을 연다.

"},{"id":"63867","en":"PhyScene3D: Physically Consistent 3D Interactive Tabletop Scene Generation","ko":"PhyScene3D: 물리적으로 일관된 3D 인터랙티브 탁상 장면 생성","authors":"Weixing Chen, Zhuoqian Feng, Yexin Zhang, Yifan Wen, Yang Liu, Yinghong Liao, Weichao Qiu, Guanbin Li, Liang Lin","pos":"#406","link":"https://openreview.net/forum?id=T4A2aYyaq9","abs":"

Generating physically consistent 3D tabletop scenes is a fundamental yet underexplored problem for interactive and generalist robotic learning. The challenge stems from dense object hierarchies and irregular affordances. Existing methods, ranging from decoupled symbolic solvers to end-to-end regression models, often suffer from error propagation or overfitting to noisy supervision containing widespread physical violations. To address these limitations, we introduce PhyScene3D, a framework that reformulates generation as a Human-Mimetic Constructive Process. The proposed Cognitive Topological Reasoning Chain (CTRC) factorizes scene synthesis into a sequential, anchor-conditioned process under the novel 3D Axis-Aligned Bounding Box (3D AABB)-based placement scheme, thereby imposing a strong structural inductive bias. To address imperfect supervision and physical infeasibility, we introduce Physics-Aware Denoising Alignment (PADA), which integrates a differentiable Signed Distance Field (SDF) with Test-Time Optimization (TTO) to project generated scenes onto a physics-feasible manifold while preserving semantic intent. Experiments demonstrate that PhyScene3D outperforms state-of-the-art approaches in both semantic accuracy and physical validity, achieving a 40\\% reduction in collision rate relative to the human-annotated training data.

","absKo":"

물리적으로 일관된 3D tabletop scene을 생성하는 일은 interactive 및 generalist robotic learning을 위한 근본적이면서도 아직 충분히 탐구되지 않은 문제이다. 이 문제는 촘촘한 object hierarchy와 불규칙한 affordance에서 비롯된다. 기존 방법은 분리된 symbolic solver부터 end-to-end regression model까지 다양하지만, 오류 전파에 취약하거나 광범위한 물리 위반을 포함한 noisy supervision에 과적합되는 경우가 많다. 이러한 한계를 해결하기 위해, 우리는 생성 문제를 Human-Mimetic Constructive Process로 재정식화한 프레임워크인 PhyScene3D를 제안한다. 제안하는 Cognitive Topological Reasoning Chain(CTRC)은 새로운 3D Axis-Aligned Bounding Box(3D AABB) 기반 배치 방식을 통해 scene synthesis를 sequential하고 anchor-conditioned된 과정으로 분해하여, 강한 구조적 inductive bias를 부여한다. 불완전한 supervision과 물리적으로 불가능한 사례를 다루기 위해, 우리는 미분 가능한 Signed Distance Field(SDF)와 Test-Time Optimization(TTO)을 통합한 Physics-Aware Denoising Alignment(PADA)를 도입하여, semantic intent를 보존하면서 생성된 scene을 physics-feasible manifold로 사상한다. 실험 결과, PhyScene3D는 semantic accuracy와 physical validity 모두에서 state-of-the-art 접근법을 능가했으며, 인간 주석 training data 대비 collision rate를 40\\% 줄였다.

"},{"id":"61757","en":"World Guidance: World Modeling in Condition Space for Action Generation","ko":"World Guidance: Action 생성을 위한 Condition Space의 World Modeling","authors":"Yue Su, Sijin Chen, Haixin Shi, Mingyu Liu, Zhengshen Zhang, Ningyuan Huang, Weiheng Zhong, Zhengbang Zhu, Yuxiao Liu, Xihui Liu","pos":"#601","link":"https://openreview.net/forum?id=nCWnYFezUm","abs":"

Leveraging future observation modeling to facilitate action generation presents a promising avenue for enhancing the capabilities of Vision-Language-Action (VLA) models. However, existing approaches struggle to strike a balance between maintaining efficient, predictable future representations and preserving sufficient fine-grained information to guide precise action generation. To address this limitation, we propose WoG (World Guidance), a framework that maps future observations into compact conditions by injecting them into the action inference pipeline. The VLA is then trained to simultaneously predict these compressed conditions alongside future actions, thereby achieving effective world modeling within the condition space for action inference. We demonstrate that modeling and predicting this condition space not only facilitates fine-grained action generation but also exhibits superior generalization capabilities. Moreover, it learns effectively from substantial human manipulation videos. Extensive experiments across both simulation and real-world environments validate that WoG significantly outperforms existing methods based on future prediction.

","absKo":"미래 관측 모델링을 활용해 action generation을 촉진하는 것은 Vision-Language-Action (VLA) model의 성능을 향상시키는 유망한 방향이다. 그러나 기존 접근법은 효율적이고 예측 가능한 미래 표현을 유지하는 것과 정밀한 action generation을 유도할 수 있을 만큼 충분한 세밀 정보를 보존하는 것 사이의 균형을 맞추는 데 어려움을 겪는다. 이러한 한계를 해결하기 위해, 우리는 미래 관측을 action inference pipeline에 주입하여 이를 compact condition으로 매핑하는 WoG (World Guidance) framework를 제안한다. 이후 VLA는 이러한 compressed condition과 미래 action을 동시에 예측하도록 학습되며, 이를 통해 condition space 내에서 효과적인 world modeling을 action inference를 위해 달성한다. 우리는 이 condition space를 모델링하고 예측하는 것이 세밀한 action generation을 촉진할 뿐 아니라 우수한 generalization capability도 보인다는 것을 입증한다. 더 나아가, 이는 대규모 human manipulation video로부터 효과적으로 학습한다. simulation과 real-world environment 모두에서 수행한 광범위한 실험은 WoG가 미래 예측 기반의 기존 방법들을 크게 능가함을 검증한다.

"},{"id":"64826","en":"XR-1: Towards Versatile Vision-Language-Action Models via Learning Unified Vision-Motion Representations","ko":"XR-1: 통합 Vision-Motion Representation 학습을 통한 다재다능한 Vision-Language-Action 모델을 향하여","authors":"Shichao Fan, Kun Wu, Yinuo Zhao, Xinhua Wang, Di Wu, Fei Liao, Ning Liu, Yixue Zhang, Zhen Zhao, Zhiyuan Xu, Meng Li, Qingjie Liu, Shanghang Zhang, Min Wan, Jian Tang","pos":"#602","link":"https://openreview.net/forum?id=JO0IsGJg16","abs":"Recent progress in large-scale robotic datasets and vision-language models (VLMs) has advanced research on vision-language-action (VLA) models. However, existing VLA models still face two fundamental challenges: (\\textit{i}) producing precise low-level actions from high-dimensional observations, (\\textit{ii}) bridging domain gaps across heterogeneous data sources, including diverse robot embodiments and human demonstrations. Existing methods often encode latent variables from either visual dynamics or robotic actions to guide policy learning, but they fail to fully exploit the complementary multi-modal knowledge present in large-scale, heterogeneous datasets. In this work, we present \\textbf{XR-1}, a novel framework for versatile and scalable VLA learning across diverse robots, tasks, and environments. At its core, XR-1 introduces the \\emph{Unified Vision-Motion Codes (UVMC)}, a discrete latent representation learned via a dual-branch VQ-VAE that jointly encodes visual dynamics and robotic motion. UVMC addresses these challenges by (\\textit{i}) serving as an intermediate representation between the observations and actions, and (\\textit{ii}) aligning multimodal dynamic information from heterogeneous data sources to capture complementary knowledge. To effectively exploit UVMC, we propose a \\emph{three-stage training paradigm}: (\\textit{i}) self-supervised UVMC learning, (\\textit{ii}) UVMC-guided pretraining on large-scale cross-embodiment robotic datasets, and (\\textit{iii}) task-specific post-training. We validate XR-1 through extensive real-world experiments with more than 12,000 rollouts on six different robot embodiments, spanning over 120 diverse manipulation tasks. XR-1 consistently outperforms state-of-the-art baselines such as $\\pi_0$ and GR00T-N1.5 while demonstrating strong generalization to novel objects, background variations, distractors, and illumination changes. Our project is at \\href{https://xr-1-vla.github.io/}{https://xr-1-vla.github.io/}, and our code will be open-sourced.","absKo":"대규모 robot dataset과 vision-language model (VLM)의 최근 발전은 vision-language-action (VLA) model 연구를 진전시켰다. 그러나 기존 VLA model은 여전히 두 가지 근본적인 과제에 직면한다. (\\textit{i}) 고차원 관측으로부터 정밀한 low-level action을 생성하는 것, (\\textit{ii}) 다양한 robot embodiment와 human demonstration을 포함한 이질적인 데이터 소스 간의 domain gap을 메우는 것이다. 기존 방법은 정책 학습을 유도하기 위해 visual dynamics 또는 robotic action 중 한쪽에서 latent variable을 인코딩하는 경우가 많지만, 대규모 이질적 dataset에 존재하는 상호보완적인 multimodal knowledge를 충분히 활용하지 못한다. 본 연구에서는 다양한 robot, task, environment 전반에서 범용적이고 확장 가능한 VLA 학습을 위한 새로운 프레임워크 \\textbf{XR-1}을 제안한다. 그 핵심에는 visual dynamics와 robotic motion을 공동으로 인코딩하는 dual-branch VQ-VAE를 통해 학습되는 discrete latent representation인 \\emph{Unified Vision-Motion Codes (UVMC)}가 있다. UVMC는 (\\textit{i}) 관측과 action 사이의 intermediate representation으로 작동하고, (\\textit{ii}) 이질적 데이터 소스의 multimodal dynamic information을 정렬하여 상호보완적 지식을 포착함으로써 이러한 과제를 해결한다. UVMC를 효과적으로 활용하기 위해, 우리는 \\emph{three-stage training paradigm}을 제안한다. (\\textit{i}) self-supervised UVMC learning, (\\textit{ii}) 대규모 cross-embodiment robotic dataset에서의 UVMC-guided pretraining, (\\textit{iii}) task-specific post-training이다. 우리는 여섯 가지 서로 다른 robot embodiment에 걸쳐 120개 이상의 다양한 manipulation task를 포함하는 12,000회 이상의 rollout을 통해 실제 환경에서 XR-1을 광범위하게 검증했다. XR-1은 $\\pi_0$와 GR00T-N1.5 같은 state-of-the-art baseline을 일관되게 능가하면서, 새로운 object, background variation, distractor, illumination 변화에 대해서도 강한 generalization을 보였다. 프로젝트는 \\href{https://xr-1-vla.github.io/}{https://xr-1-vla.github.io/}에 있으며, 코드는 오픈소스로 공개될 예정이다."},{"id":"63796","en":"Zero-Shot 3D Question Answering via Hierarchical View-to-Token Transportation","ko":"계층적 view-to-token transportation을 통한 zero-shot 3D 질문응답","authors":"Dongsheng Wang, Dawei Su, Hui Huang","pos":"#603","link":"https://openreview.net/forum?id=Tg9q3RHCdM","abs":"

Recently, zero-shot 3D scene understanding via 2D Vision-Language Models (VLMs) has gained increasing research interest due to their promising spatial reasoning capabilities. Typically, multiple 2D views are sampled from a 3D point cloud and fed into pre-trained VLMs to answer a given question. This paradigm highlights the critical role of input context quality and raises the challenge of retaining as many task-relevant 3D details as possible under a limited input budget. We propose \\texttt{KeyVT}, a hierarchical approach for input context collection at both the view and token levels. Specifically, we combine pixel features with camera parameters and assess view importance based on both semantic content and geometric position, resulting in spatially consistent and task-relevant views. Furthermore, we address redundancy among patches across selected views by identifying representative tokens under the optimal transport (OT) framework, where view tokens and key tokens are formulated as two discrete distributions in the embedding space. These key tokens are expected to cover all view features by minimizing the OT distance. We evaluate our framework on three widely used benchmarks, demonstrating significant improvements over existing tuning-free methods and performance comparable to training-based approaches.

","absKo":"

최근 2D Vision-Language Models (VLMs)를 통한 zero-shot 3D scene understanding은 유망한 spatial reasoning 능력 덕분에 연구 관심이 높아지고 있다. 일반적으로 하나의 3D point cloud에서 여러 개의 2D view를 샘플링한 뒤, 이를 pre-trained VLM에 입력하여 주어진 질문에 답한다. 이러한 paradigm은 입력 context 품질의 중요성을 부각시키며, 제한된 input budget 하에서 task와 관련된 3D 세부 정보를 가능한 한 많이 보존하는 과제를 제기한다. 우리는 view 수준과 token 수준 모두에서 입력 context를 수집하는 계층적 접근법인 \\texttt{KeyVT}를 제안한다. 구체적으로, pixel feature와 camera parameter를 결합하고 semantic content와 geometric position을 모두 고려하여 view 중요도를 평가함으로써, 공간적으로 일관되고 task 관련성이 높은 view를 얻는다. 나아가 선택된 view들 사이의 patch 중복을 해결하기 위해 optimal transport (OT) framework 하에서 대표 token을 식별한다. 여기서 view token과 key token은 embedding space에서 두 개의 discrete distribution으로 정식화된다. 이러한 key token은 OT distance를 최소화함으로써 모든 view feature를 포괄할 것으로 기대된다. 우리는 세 가지 널리 사용되는 benchmark에서 본 framework를 평가했으며, 기존 tuning-free 방법보다 유의미한 향상을 보이고 training-based 접근법과도 견줄 만한 성능을 달성함을 확인했다.

"},{"id":"61703","en":"WestWorld: A Knowledge-Encoded Scalable Trajectory World Model for Diverse Robotic Systems","ko":"WestWorld: 다양한 로봇 시스템을 위한 지식 인코딩 확장 가능 Trajectory World Model","authors":"Yuchen Wang, Jiangtao Kong, Sizhe Wei, Xiaochang Li, Haohong Lin, Hongjue Zhao, Tianyi Zhou, Lu Gan, Huajie Shao","pos":"#700","link":"https://openreview.net/forum?id=ncRRCG4BfP","abs":"

Trajectory world models play a crucial role in robotic dynamics learning, planning, and control. While recent works have explored trajectory world models for diverse robotic systems, they struggle to scale to a large number of distinct system dynamics and overlook domain knowledge of physical structures. To address these limitations, we introduce WestWorld, a knoWledge-Encoded Scalable Trajectory World model for diverse robotic systems. To tackle the scalability challenge, we propose a novel system-aware Mixture-of-Experts (Sys-MoE) that dynamically combines and routes specialized experts for different robotic systems via a learnable system embedding. To further enhance zero-shot generalization, we incorporate domain knowledge of robot physical structures by introducing a structural embedding that aligns trajectory representations with morphological information. After pretraining on 89 complex environments spanning diverse morphologies across both simulation and real-world settings, WestWorld achieves significant improvements over competitive baselines in zero- and few-shot trajectory prediction. Additionally, it shows strong scalability across a wide range of robotic environments and significantly improves performance on downstream model-based control for different robots. Finally, we deploy our model on a real-world Unitree Go1, where it demonstrates stable locomotion performance (see our demo on the website).

","absKo":"

Trajectory world model은 로봇 dynamics learning, planning, control에서 핵심적인 역할을 한다. 최근 연구들은 다양한 robotic system을 위한 trajectory world model을 탐구해 왔지만, 서로 다른 system dynamics가 많은 환경으로 확장하는 데 어려움을 겪고 있으며 물리적 구조에 대한 domain knowledge를 간과하고 있다. 이러한 한계를 해결하기 위해 우리는 다양한 robotic system을 위한 knoWledge-Encoded Scalable Trajectory World model인 WestWorld를 제안한다. 확장성 문제를 해결하기 위해, 우리는 learnable system embedding을 통해 서로 다른 robotic system에 특화된 expert를 동적으로 결합하고 라우팅하는 새로운 system-aware Mixture-of-Experts (Sys-MoE)를 제안한다. zero-shot generalization을 더욱 향상시키기 위해, robot의 physical structure에 대한 domain knowledge를 구조적 embedding으로 도입하여 trajectory representation과 morphological information을 정렬한다. 시뮬레이션과 실제 환경을 아우르는 서로 다른 morphology의 89개 복잡한 환경에서 사전학습한 후, WestWorld는 zero-shot 및 few-shot trajectory prediction에서 경쟁적 baseline을 크게 능가한다. 또한 광범위한 robotic environment 전반에서 뛰어난 확장성을 보이며, 서로 다른 robot에 대한 downstream model-based control 성능도 크게 향상시킨다. 마지막으로, 실제 Unitree Go1에 모델을 배치했으며 안정적인 locomotion 성능을 보여주었다(웹사이트의 demo를 참조).

"},{"id":"62813","en":"Vision-Language-Action Pretraining from Large-Scale Human Videos","ko":"대규모 인간 비디오 기반 Vision-Language-Action 사전학습","authors":"Hao Luo, Yicheng Feng, Wanpeng Zhang, Sipeng Zheng, Ye Wang, Haoqi Yuan, jiazheng liu, Chaoyi Xu, Haiweng Xu, Qin Jin, Zongqing Lu","pos":"#701","link":"https://openreview.net/forum?id=d2TFven78P","abs":"

Existing Vision-Language-Action (VLA) models struggle with complex manipulation tasks requiring high dexterity and generalization, primarily due to their reliance on synthetic data with significant sim-to-real gaps or limited teleoperated demonstrations. To address this bottleneck, we propose leveraging human hands as a manipulator template, capitalizing on the rich dexterity and scalability present in web data of human manipulation. Our approach introduces physical instruction tuning, a novel training paradigm that combines large-scale VLA pretraining from human videos, perspective spatial alignment for reasoning in a unified physical space, and post-training adaptation in physical environments. Additionally, we introduce a part-level motion tokenization method that achieves millimeter-level reconstruction accuracy to model precise hand trajectories serving as scalable motion primitives. To support our paradigm, we develop a comprehensive data curation pipeline that integrates heterogeneous sources into a large-scale dataset with millions of motion-based instructional instances. Empirically, our model demonstrates superior performance in hand motion generation and instruction following, adhering to favorable scaling laws with respect to model and data sizes. Importantly, we demonstrate promising capabilities to robotic dexterous manipulation, validating the effectiveness of bridging the human-robot embodiment gap.

","absKo":"

기존 Vision-Language-Action (VLA) model은 높은 dexterity와 generalization을 요구하는 복잡한 manipulation task에서 여전히 어려움을 겪는다. 주된 이유는 큰 sim-to-real gap을 지닌 synthetic data 또는 제한적인 teleoperated demonstration에 의존하기 때문이다. 이 병목을 해결하기 위해, 우리는 인간 손을 manipulator template로 활용하는 방안을 제안하며, human manipulation의 web data에 존재하는 풍부한 dexterity와 scalability를 적극 활용한다. 우리의 접근법은 physical instruction tuning을 도입한다. 이는 인간 video로부터의 대규모 VLA pretraining, 통합된 physical space에서 추론하기 위한 perspective spatial alignment, 그리고 physical environment에서의 post-training adaptation을 결합한 새로운 training paradigm이다. 또한 우리는 part-level motion tokenization method를 도입하여 millimeter 수준의 reconstruction accuracy로 정밀한 손 trajectory를 모델링하며, 이를 확장 가능한 motion primitive로 사용한다. 이 paradigm을 지원하기 위해, 우리는 이질적인 source를 통합하여 수백만 개의 motion-based instruction instance를 포함하는 대규모 dataset으로 만드는 포괄적 data curation pipeline을 구축했다. 실험적으로, 우리 model은 손 motion generation과 instruction following에서 더 우수한 성능을 보이며, model 및 data size에 대해 유리한 scaling law를 따른다. 특히, 우리는 로봇 dexterous manipulation에 대한 유망한 capability를 입증하여 human-robot embodiment gap을 메우는 접근의 효과를 검증한다.

"},{"id":"60500","en":"TapSampling: Inference-Time Sampling with a Task-Progress-Understanding Verifier for Robotic Manipulation","ko":"TapSampling: Robotic Manipulation을 위한 Task-Progress-Understanding Verifier 기반 Inference-Time Sampling","authors":"Sizhe Zhao, Shengping Zhang, Shuo Yang, Weiyu Zhao, Shuigen Wang, Xiangyang Ji","pos":"#705","link":"https://openreview.net/forum?id=zeIa4sZCD1","abs":"

Existing embodied control research demonstrates remarkable performance improvements by scaling training data and model size. We instead explore inference-time strategy as an alternative axis. Non-deterministic generative models, such as diffusion and autoregressive models, have been widely adopted in the field of embodied control. However, the single-shot inference paradigm limits their performance. In this paper, we propose \\textbf{TapSampling}, a plug-and-play framework for inference-time sampling. First, we introduce an Action-VAE to represent actions in a low-dimensional latent space. The Action-VAE maps initial actions from policies into a compressed posterior distribution, from which an arbitrary number of latent samples can be drawn and decoded into candidate actions that approximately follow the true action distribution. Second, we formulate action verification as task-progress outcome prediction and train the verifier by leveraging the intrinsic sequential information of robotic datasets. The predicted scores have clear semantic grounding, enabling interpretable action selection. Furthermore, TapSampling is a policy-agnostic framework. Extensive experiments in both simulated and real-world environments demonstrate that our method effectively improves multiple generalist policies substantially without further finetuning the policy models.

","absKo":"

기존 embodied control 연구는 training data와 model size를 확장함으로써 놀라운 성능 향상을 보여준다. 우리는 대신 inference-time strategy를 대안적인 축으로 탐구한다. diffusion과 autoregressive model 같은 비결정적 생성 모델은 embodied control 분야에서 널리 채택되어 왔다. 그러나 single-shot inference paradigm은 이들의 성능을 제한한다. 본 논문에서는 inference-time sampling을 위한 plug-and-play framework인 \\textbf{TapSampling}을 제안한다. 먼저, 우리는 action을 저차원 latent space에서 표현하기 위한 Action-VAE를 도입한다. Action-VAE는 policy로부터의 초기 action을 압축된 posterior distribution으로 매핑하며, 여기서 임의의 수의 latent sample을 뽑아 실제 action distribution을 근사적으로 따르는 candidate action으로 decode할 수 있다. 둘째, 우리는 action verification을 task-progress outcome prediction으로 정식화하고, robot dataset의 본질적인 sequential information을 활용하여 verifier를 학습한다. 예측된 score는 명확한 semantic grounding을 가지므로, 해석 가능한 action selection이 가능하다. 더 나아가 TapSampling은 policy-agnostic framework이다. simulated 및 real-world 환경 모두에서 수행한 대규모 실험은, 우리의 방법이 policy model을 추가로 finetuning하지 않고도 여러 generalist policy의 성능을 효과적으로 크게 향상시킴을 보여준다.

"},{"id":"65669","en":"Tabero: Learning Gentle Manipulation with Closed-Loop Force Feedback from Vision, Touch, and Language","ko":"Tabero: 시각·촉각·언어로부터의 폐루프 힘 피드백을 통한 부드러운 조작 학습","authors":"Qiwei Wu, Rui Zhang, Xin Xiang, Tao Li, Weihua Zhang, Junjie Lai, Renjing Xu","pos":"#706","link":"https://openreview.net/forum?id=BOWWKD9xEF","abs":"

Tactile sensing is essential for robots to achieve human-like gentle manipulation capabilities. However, existing Vision-Language-Action (VLA) models struggle to exploit tactile feedback for gentle manipulation due to the scarcity of aligned vision-tactile-language data and the lack of effective closed-loop force feedback mechanisms. To address these challenges, we introduce Tabero, a benchmark and model suite for gentle, language-conditioned robotic manipulation that demands fine-grained contact force perception. First, the Tabero benchmark addresses the scarcity of tactile data by presenting a data-efficient pipeline that repurposes open-source robot manipulation trajectories to generate a diverse set of vision-tactile-language tasks, and establishes a multidimensional evaluation protocol that measures task success alongside physical interaction quality. Second, we propose Tabero-VTLA, a Vision-Tactile-Language-Action architecture featuring a decoupled force-position command interface; the resulting force-position commands are executed by a fixed hybrid controller to enable real-time, force-aware manipulation. Evaluated on Tabero, our model maintains high task success while reducing average grip force by over 70% under gentle instructions, demonstrating its ability to modulate interaction forces based on multimodal experience.

","absKo":"

촉각 sensing은 로봇이 사람과 유사한 섬세한 조작 능력을 달성하는 데 필수적이다. 그러나 기존 Vision-Language-Action(VLA) 모델은 정렬된 vision-tactile-language 데이터의 부족과 효과적인 closed-loop force feedback mechanism의 부재로 인해, gentle manipulation에 촉각 피드백을 충분히 활용하지 못한다. 이러한 문제를 해결하기 위해, 우리는 정교한 contact force perception을 요구하는 언어 조건부 섬세 로봇 조작을 위한 benchmark와 model suite인 Tabero를 소개한다. 먼저, Tabero benchmark는 오픈소스 robot manipulation trajectory를 재활용하여 다양한 vision-tactile-language task를 생성하는 data-efficient pipeline을 제시함으로써 촉각 데이터의 부족 문제를 해결하고, task success와 물리적 상호작용 품질을 함께 측정하는 다차원 평가 프로토콜을 구축한다. 둘째, 우리는 분리된 force-position command interface를 갖는 Vision-Tactile-Language-Action architecture인 Tabero-VTLA를 제안하며, 생성된 force-position command는 고정된 hybrid controller에 의해 실행되어 실시간 force-aware manipulation을 가능하게 한다. Tabero에서 평가한 결과, 우리 모델은 높은 task success를 유지하면서 gentle instruction 하에서 평균 grip force를 70% 이상 감소시켜, 다중 모달 경험에 기반해 상호작용 힘을 조절할 수 있음을 보여준다.

"},{"id":"63054","en":"Structured 4D Latent World Model for Robot Planning","ko":"로봇 계획을 위한 구조화된 4D 잠재 월드 모델","authors":"Zhiyi Li, Peilin Wu, Xiaoshen Han, Ruojin Cai, Yilun Du","pos":"#707","link":"https://openreview.net/forum?id=aXAgpGfHGc","abs":"

Learned world models are emerging as a powerful paradigm in robotics, offering a promising path toward task generalization, long-horizon planning, and flexible decision-making. However, prevailing approaches often operate on 2D video sequences, inherently lacking the 3D geometric understanding necessary for precise spatial reasoning and physical consistency. We introduce a Structured 4D Latent World Model, which predicts the evolution of a scene’s 3D structure in a structured latent space conditioned on observations and textual instructions. Our representation encodes the scene holistically and can be decoded into diverse 3D formats, enabling a more complete and physically consistent scene understanding. This structured 4D latent world model serves as a planner, generating future scenes that are translated into executable actions by a goal-conditioned inverse dynamics module. Experiments demonstrate that our model generates futures with superior visual quality, physical consistency, and multi-view coherence compared to state-of-the-art video-based planners. Consequently, our full planning pipeline achieves superior performance on complex manipulation tasks, exhibits robust generalization to novel visual conditions, and proves effective on real-world robotic platforms. Our website is available at https://icml2026-18617.github.io/.

","absKo":"

학습된 world model은 robotics에서 강력한 paradigm으로 부상하고 있으며, task generalization, long-horizon planning, flexible decision-making으로 가는 유망한 경로를 제공한다. 그러나 기존 접근법은 주로 2D video sequence에서 동작하므로, 정밀한 spatial reasoning과 물리적 일관성에 필요한 3D geometric understanding이 본질적으로 부족하다. 우리는 관측과 textual instruction에 조건화된 structured latent space에서 scene의 3D structure evolution을 예측하는 Structured 4D Latent World Model을 소개한다. 우리의 representation은 scene을 holistic하게 인코딩하며 다양한 3D format으로 decode될 수 있어, 더 완전하고 물리적으로 일관된 scene understanding을 가능하게 한다. 이 structured 4D latent world model은 planner로서 기능하며, 미래 scene을 생성하고 이를 goal-conditioned inverse dynamics module이 실행 가능한 action으로 변환한다. 실험 결과, 우리 model은 state-of-the-art video-based planner와 비교해 더 뛰어난 visual quality, physical consistency, multi-view coherence를 갖는 future를 생성함을 보인다. 그 결과, 우리의 전체 planning pipeline은 복잡한 manipulation task에서 더 우수한 성능을 달성하고, 새로운 visual condition에 대한 강건한 generalization을 보이며, 실제 robotic platform에서도 효과적임이 입증되었다. 우리 website는 https://icml2026-18617.github.io/ 에서 확인할 수 있다.

"},{"id":"63548","en":"SpikeVLA: Vision-Language-Action Models with Spiking Neural Networks","ko":"SpikeVLA: Spiking Neural Networks를 활용한 Vision-Language-Action 모델","authors":"Ruiqi Song, Dujun Nie, Siyu Teng, Baiyong Ding, Xiaotong Zhang, Dong Li, Chenming Zhang, Yuchen Li, Hangbin Wu, Long Chen","pos":"#708","link":"https://openreview.net/forum?id=W86R5sIsxE","abs":"

Vision-Language-Action (VLA) models have become a central paradigm for embodied intelligence. However, most existing approaches are built on large-scale Transformers, resulting in substantial inference latency and energy consumption that limit their practical deployment in low-power, real-time scenarios. We propose SpikeVLA, an end-to-end spiking VLA framework for embodied navigation with energy-efficient inference, consisting of three key components. (i) a spiking vision encoder, Spike-V, that replaces dense continuous computation with event-driven spiking representations to reduce the energy cost of visual representation learning, (ii) a multimodal spiking large language model, Spike-L, that reformulates cross-modal reasoning with spiking dynamics and token-level event-driven sparsity to further lower inference overhead, and (iii) a spiking action policy network, Spike-A, that uses Laplacian-kernel population coding and end-to-end reinforcement learning to produce stable, robust continuous control under low-energy constraints. Experiments on multimodal interaction and robotic control tasks show that SpikeVLA significantly reduces energy consumption and computational overhead while maintaining competitive performance, highlighting its potential for low-power, real-time embodied intelligence.

","absKo":"

Vision-Language-Action (VLA) 모델은 embodied intelligence의 중심 paradigm이 되었다. 그러나 기존의 대부분 접근법은 대규모 Transformer를 기반으로 구축되어 있어, low-power real-time 시나리오에서의 실제 배치를 제한하는 상당한 inference latency와 energy consumption을 초래한다. 우리는 energy-efficient inference를 위한 end-to-end spiking VLA framework인 SpikeVLA를 제안하며, 이는 세 가지 핵심 구성 요소로 이루어진다. (i) spiking vision encoder인 Spike-V는 dense한 연속 계산을 event-driven spiking representation으로 대체하여 visual representation learning의 energy cost를 줄이고, (ii) multimodal spiking large language model인 Spike-L은 cross-modal reasoning을 spiking dynamics와 token-level event-driven sparsity로 재구성하여 inference overhead를 추가로 낮추며, (iii) spiking action policy network인 Spike-A는 Laplacian-kernel population coding과 end-to-end reinforcement learning을 사용하여 낮은 에너지 제약 하에서도 안정적이고 강건한 continuous control을 생성한다. multimodal interaction과 robotic control task에서의 실험은 SpikeVLA가 경쟁력 있는 성능을 유지하면서도 energy consumption과 computational overhead를 크게 줄임을 보여주며, low-power real-time embodied intelligence로서의 잠재력을 강조한다.

"},{"id":"66214","en":"Spatial Memory for Out-of-Vision Manipulation in Vision-Language-Action","ko":"Vision-Language-Action에서 시야 밖 조작을 위한 공간 기억","authors":"Pengteng Li, Weiyu Guo, He ZHANG, Tiefu Cai, Xiao He, Yandong Guo, Hui Xiong","pos":"#709","link":"https://openreview.net/forum?id=5i888dLp8N","abs":"

We introduce SOMA, the Spatial Memory framework for Out-of-Vision Manipulation in Vision-Language-Action (VLA) models. Most existing VLAs implicitly assume that task-relevant objects are always visible, leading to brittle and reactive behaviors when targets fall outside the camera’s field of view. SOMA addresses this limitation by equipping VLAs with a persistent spatial memory constructed from multi-view observations acquired via a movable head camera, enabling reasoning beyond the current visual frustum. The framework consists of three components: Spatial Memory Construction, which aggregates angular-wise observations into a unified spatial–semantic representation through scanning; Dynamic Memory Refinement, which maintains global consistency over time; and Contextual Memory Retrieval, which activates instruction-relevant spatial cues during manipulation. We evaluate SOMA on five challenging real-world out-of-vision manipulation tasks, including multi-step and dual-arm scenarios where target objects are initially invisible. Experimental results show that SOMA not only improves task success rates, but also induces qualitatively different manipulation behaviors, with faster target localization, reduced viewpoint search, and near one-shot grasping under partial observability. Additional experiments on RoboCasa GR1 and SimplerEnv further validate the effectiveness of SOMA’s memory design under conventional fully observable settings.

","absKo":"

우리는 Vision-Language-Action(VLA) 모델에서 시야 밖 조작(out-of-vision manipulation)을 위한 Spatial Memory 프레임워크인 SOMA를 소개한다. 기존 VLA의 대부분은 작업 관련 객체가 항상 보인다고 암묵적으로 가정하기 때문에, 대상이 카메라의 시야 밖으로 벗어나면 취약하고 반응적인 행동을 보인다. SOMA는 movable head camera를 통해 얻은 multi-view observation으로 구성된 지속적인 spatial memory를 VLA에 탑재함으로써 이 한계를 해결하고, 현재 시각 frustum을 넘어 추론할 수 있게 한다. 이 프레임워크는 세 가지 구성요소로 이루어진다. Spatial Memory Construction은 scanning을 통해 각도별 관측을 통합된 spatial-semantic representation으로 집계하고, Dynamic Memory Refinement는 시간에 걸친 global consistency를 유지하며, Contextual Memory Retrieval은 조작 중 지시문과 관련된 spatial cue를 활성화한다. 우리는 대상 객체가 처음에는 보이지 않는 multi-step 및 dual-arm 시나리오를 포함한 다섯 가지 어려운 실제 out-of-vision manipulation task에서 SOMA를 평가한다. 실험 결과, SOMA는 task success rate를 향상시킬 뿐 아니라, 더 빠른 target localization, 감소된 viewpoint search, 부분 관측하에서의 거의 one-shot grasping과 같은 질적으로 다른 조작 행동을 유도함을 보인다. RoboCasa GR1와 SimplerEnv에서의 추가 실험은 전통적인 fully observable setting에서도 SOMA의 memory design이 효과적임을 추가로 검증한다.

"},{"id":"61584","en":"SafeLab: An Interactive High-Fidelity Benchmark for Embodied Safety in Scientific Robotics","ko":"SafeLab: 과학 로보틱스의 Embodied Safety를 위한 대화형 High-Fidelity Benchmark","authors":"Fengshuo Bai, Yufeng Li, Ruihai Wu, Peishuo Wang, Yuhan Wang, Bernie Zhu, Yuanfei Wang, Tawei Chou, Gao, Runchuan Zhu, Ying Wen, Yaodong Yang, Yuanpei Chen","pos":"#711","link":"https://openreview.net/forum?id=ojwpa0wXo9","abs":"

Laboratory automation driven by scientific embodied agents represents a critical frontier in modern laboratories. Unlike conventional robotic domains, laboratory environments impose zero-tolerance constraints on manipulation precision and collision, as minor deviations can lead to irreversible chemical hazards or equipment damage. This naturally makes the automated laboratory an ideal testbed for advancing embodied safety. However, existing benchmarks predominantly feature high-tolerance manipulation tasks where intermediate failures are largely reversible. More critically, current Vision-Language-Action (VLA) models trained via static imitation learning cannot satisfy these strict constraints. Because they merely mimic successful demonstrations, they lack the ability to recover from execution drift, leading to catastrophic compounding errors in precision-critical domains. Overcoming this limitation requires transitioning from static datasets to interactive environments that support Reinforcement Learning (RL) for dynamic error recovery. To this end, we introduce SafeLab, a generative simulation benchmark designed for the full lifecycle of safe robot learning. Grounded in a high-fidelity chemistry lab, our framework integrates an {LLM engine} for procedural task synthesis, an {automated expert} for scalable demonstration collection, and an {interactive environment} for continuous RL refinement. Leveraging this infrastructure, we release a dataset of 6,000+ {complex trajectories} to evaluate state-of-the-art VLA models. Experiments reveal that current embodied agents fail significantly under these safety constraints. In contrast, our RL post-training pipeline enables agents to learn active error correction, mitigating hazardous failures and improving success rates by 37\\%, thereby establishing SafeLab as a critical platform for developing reliable and safe generalist agents.

","absKo":"

scientific embodied agent가 이끄는 laboratory automation은 현대 실험실에서 중요한 최전선이다. 기존 robotic domain과 달리, laboratory environment는 manipulation precision과 collision에 대해 zero-tolerance constraint를 부과하며, 작은 편차도 돌이킬 수 없는 chemical hazard나 장비 손상으로 이어질 수 있다. 이러한 특성은 자동화된 실험실을 embodied safety를 발전시키기 위한 이상적인 testbed로 만든다. 그러나 기존 benchmark는 대부분 중간 실패가 대체로 되돌릴 수 있는 high-tolerance manipulation task를 중심으로 구성되어 있다. 더 중요하게는, static imitation learning으로 학습된 현재의 Vision-Language-Action (VLA) model은 이러한 엄격한 constraint를 만족시킬 수 없다. 이들은 성공적인 demonstration을 단순히 모방할 뿐 execution drift로부터 복구하는 능력이 부족하여, precision-critical domain에서 치명적인 error 누적을 초래한다. 이러한 한계를 극복하려면 static dataset에서 벗어나 dynamic error recovery를 지원하는 Reinforcement Learning (RL)을 위한 interactive environment로 전환해야 한다. 이를 위해 우리는 safe robot learning의 전체 lifecycle을 위한 generative simulation benchmark인 SafeLab을 제안한다. 고정밀 chemistry lab을 기반으로, 우리의 framework는 procedural task synthesis를 위한 {LLM engine}, 확장 가능한 demonstration 수집을 위한 {automated expert}, 지속적인 RL refinement를 위한 {interactive environment}를 통합한다. 이 인프라를 활용해 우리는 최첨단 VLA model을 평가하기 위한 6,000개 이상의 {complex trajectories} 데이터셋을 공개한다. 실험은 현재의 embodied agent가 이러한 safety constraint 하에서 크게 실패함을 보여준다. 반면, 우리의 RL post-training pipeline은 agent가 능동적인 error correction을 학습하도록 하여 위험한 실패를 완화하고 성공률을 37\\% 향상시켰으며, 이를 통해 SafeLab을 신뢰할 수 있고 안전한 generalist agent 개발을 위한 중요한 플랫폼으로 확립한다.

"},{"id":"66066","en":"SCALE: Self-uncertainty Conditioned Adaptive Looking and Execution for Vision-Language-Action Models","ko":"SCALE: Vision-Language-Action 모델을 위한 자기 불확실성 조건부 적응형 관찰 및 실행","authors":"Hyeonbeom Choi, Daechul Ahn, Youhan Lee, Taewook Kang, Seongwon Cho, Jonghyun Choi","pos":"#713","link":"https://openreview.net/forum?id=7MlfE2Da2W","abs":"

Vision-Language-Action (VLA) models have emerged as a promising paradigm for general-purpose robotic control, with test-time scaling (TTS) gaining attention to enhance robustness beyond training. However, existing TTS methods for VLAs require additional training, verifiers, and multiple forward passes, making them impractical for deployment. Moreover, they intervene only at action decoding while keeping visual representations fixed—insufficient under perceptual ambiguity, where reconsidering how to perceive is as important as deciding what to do. To address these limitations, we propose SCALE, a simple inference strategy that jointly modulates visual perception and action based on 'self-uncertainty', inspired by uncertainty-driven exploration in Active Inference theory—requiring no additional training, no verifier, and only a single forward pass. SCALE broadens exploration in both perception and action under high uncertainty, while focusing on exploitation when confident—enabling adaptive execution across varying conditions. Experiments on simulated and real-world benchmarks demonstrate that \\method improves state-of-the-art VLAs and outperforms existing TTS methods while maintaining single-pass efficiency.

","absKo":"

Vision-Language-Action (VLA) 모델은 범용 로봇 제어를 위한 유망한 패러다임으로 부상했으며, test-time scaling (TTS)은 학습을 넘어서는 강건성을 높이기 위한 방법으로 주목받고 있다. 그러나 기존의 VLA용 TTS 방법은 추가 학습, verifier, 그리고 여러 번의 forward pass를 필요로 하므로 실제 배포에는 비실용적이다. 더 나아가, 이들 방법은 action decoding에만 개입하고 visual representation은 고정한 채 두는데, 이는 지각적 모호성 하에서는 불충분하다. 무엇을 할지 결정하는 것만큼이나 어떻게 지각할지를 다시 고려하는 것이 중요하기 때문이다. 이러한 한계를 해결하기 위해, 우리는 Active Inference theory의 uncertainty-driven exploration에서 영감을 받은 간단한 inference strategy인 SCALE을 제안한다. SCALE은 추가 학습, verifier, 단일 forward pass만으로 동작하며, 'self-uncertainty'에 기반해 visual perception과 action을 jointly modulate한다. SCALE은 불확실성이 높을 때는 perception과 action 모두에서 exploration을 확장하고, 확신이 있을 때는 exploitation에 집중함으로써 다양한 조건에서 적응적 실행을 가능하게 한다. 시뮬레이션 및 실제 world benchmark 실험은 \\method가 state-of-the-art VLA를 향상시키고, 기존 TTS 방법들을 상회하면서도 single-pass 효율성은 유지함을 보여준다.

"},{"id":"61401","en":"RoboOmni: Actions Are Just Another Modality for Your Vision-Language Models","ko":"RoboOmni: Actions도 Vision-Language Models를 위한 또 하나의 Modality일 뿐이다","authors":"Dong Wang, Zilong Chen, Jirong Liu, Ziqing Qiao, Xin Xiao, Bingyi Kang, Hongtao Wu, Xiao Ma, Tao Kong, Huaping Liu","pos":"#714","link":"https://openreview.net/forum?id=qdXOfyGMuB","abs":"Integrating Vision-Language Models (VLMs) into robotics has facilitated the development of generalizable Vision-Language Action (VLA) policies. However, unified discrete frameworks lag behind decoupled continuous designs due to limitations in action chunking and temporal modeling. To address this, we introduce **RoboOmni**, a unified multi-modal next-token prediction framework. Challenging the assumption that continuous modeling is essential for high-performance manipulation, **RoboOmni** demonstrates that *actions are just another modality* capable of being effectively modeled discretely. At the core of our method is Multi-Token Action Prediction (MTAP), which integrates action chunking directly into the discrete tokenizer. This design resolves temporal modeling bottlenecks and significantly reduces distribution shift between training and inference. By preserving the native VLM training and inference pipeline, **RoboOmni** naturally benefits from large-scale multimodal co-training and modern decoding optimizations. Extensive evaluations on the CALVIN, SimplerEnv, and real-world platforms confirm that **RoboOmni** establishes new state-of-the-art performance, significantly outperforming diffusion-based baselines such as $\\pi_0$. Notably, combining our proposed MTAP with the FAST tokenizer achieves a 94.4\\% average success rate on CALVIN, while the Bin tokenizer implementation attains a 27$\\times$ inference speedup compared to OpenVLA.","absKo":"Vision-Language Models (VLMs)를 robotics에 통합함으로써 일반화 가능한 Vision-Language Action (VLA) policy의 개발이 가능해졌다. 그러나 unified discrete framework는 action chunking과 temporal modeling의 한계로 인해 decoupled continuous design에 뒤처진다. 이를 해결하기 위해 우리는 통합 multi-modal next-token prediction framework인 RoboOmni를 소개한다. 고성능 manipulation에 continuous modeling이 필수라는 가정에 도전하며, RoboOmni는 *actions are just another modality*로서 discrete하게 효과적으로 모델링될 수 있음을 보여준다. 이 방법의 핵심은 Multi-Token Action Prediction (MTAP)으로, action chunking을 discrete tokenizer에 직접 통합한다. 이 설계는 temporal modeling 병목을 해소하고 training과 inference 간 distribution shift를 크게 줄인다. native VLM training 및 inference pipeline을 유지함으로써 RoboOmni는 대규모 multimodal co-training과 현대적 decoding optimization의 이점을 자연스럽게 누린다. CALVIN, SimplerEnv, 그리고 real-world platform에서의 광범위한 평가를 통해 RoboOmni가 새로운 state-of-the-art 성능을 확립하고, $\\pi_0$와 같은 diffusion-based baseline을 크게 능가함이 확인되었다. 특히, 제안한 MTAP을 FAST tokenizer와 결합하면 CALVIN에서 평균 94.4\\% success rate를 달성하며, Bin tokenizer 구현은 OpenVLA 대비 27$\\times$의 inference 속도 향상을 달성한다."},{"id":"62543","en":"RoboFlow4D: A Lightweight Flow World Model Toward Real-Time Flow-Guided Robotic Manipulation","ko":"RoboFlow4D: 실시간 Flow-guided 로봇 조작을 위한 경량 Flow World Model","authors":"Sixu Lin, Huaiyuan Xu, Junliang Chen, Zhuohao Li, Guangming Wang, Yixiong Jing, Sheng Xu, Runyi Zhao, Brian Sheil, Lap-Pui Chau, Guiliang Liu","pos":"#715","link":"https://openreview.net/forum?id=fh6XBnjFlv","abs":"

Planning and acting in 3D environments is a fundamental capability for robotic manipulation in the real world. Although prior work has explored predictive flow planners to guide 3D manipulation, existing approaches often rely on modular pipelines stacking multiple submodels, resulting in high computational overhead and limited real-time performance. To address these challenges, we introduce RoboFlow4D, a lightweight flow world model that unifies perception and planning by estimating temporal motion in physical 3D space. As an end-to-end framework, RoboFlow4D directly predicts multi-frame 3D flows from visual observations and textual instructions, providing explicit flow-based planning to guide action generation. This design allows seamless integration with general action policies, forming an efficient observation–planning–execution closed loop. Through slow–fast collaboration between flow prediction and action control, RoboFlow4D enables real-time and resource-efficient manipulation. Extensive experiments in both simulation and real-world settings demonstrate that RoboFlow4D consistently improves manipulation success rates and computational efficiency, advancing flow-guided planning for embodied intelligence.

","absKo":"

3D 환경에서의 계획과 행동은 현실 세계 로봇 manipulation을 위한 근본적인 능력이다. 이전 연구들이 3D manipulation을 유도하기 위해 predictive flow planners를 탐구해 왔지만, 기존 접근은 여러 submodel을 쌓는 modular pipeline에 의존하는 경우가 많아 계산 오버헤드가 크고 실시간 성능이 제한된다. 이러한 문제를 해결하기 위해 우리는 물리적 3D 공간에서의 temporal motion을 추정함으로써 perception과 planning을 통합하는 경량 flow world model인 RoboFlow4D를 제안한다. end-to-end framework로서 RoboFlow4D는 시각 관측과 텍스트 지시로부터 multi-frame 3D flows를 직접 예측하고, action generation을 유도하기 위한 explicit flow-based planning을 제공한다. 이 설계는 일반적인 action policy와의 매끄러운 통합을 가능하게 하여, 효율적인 observation–planning–execution closed loop를 형성한다. flow prediction과 action control 간의 slow–fast collaboration을 통해 RoboFlow4D는 실시간이며 자원 효율적인 manipulation을 가능하게 한다. simulation과 real-world setting 모두에서 수행한 광범위한 실험은 RoboFlow4D가 일관되게 manipulation 성공률과 계산 효율을 향상시켜, embodied intelligence를 위한 flow-guided planning을 진전시킴을 보여준다.

"},{"id":"66662","en":"FlatLab: A Unified Methodology Framework and Simulation-Based Benchmark for Robotic Manipulation of Flat Objects","ko":"FlatLab: 평면 물체의 로봇 조작을 위한 통합 방법론 프레임워크 및 시뮬레이션 기반 벤치마크","authors":"Xingyu Zhu, Wenshuo Han, Zhouyu Wang, Yuran Wang, Ruihai Wu, Hao Dong, Fan Tang, Hechang Chen, Hyung Jin Chang, Yixing Gao","pos":"#800","link":"https://openreview.net/forum?id=1dWG9PJSVp","abs":"

Robotic manipulation of flat objects is challenging due to the ungraspable configurations and strong variations in object geometry and material. Existing methods rely on heuristic pre-manipulation and are often evaluated in closed settings with limited generalization. We propose a unified framework that decouples the manipulation into a strategy generator and an action execution module. The strategy generator predicts appropriate manipulation strategies from object point clouds by learning strategy-centric, object-invariant representations via simulated data transformation and contrastive learning. Conditioned on the predicted strategy, the execution module decomposes long-horizon manipulation into reusable action primitives and dynamically composes them to generate stable trajectories. To enable systematic evaluation, we introduce FlatLab, a comprehensive simulation benchmark for robotic flat object manipulation. FlatLab provides high-fidelity physical simulation of diverse rigid and deformable flat objects, automated multi-modal data collection, and standardized task definitions and evaluation protocols. Experiments conducted in FlatLab demonstrate that our approach generalizes effectively to unseen objects and categories, outperforming existing baselines. The project page is provided at \\url{https://flatlab-web.github.io/}, and the code will be publicly released.

","absKo":"

평면 객체의 로봇 조작은 grasp할 수 없는 구성과 객체 기하 및 재질의 큰 변화 때문에 어렵다. 기존 방법들은 휴리스틱한 사전 조작에 의존하며, 종종 일반화가 제한된 폐쇄된 설정에서 평가된다. 우리는 조작을 전략 생성기와 행동 실행 모듈로 분해하는 통합 프레임워크를 제안한다. 전략 생성기는 시뮬레이션 데이터 변환과 contrastive learning을 통해 strategy-centric하고 object-invariant한 표현을 학습함으로써 객체 point cloud로부터 적절한 조작 전략을 예측한다. 예측된 전략에 조건화된 실행 모듈은 장기 지평의 조작을 재사용 가능한 action primitive로 분해하고, 이를 동적으로 조합하여 안정적인 trajectory를 생성한다. 체계적인 평가를 가능하게 하기 위해, 우리는 로봇 평면 객체 조작을 위한 종합적인 simulation benchmark인 FlatLab을 소개한다. FlatLab은 다양한 rigid 및 deformable 평면 객체에 대한 고충실도 물리 시뮬레이션, 자동화된 multi-modal 데이터 수집, 표준화된 task 정의와 평가 프로토콜을 제공한다. FlatLab에서 수행한 실험은 우리의 접근이 보지 못한 객체와 범주에 효과적으로 일반화하며 기존 baselines를 능가함을 보여준다. 프로젝트 페이지는 \\url{https://flatlab-web.github.io/}이며, 코드는 공개될 예정이다.

"},{"id":"63621","en":"From Pixels to Tokens: A Systematic Study of Latent Action Supervision for Vision-Language-Action Models","ko":"픽셀에서 토큰으로: Vision-Language-Action 모델을 위한 latent action supervision의 체계적 연구","authors":"Yihan Lin, Haoyang Li, Yang Li, Haitao Shen, Yihan Zhao, Chao Shao, Jing Zhang","pos":"#803","link":"https://openreview.net/forum?id=VMsumctGvg","abs":"

Latent actions serve as an intermediate representation that enables consistent modeling of vision-language-action (VLA) models across heterogeneous datasets. However, approaches to supervising VLAs with latent actions are fragmented and lack a systematic comparison. This work structures the study of latent action supervision from two perspectives: (i) regularizing the trajectory via image-based latent actions, and (ii) unifying the target space with action-based latent actions. Under a unified VLA baseline, we instantiate and compare four representative integration strategies. Our results reveal a formulation-task correspondence: image-based latent actions benefit long-horizon reasoning, whereas action-based latent actions excel at complex motor coordination. Furthermore, we find that directly supervising the VLM with discrete latent action tokens yields the most effective performance. Finally, our experiments offer initial insights into the benefits of latent action supervision in mixed-data, suggesting a promising direction for VLA training.

","absKo":"

Latent action은 heterogeneous dataset 전반에서 vision-language-action (VLA) model을 일관되게 모델링할 수 있게 해주는 intermediate representation으로 작동한다. 그러나 latent action을 이용한 VLA supervision 방법들은 분절되어 있으며, 체계적인 비교가 부족하다. 본 연구는 latent action supervision의 연구를 두 가지 관점에서 구조화한다: (i) image-based latent action을 통한 trajectory regularization, 그리고 (ii) action-based latent action을 통한 target space 통합. 통합된 VLA baseline 아래에서, 우리는 네 가지 대표적인 integration strategy를 구현하고 비교한다. 그 결과 formulation-task correspondence가 드러난다. image-based latent action은 long-horizon reasoning에 유리한 반면, action-based latent action은 복잡한 motor coordination에서 뛰어나다. 또한 discrete latent action token으로 VLM을 직접 supervision하는 것이 가장 효과적인 성능을 보인다는 사실을 발견했다. 마지막으로, 우리의 실험은 mixed-data에서 latent action supervision의 이점에 대한 초기 통찰을 제공하며, VLA training을 위한 유망한 방향을 시사한다.

"},{"id":"65375","en":"Generalizable and Actionable Parts Pose Estimation with Symmetry Annotation-Free Learning Strategy","ko":"대칭 어노테이션 없는 학습 전략을 통한 일반화 가능하고 실행 가능한 부품 포즈 추정","authors":"wenxiao chen, Xueyu Yuan, Liu Liu, Di Wu, Dan Guo","pos":"#805","link":"https://openreview.net/forum?id=EEgK2cD7gX","abs":"

Urgently needed generalizable robot object interaction and manipulation requires high-quality Cross-Category object perception. As a pioneer of this area, Generalizable and Actionable Parts (GAParts) understanding has attracted increasing attention from relevant researchers. However, most recent works either have insufficient design regarding the symmetry issue or require rich symmetry annotation, which severely impedes precise GAPart pose estimation in data-lacking scenarios. In this paper, we propose SAFAG, a novel Symmetry Annotation-Free framework for Generalizable and Actionable Parts Pose Estimation. Specifically, we suggest a stepwise refinement two-stage framework for candidate-to-final quaternion regression, and tackle the symmetry prediction as a probability distribution problem with self-supervised learning strategy. The experimental results demonstrate the superior performance and robustness of our SAFAG. We believe that our work has the enormous potential to be applied in many areas of embodied AI system.

","absKo":"

긴급히 필요한 일반화 가능한 robot object interaction 및 manipulation을 위해서는 고품질의 Cross-Category object perception이 필요하다. 이 분야의 선구자로서 Generalizable and Actionable Parts (GAParts) 이해는 관련 연구자들의 관심을 점점 더 많이 받아 왔다. 그러나 최근의 대부분 연구는 symmetry 문제에 대한 설계가 충분하지 않거나 풍부한 symmetry annotation을 요구하여, 데이터가 부족한 상황에서 정확한 GAPart pose estimation을 심각하게 저해한다. 본 논문에서는 Generalizable and Actionable Parts Pose Estimation을 위한 새로운 Symmetry Annotation-Free framework인 SAFAG를 제안한다. 구체적으로, 우리는 candidate-to-final quaternion regression을 위한 단계적 refinement two-stage framework를 제안하고, self-supervised learning strategy를 통해 symmetry prediction을 probability distribution problem으로 다룬다. 실험 결과는 SAFAG의 우수한 성능과 견고성을 보여준다. 우리는 우리의 연구가 embodied AI system의 여러 영역에 적용될 수 있는 큰 잠재력을 지닌다고 믿는다.

"},{"id":"61610","en":"Joint Navigation and Manipulation Planning with 3D Interaction Chains","ko":"3D Interaction Chains를 활용한 Joint Navigation and Manipulation Planning","authors":"Keming Zhang, Sixian Zhang, Xinhang Song, Hongyu Wang, Yiyao Wang, Yingjie Wang, Shuqiang Jiang","pos":"#809","link":"https://openreview.net/forum?id=oVB2xYWvpv","abs":"

Open-vocabulary mobile manipulation (OVMM) requires long-horizon navigation in unseen environments and object-centric manipulation. Most existing methods treat navigation and manipulation as separate stages, which can yield navigation endpoints that are poor for manipulation or manipulation-friendly poses that are globally inefficient. We address this mismatch with 3D Interaction Chains (3D-IC), a unified framework that couples multi-stage navigation and manipulation planning. 3D-IC maintains a shared 3D feature map for both skills, generates stage-aligned interaction waypoints, and links them into candidate multi-stage chains. A hierarchical policy then scores these chains by jointly considering feasibility (via VLM reasoning over waypoint-centric 3D features) and transition cost, selecting the best trade-off between success and path efficiency. The robot executes the next waypoint and replans as new observations arrive. Experiments in simulation and on a real Stretch 3 robot demonstrate consistent gains in both task success and trajectory efficiency.

","absKo":"

Open-vocabulary mobile manipulation(OVMM)은 미지의 환경에서의 장거리 내비게이션과 object-centric manipulation을 요구한다. 기존 방법의 대부분은 내비게이션과 조작을 서로 분리된 단계로 다루어, 조작에는 부적합한 내비게이션 종착점이나 전역적으로는 비효율적인 조작 친화적 자세를 초래할 수 있다. 우리는 이러한 불일치를 해결하기 위해 다단계 내비게이션과 조작 계획을 결합하는 통합 프레임워크인 3D Interaction Chains(3D-IC)를 제안한다. 3D-IC는 두 스킬 모두를 위해 공유 3D feature map을 유지하고, 단계 정렬(stage-aligned) interaction waypoint를 생성하며, 이를 후보 다단계 chain으로 연결한다. 이후 계층적 policy는 waypoint 중심 3D feature에 대한 VLM reasoning을 통한 실행 가능성과 transition cost를 함께 고려하여 이러한 chain을 점수화하고, 성공률과 경로 효율성 사이의 최적 균형을 선택한다. 로봇은 다음 waypoint를 실행하고 새로운 observation이 도착하면 재계획한다. simulation과 실제 Stretch 3 robot에서의 실험은 task success와 trajectory efficiency 모두에서 일관된 향상을 보여준다.

"},{"id":"64902","en":"MapDream: Task-Driven Map Learning for Vision-Language Navigation","ko":"MapDream: Vision-Language Navigation을 위한 Task-Driven Map 학습","authors":"Guoxin Lian, Shuo Wang, Yucheng Wang, Yongcai Wang, Maiyue Chen, kaihui.wang, Bo Zhang, Zhizhong Su, Deying Li, Zhaoxin Fan","pos":"#810","link":"https://openreview.net/forum?id=IkXFH6alZN","abs":"

Vision-Language Navigation (VLN) requires agents to follow natural language instructions in partially observed 3D environments, motivating map representations that aggregate spatial context beyond local perception. However, most existing approaches rely on hand-crafted maps constructed independently of the navigation policy. We argue that maps should instead be learned representations shaped directly by navigation objectives rather than exhaustive reconstructions. Based on this insight, we propose MapDream, a map-in-the-loop framework that formulates map construction as autoregressive bird’s-eye-view (BEV) image synthesis. The framework jointly learns map generation and action prediction, distilling environmental context into a compact three-channel BEV map that preserves only navigation-critical affordances. Supervised pre-training bootstraps a reliable mapping-to-control interface, while the autoregressive design enables end-to-end joint optimization through reinforcement fine-tuning. Experiments on R2R-CE and RxR-CE achieve state-of-the-art monocular performance, validating task-driven generative map learning.

","absKo":"

Vision-Language Navigation(VLN)은 agent가 부분적으로 관측되는 3D 환경에서 자연어 instruction을 따르도록 요구하며, 이에 따라 국소적 perception을 넘어 공간적 context를 집약하는 map representation이 필요하다. 그러나 기존의 대부분 접근법은 navigation policy와 독립적으로 구성된 hand-crafted map에 의존한다. 우리는 map이 exhaustive reconstruction이 아니라 navigation objective에 의해 직접 형성되는 learned representation이어야 한다고 주장한다. 이러한 통찰을 바탕으로, 우리는 map construction을 autoregressive bird’s-eye-view(BEV) image synthesis로 정식화하는 map-in-the-loop framework인 MapDream을 제안한다. 이 framework는 map generation과 action prediction을 공동으로 학습하여, 환경 context를 navigation에 필수적인 affordance만 보존하는 compact한 3-channel BEV map으로 distill한다. supervised pre-training은 신뢰할 수 있는 mapping-to-control interface를 bootstrap하고, autoregressive design은 reinforcement fine-tuning을 통한 end-to-end joint optimization을 가능하게 한다. R2R-CE와 RxR-CE에서의 실험은 state-of-the-art monocular performance를 달성하며, task-driven generative map learning을 검증한다.

"},{"id":"64538","en":"Neural Implicit Action Fields: From Discrete Waypoints to Continuous Functions for Vision-Language-Action Models","ko":"Neural Implicit Action Fields: Vision-Language-Action 모델을 위한 이산 경유점에서 연속 함수로","authors":"Haoyun Liu, Jianzhuang Zhao, Xinyuan Chang, Tianle Shi, Chuanzhang Meng, Jiayuan Tan, Feng Xiong, Tong Lin, Dongjie Huo, Mu Xu, SongLin Dong, Zhiheng Ma, Yihong Gong, Sheng Zhong","pos":"#811","link":"https://openreview.net/forum?id=MWEh9IPz8N","abs":"

Despite the rapid progress of Vision-Language-Action (VLA) models, the prevailing paradigm of predicting discrete waypoints remains fundamentally misaligned with the intrinsic continuity of physical motion. This discretization imposes rigid sampling rates, lacks high-order differentiability, and introduces quantization artifacts that hinder precise, compliant interaction. We propose Neural Implicit Action Fields (NIAF), a paradigm shift that reformulates action prediction from discrete waypoints to continuous action function regression. By utilizing an MLLM as a hierarchical spectral modulator over a learnable motion prior, NIAF synthesizes infinite-resolution trajectories as continuous-time manifolds. This formulation enables Analytical Differentiability, allowing for explicit supervision of velocity, acceleration, and jerk to ensure mathematical consistency and physical plausibility. Our approach achieves state-of-the-art results on CALVIN and LIBERO benchmarks across diverse backbones. Furthermore, real-world experiments demonstrate that NIAF enables stable impedance control, bridging the gap between high-level semantic understanding and low-level dynamic execution.

","absKo":"

Vision-Language-Action (VLA) model의 급속한 발전에도 불구하고, 이산적인 waypoint를 예측하는 지배적 패러다임은 물리적 운동의 본질적 연속성과 근본적으로 맞지 않는다. 이러한 이산화는 경직된 sampling rate를 강요하고, 고차 미분 가능성이 부족하며, 정밀하고 순응적인 상호작용을 방해하는 quantization artifact를 유발한다. 우리는 discrete waypoint에서 continuous action function regression으로 action prediction을 재정식화하는 패러다임 전환인 Neural Implicit Action Fields (NIAF)를 제안한다. 학습 가능한 motion prior 위에서 MLLM을 계층적 spectral modulator로 활용함으로써, NIAF는 continuous-time manifold로서 무한 해상도의 trajectory를 합성한다. 이 formulation은 Analytical Differentiability를 가능하게 하여, 속도, 가속도, jerk에 대한 명시적 supervision을 통해 수학적 일관성과 물리적 타당성을 보장한다. 우리의 접근법은 다양한 backbone에 걸쳐 CALVIN 및 LIBERO benchmark에서 state-of-the-art 결과를 달성한다. 나아가, 실제 세계 실험은 NIAF가 안정적인 impedance control을 가능하게 하여, 고수준 의미 이해와 저수준 동적 실행 사이의 간극을 메운다는 것을 보여준다.

"},{"id":"63338","en":"Reflex: Real-Time Vision-Language-Action Control through Streaming Inference","ko":"Reflex: 스트리밍 추론을 통한 실시간 Vision-Language-Action 제어","authors":"Yuanchun Guo, Bingyan Liu","pos":"#815","link":"https://openreview.net/forum?id=XzsZK1hvpv","abs":"Flow matching Vision-Language-Action (VLA) models promise precise continuous control, but their iterative denoising nature introduces fundamental incompatibilities with real-time robotics: global timestep injection invalidates KV-caching, forcing a choice between slow $O(N^2)$ re-computation or mathematically incorrect cache reuse. We present \\textbf{Reflex}, a framework that enables \\textit{real-time streaming inference} for flow matching policies by exploiting the \\textit{Timestep-Invariance Property}---that perception encoders are functionally independent of the denoising loop. Reflex partitions the attention context into static, sliding, and dynamic regions, enabling $O(1)$ incremental cache updates that guarantee outputs identical to full-batch inference. To ensure stability under continuous high-frequency inference, we introduce \\textit{AdaRMSNorm}, an adaptive normalization layer that prevents BFloat16 numerical collapse by gating on flow phase. We further maximize throughput through an \\textit{async pipeline} that decouples visual encoding from action generation, combined with \\textit{operator fusion} that reduces kernel overhead. On LIBERO and Kinetix benchmarks, Reflex achieves a 2.58$\\times$ inference speedup and 50Hz stable streaming, reducing reaction latency by up to 54\\% and enabling efficient deployment without performance degradation.","absKo":"Flow matching Vision-Language-Action(VLA) model은 정밀한 continuous control을 약속하지만, 그 반복적 denoising 특성은 real-time robotics와 근본적으로 맞지 않는다. 전역 timestep 주입은 KV-caching을 무효화하여, 느린 $O(N^2)$ 재계산과 수학적으로 부정확한 cache 재사용 사이에서 선택을 강요한다. 우리는 \\textbf{Reflex}를 제안하는데, 이는 perception encoder가 denoising loop와 functionally independent하다는 \\textit{Timestep-Invariance Property}를 활용하여 flow matching policy에 \\textit{real-time streaming inference}를 가능하게 하는 framework이다. Reflex는 attention context를 static, sliding, dynamic region으로 분할하여, full-batch inference와 동일한 출력을 보장하는 $O(1)$ incremental cache update를 가능하게 한다. continuous high-frequency inference 하에서의 안정성을 보장하기 위해, flow phase에 따라 gating하여 BFloat16 numerical collapse를 방지하는 adaptive normalization layer인 \\textit{AdaRMSNorm}을 도입한다. 또한 visual encoding과 action generation을 분리하는 \\textit{async pipeline}과 kernel overhead를 줄이는 \\textit{operator fusion}을 결합하여 throughput을 극대화한다. LIBERO와 Kinetix benchmark에서 Reflex는 2.58$\\times$ inference speedup과 50Hz의 안정적인 streaming을 달성하며, reaction latency를 최대 54\\%까지 줄이고 performance degradation 없이 효율적인 deployment를 가능하게 한다."},{"id":"63193","en":"Embodied-DETR: End-to-End Temporal 3D Object Detection in Egocentric Views","ko":"Embodied-DETR: 자기중심 시점에서의 종단간 시간적 3D 객체 탐지","authors":"Ziheng Ding, Xiaze Zhang, Yuejie Zhang, lifeng chen, Rui Feng","pos":"#901","link":"https://openreview.net/forum?id=ZEODIzlKwZ","abs":"

Embodied 3D object detection is a fundamental perception capability for embodied agents, where observations are partial, heavily occluded, and sequential, requiring modeling of temporal continuity. However, existing benchmarks and methods are primarily designed for fully reconstructed global scenes and fail to capture temporal scene context and instance evolution in first-person perception. We introduce Embodied-Det, a new benchmark for egocentric 3D object detection that evaluates detection accuracy, temporal stability, and consistency under embodied settings. Building on this benchmark, we propose Embodied-DETR, an end-to-end temporal detection framework that models scene-level context and instance-level continuity through two complementary temporal modules, Scene-aware Feature Aggregation and Instance-aware Query Embedding. Experiments on Embodied-Det show that existing methods suffer substantial performance degradation in egocentric temporal settings, while Embodied-DETR achieves superior accuracy and temporal consistency, demonstrating the effectiveness of temporal modeling for embodied 3D perception.

","absKo":"

Embodied 3D object detection은 embodied agent의 기본적인 perception 능력으로, 관측이 부분적이고, 심하게 occlusion되어 있으며, 순차적으로 주어지므로 temporal continuity를 모델링해야 한다. 그러나 기존 benchmark와 방법은 주로 완전히 재구성된 global scene을 위해 설계되어 first-person perception에서 temporal scene context와 instance evolution을 포착하지 못한다. 우리는 embodied setting에서 detection accuracy, temporal stability, consistency를 평가하는 egocentric 3D object detection을 위한 새로운 benchmark인 Embodied-Det를 소개한다. 이 benchmark를 바탕으로, 우리는 Scene-aware Feature AggregationInstance-aware Query Embedding이라는 상호 보완적인 두 temporal module을 통해 scene-level context와 instance-level continuity를 모델링하는 end-to-end temporal detection framework인 Embodied-DETR를 제안한다. Embodied-Det에서의 실험은 기존 방법들이 egocentric temporal setting에서 상당한 성능 저하를 겪는 반면, Embodied-DETR은 더 우수한 정확도와 temporal consistency를 달성하여 embodied 3D perception에서 temporal modeling의 효과를 입증함을 보여준다.

"},{"id":"64447","en":"Dual-Stream Diffusion for World-Model Augmented Vision-Language-Action Model","ko":"세계 모델 증강 Vision-Language-Action 모델을 위한 이중 스트림 Diffusion","authors":"John Won, Kyungmin Lee, Huiwon Jang, Dongyoung Kim, Jinwoo Shin","pos":"#906","link":"https://openreview.net/forum?id=NWApJP7UKT","abs":"

Augmenting Vision-Language-Action models (VLAs) with world models is promising for robotic policy learning but faces challenges in jointly predicting states and actions due to the modality gap. To address this, we propose DUal-STream diffusion (DUST), a world-model augmented VLA framework featuring a multimodal diffusion transformer that maintains separate modality streams while enabling cross-modal knowledge sharing. In addition, DUST utilizes independent noise perturbations and a decoupled flow matching loss to learn cross-modal causal relationships. We further introduce an asynchronous sampling method for action and vision tokens that enhances performance through inference-time scaling. Experimental results on simulated benchmarks like RoboCasa and GR-1 show that DUST achieves up to 6\\% gains over state-of-the-art VLA and world-modeling baselines, with inference-time scaling providing an additional 2–5\\% improvement. In real-world tasks using the Franka Research 3, DUST outperforms baselines by 10\\% in success rate. Finally, we demonstrate that DUST enables effective transfer learning through both pretraining on action-free videos and joint-training with heterogeneous robot and human datasets.

","absKo":"

Vision-Language-Action model (VLA)에 world model을 결합하는 것은 robotic policy learning에 유망하지만, modality gap 때문에 state와 action을 jointly prediction하는 데 어려움이 있다. 이를 해결하기 위해 우리는 DUal-STream diffusion (DUST)를 제안한다. 이는 cross-modal knowledge sharing을 가능하게 하면서도 separate modality stream을 유지하는 multimodal diffusion transformer를 특징으로 하는 world-model augmented VLA framework이다. 또한 DUST는 independent noise perturbation과 decoupled flow matching loss를 활용하여 cross-modal causal relationship을 학습한다. 우리는 inference-time scaling을 통해 성능을 향상시키는 action 및 vision token에 대한 asynchronous sampling method도 추가로 도입한다. RoboCasa와 GR-1 같은 simulated benchmark에서의 실험 결과, DUST는 state-of-the-art VLA 및 world-modeling baseline 대비 최대 6\\% 향상을 달성했으며, inference-time scaling은 추가로 2–5\\%의 개선을 제공했다. Franka Research 3를 사용한 실제 세계 작업에서는 DUST가 success rate에서 baseline을 10\\% 앞섰다. 마지막으로, DUST가 action-free video에 대한 pretraining과 heterogeneous robot 및 human dataset과의 joint-training을 통해 효과적인 transfer learning을 가능하게 함을 보였다.

"},{"id":"62902","en":"Discrete Diffusion VLA: Bringing Discrete Diffusion to Action Decoding in Vision-Language-Action Policies","ko":"Discrete Diffusion VLA: Vision-Language-Action 정책에서 행동 디코딩에 Discrete Diffusion 적용","authors":"Zhixuan Liang, Yizhuo Li, Tianshuo Yang, CHENGYUE WU, Sitong Mao, Liuao Pei, Tian Nian, Shunbo Zhou, Xiaokang Yang, Jiangmiao Pang, Yao Mu, Ping Luo","pos":"#908","link":"https://openreview.net/forum?id=c3BVcHcSiR","abs":"

Vision–Language–Action (VLA) models adapt large vision–language backbones to map images and instructions into robot actions. However, prevailing VLAs either generate actions autoregressively in a fixed left-to-right order or attach separate diffusion heads outside the backbone, fragmenting information pathways and hindering unified, scalable architectures. We present Discrete Diffusion VLA, a unified-transformer policy that models discretized action chunks with discrete diffusion retaining progressive refinement inside the VLM backbone. Our method achieves an adaptive decoding order that resolves high-confidence (easy) action elements before harder ones and employs secondary re-masking to revisit uncertain predictions, enabling robust error correction. This design preserves pretrained vision-language priors, supports parallel decoding, and improves the efficiency. Discrete Diffusion VLA achieves 96.5% avg.~success on LIBERO, 71.2% visual matching on SimplerEnv-Fractal, and 54.2% overall on SimplerEnv-Bridge. On out-of-distribution benchmarks, our method exhibits only 1.4% language degradation versus 8.0% for parallel decoding, and 21.0% vision degradation versus 29.0% for continuous diffusion, demonstrating well retention of pretrained vision-language capabilities. Visualization analysis confirms the learned decoding order adaptively prioritizes high-confidence tokens, validating our refinement strategy.

","absKo":"

Vision–Language–Action (VLA) model은 대규모 vision–language backbone을 적응시켜 image와 instruction을 robot action으로 매핑한다. 그러나 현재의 VLA는 고정된 left-to-right 순서로 autoregressive하게 action을 생성하거나, backbone 외부에 별도의 diffusion head를 붙여 information pathway를 분절시키고 통합적이며 확장 가능한 architecture를 저해한다. 우리는 VLM backbone 내부에서 점진적 refinement를 유지하는 discrete diffusion으로 discretized action chunk를 모델링하는 unified-transformer policy인 Discrete Diffusion VLA를 제시한다. 우리의 방법은 높은 confidence를 가진(쉬운) action element를 더 어려운 것보다 먼저 해소하는 adaptive decoding order를 달성하며, secondary re-masking을 통해 불확실한 prediction을 다시 검토하여 견고한 error correction을 가능하게 한다. 이 설계는 pretrained vision-language prior를 보존하고, parallel decoding을 지원하며, 효율성을 향상시킨다. Discrete Diffusion VLA는 LIBERO에서 평균 96.5% 성공률, SimplerEnv-Fractal에서 71.2% visual matching, SimplerEnv-Bridge에서 54.2% overall 성능을 달성한다. out-of-distribution benchmark에서는 우리의 방법이 parallel decoding 대비 language degradation은 1.4%에 그치는 반면 8.0%였고, continuous diffusion 대비 vision degradation은 21.0%로 29.0%보다 낮아, pretrained vision-language capability를 잘 보존함을 보여준다. 시각화 분석은 학습된 decoding order가 높은 confidence token을 적응적으로 우선시함을 확인하여, 우리의 refinement strategy를 검증한다.

"},{"id":"63653","en":"DecoVer: A Decompose-and-Verify Neuro-Symbolic Framework for Embodied Task Planning with BC+","ko":"DecoVer: BC+를 활용한 embodied task planning을 위한 분해-검증 neuro-symbolic 프레임워크","authors":"YiXiang Jiang, Binqian Xu, Xiangbo Shu","pos":"#910","link":"https://openreview.net/forum?id=Uxw8jLtzCY","abs":"

Despite their remarkable general capabilities, Large Language Models (LLMs) struggle with the precise grounding required for embodied task planning. To bridge this gap, neuro-symbolic approaches have emerged, leveraging action languages like BC+ for their formal expressiveness and reasoning flexibility. However, prior methods that naively couple LLMs with BC+ typically depend on one-shot program generation, which is brittle in dynamic environments and prone to sequential omission and causal inconsistency. To address these limitations, we propose DecoVer, a Decompose-and-Verify neuro-symbolic framework that systematically adapts BC+ to embodied task planning. Specifically, DecoVer employs a cascading decomposition strategy to partition complex knowledge into hierarchical subspaces and integrates a dual verification mechanism for syntactic and executable correctness. Extensive experiments demonstrate that DecoVer consistently outperforms LLM-based baselines across the majority of evaluation metrics, achieving a 12.9% success rate gain over the highly capable Gemini-3-Pro-Preview and a 60.9% improvement over GPT-5.1 on logically complex test cases.

","absKo":"

LLM은 뛰어난 일반 능력에도 불구하고 embodied task planning에 필요한 정밀한 grounding에서는 어려움을 겪는다. 이 격차를 메우기 위해, BC+ 같은 action language의 formal expressiveness와 reasoning flexibility를 활용하는 neuro-symbolic approach가 등장했다. 그러나 LLM과 BC+를 단순하게 결합하는 기존 방법은 대개 one-shot program generation에 의존하며, 이는 동적인 environment에서 취약하고 sequential omission 및 causal inconsistency에 취약하다. 이러한 한계를 해결하기 위해, 우리는 BC+를 embodied task planning에 체계적으로 적응시키는 Decompose-and-Verify neuro-symbolic framework인 DecoVer를 제안한다. 구체적으로 DecoVer는 복잡한 knowledge를 계층적 subspace로 분할하는 cascading decomposition strategy를 사용하고, syntactic correctness와 executable correctness를 위한 dual verification mechanism을 통합한다. 광범위한 실험은 DecoVer가 평가 metric의 대부분에서 LLM 기반 baseline을 일관되게 능가함을 보여주며, 매우 능력 있는 Gemini-3-Pro-Preview 대비 12.9%의 success rate 향상을, 논리적으로 복잡한 테스트 케이스에서는 GPT-5.1 대비 60.9%의 향상을 달성한다.

"},{"id":"66635","en":"Debate2Create: Robot Co-design via Multi-Agent LLM Debate","ko":"Debate2Create: 다중 에이전트 LLM 토론을 통한 로봇 공동 설계","authors":"Kevin Qiu, Marek Cygan","pos":"#911","link":"https://openreview.net/forum?id=1ufVo73uzD","abs":"

We introduce Debate2Create (D2C), a multi-agent LLM framework that formulates robot co-design as structured, iterative debate grounded in physics-based evaluation. A design agent and control agent engage in a thesis-antithesis-synthesis loop, while pluralistic LLM judges provide multi-objective feedback to steer exploration. Across five MuJoCo locomotion benchmarks, D2C achieves up to 3.2× the default Ant score and ~9× on Swimmer, outperforming prior LLM-based methods and black-box optimization. Iterative debate yields 18–35% gains over compute-matched zero-shot generation, and D2C-generated rewards transfer to default morphologies in 4/5 tasks. Our results demonstrate that structured multi-agent debate offers an effective alternative to hand-designed objectives for joint morphology-reward optimization.

","absKo":"

우리는 Debate2Create (D2C)를 소개한다. 이는 물리 기반 평가에 근거한 구조화된 반복적 debate로 robot co-design을 정식화하는 multi-agent LLM framework이다. design agent와 control agent는 thesis-antithesis-synthesis loop를 수행하며, pluralistic LLM judge는 탐색을 유도하는 multi-objective feedback을 제공한다. 다섯 개의 MuJoCo locomotion benchmark 전반에서 D2C는 기본 Ant 점수의 최대 3.2×, Swimmer에서는 약 9×를 달성하며, 기존 LLM 기반 방법과 black-box optimization을 능가한다. 반복적 debate는 compute를 맞춘 zero-shot generation 대비 18–35%의 향상을 제공하고, D2C가 생성한 reward는 4/5 task에서 기본 morphology로 전이된다. 우리의 결과는 구조화된 multi-agent debate가 joint morphology-reward optimization을 위한 hand-designed objective의 효과적인 대안임을 보여준다.

"},{"id":"60472","en":"Any3D-VLA: Enhancing VLA Robustness via Diverse Point Clouds","ko":"Any3D-VLA: 다양한 Point Cloud를 통한 VLA 강건성 향상","authors":"Xianzhe Fan, Shengliang Deng, Xiaoyang Wu, Yuxiang Lu, Zhuoling Li, Mi Yan, Yujia Zhang, Zhizheng Zhang, He Wang, Hengshuang Zhao","pos":"#912","link":"https://openreview.net/forum?id=zyMvoKYWMZ","abs":"

Existing Vision-Language-Action (VLA) models typically take 2D images as visual input, which limits their spatial understanding in complex scenes. How can we incorporate 3D information to enhance VLA capabilities? We conduct a pilot study across different observation spaces and visual representations. The results show that explicitly lifting visual input into point clouds yields representations that better complement their corresponding 2D representations. To address the challenges of (1) scarce 3D data and (2) the domain gap induced by cross-environment differences and depth-scale biases, we propose Any3D-VLA. It unifies the simulator, sensor, and model-estimated point clouds within a training pipeline, constructs diverse inputs, and learns domain-agnostic 3D representations that are fused with the corresponding 2D representations. Simulation and real-world experiments demonstrate Any3D-VLA's advantages in improving performance and mitigating the domain gap. Our code is available at https://anonymous.4open.science/r/Any3D-VLA-anonymous.

","absKo":"

기존 Vision-Language-Action (VLA) 모델은 일반적으로 2D 이미지를 시각 입력으로 사용하므로, 복잡한 장면에서의 공간적 이해가 제한된다. 3D 정보를 어떻게 통합하여 VLA 능력을 향상시킬 수 있을까? 우리는 서로 다른 observation space와 visual representation 전반에 걸쳐 pilot study를 수행했다. 그 결과, 시각 입력을 point cloud로 명시적으로 lifting하면 해당 2D representation을 더 잘 보완하는 representation을 얻을 수 있음을 보였다. (1) 3D 데이터의 희소성과 (2) cross-environment 차이와 depth-scale bias로 인해 유발되는 domain gap이라는 과제를 해결하기 위해, 우리는 Any3D-VLA를 제안한다. 이 방법은 training pipeline 내에서 simulator, sensor, model-estimated point cloud를 하나로 통합하고, 다양한 입력을 구성하며, 대응되는 2D representation과 fuse되는 domain-agnostic 3D representation을 학습한다. 시뮬레이션 및 실제 환경 실험은 Any3D-VLA가 성능 향상과 domain gap 완화에서 우수함을 보여준다. 코드는 https://anonymous.4open.science/r/Any3D-VLA-anonymous 에서 확인할 수 있다.

"},{"id":"62270","en":"A Generalist Pair-wise Progress Critic Model for Vision-Language-Action Robots","ko":"Vision-Language-Action 로봇을 위한 Generalist Pair-wise Progress Critic Model","authors":"Qi Zhang, shaopeng zhai, Shengzhe Zhang, Litao Liu, TianyiZhang, huang, Ming Zhou","pos":"#914","link":"https://openreview.net/forum?id=i7mfaYYLDf","abs":"

Recent advances in Vision-Language-Action (VLA) models have significantly improved robotic perception and manipulation capabilities, but still struggling to adapt in dynamic, open-ended real-world environments due to a lack of reliable task progress feedback and improvement mechanisms. To address these challenges, we propose a generalist Vision Language Action-Critic model, VLAC, which can integrate both human and robot data, and unify action policy and task progress critic within a single autoregressive architecture. Specifically, we propose a scalable and generalizable pair-wise progress understanding approach that can predict the delta of task progress between two steps in a trajectory and generate correct actions to complete the task. Then, we trained the model on large-scale, multi-source human, robot, and general vision-language data for a generalist. Furthermore, we deploy reinforcement learning where VLAC can autonomously evaluate task progress to provide intrinsic rewards. Extensive evaluations demonstrate that our model generalizes effectively across diverse tasks and environments, leveraging its pair-wise progress understanding to provide reliable dense rewards, robust action generation, and significant improvements in real-world reinforcement learning.

","absKo":"

최근 Vision-Language-Action (VLA) model의 발전은 로봇의 perception 및 manipulation 능력을 크게 향상시켰지만, 신뢰할 수 있는 task progress feedback과 개선 메커니즘의 부재로 인해 동적이고 open-ended한 실제 환경에 적응하는 데 여전히 어려움을 겪고 있다. 이러한 문제를 해결하기 위해, 우리는 인간과 로봇 데이터를 모두 통합하고 action policy와 task progress critic을 단일 autoregressive architecture 안에서 통합하는 generalist Vision Language Action-Critic model인 VLAC을 제안한다. 구체적으로, 우리는 trajectory의 두 step 사이에서 task progress의 delta를 예측하고 task 완료에 필요한 올바른 action을 생성할 수 있는 확장 가능하고 일반화 가능한 pair-wise progress understanding 접근법을 제안한다. 이후 우리는 generalist을 위해 대규모 multi-source human, robot, general vision-language data로 model을 학습시켰다. 더 나아가, VLAC이 task progress를 자율적으로 평가해 intrinsic reward를 제공할 수 있도록 reinforcement learning을 적용한다. 광범위한 평가 결과, 우리 model은 pair-wise progress understanding을 활용해 신뢰할 수 있는 dense reward, robust action generation, 그리고 실제 reinforcement learning에서의 유의미한 성능 향상을 제공하며, 다양한 task와 환경 전반에 효과적으로 일반화함을 보여준다.

"},{"id":"60629","en":"Towards Practical World Model-based Reinforcement Learning for Vision-Language-Action Models","ko":"Vision-Language-Action Model을 위한 실용적 World Model 기반 Reinforcement Learning을 향하여","authors":"Zhilong Zhang, Haoxiang Ren, Yihao Sun, Yifei Sheng, Haonan Wang, Zhichao Wu, Haoxin Lin, Pierre-Luc Bacon, Yang Yu","pos":"#207","link":"https://openreview.net/forum?id=yKQ8GrwEhr","abs":"

Vision-Language-Action (VLA) models show strong generalization for robotic control, but finetuning them with reinforcement learning (RL) is constrained by the high cost and safety risks of real-world interaction. Training VLA models in interactive world models avoids these issues but introduces several challenges, including pixel-level world modeling, multi-view consistency, and compounding errors under sparse rewards. Building on recent advances across large multimodal models and model-based RL, we propose VLA-MBPO, a practical framework to tackle these problems in VLA finetuning. Our approach has three key design choices: (i) adapting unified multimodal models (UMMs) for data-efficient world modeling; (ii) an interleaved view decoding mechanism to enforce multi-view consistency; and (iii) chunk-level branched rollout to mitigate error compounding. Theoretical analysis and experiments across simulation and real-world tasks demonstrate that VLA-MBPO significantly improves policy performance and sample efficiency. Crucially, our method maintains a universal set of hyperparameters across all tasks, underscoring its robustness and scalability for real-world robotic deployment.

","absKo":"

Vision-Language-Action (VLA) 모델은 로봇 제어에서 강한 generalization을 보이지만, reinforcement learning (RL)으로 이를 finetuning하는 일은 실제 환경 상호작용의 높은 비용과 안전 위험에 의해 제약된다. VLA 모델을 interactive world model에서 학습하면 이러한 문제를 피할 수 있지만, pixel-level world modeling, multi-view consistency, sparse reward 하에서의 누적 오차 등 여러 도전 과제가 발생한다. large multimodal model과 model-based RL의 최근 발전을 바탕으로, 우리는 VLA finetuning에서 이러한 문제를 해결하기 위한 실용적 프레임워크 VLA-MBPO를 제안한다. 우리의 접근법은 세 가지 핵심 설계 선택을 가진다: (i) data-efficient world modeling을 위해 unified multimodal models (UMMs)를 적응시키는 것, (ii) multi-view consistency를 강제하는 interleaved view decoding mechanism, 그리고 (iii) error compounding을 완화하는 chunk-level branched rollout. 이론적 분석과 시뮬레이션 및 실제 세계 task에 대한 실험은 VLA-MBPO가 policy 성능과 sample efficiency를 크게 향상시킴을 보여준다. 결정적으로, 우리의 방법은 모든 task에 걸쳐 동일한 hyperparameter 집합을 유지하며, 실제 로봇 배치를 위한 robustness와 scalability를 강조한다.

"},{"id":"62843","en":"EvoMAS: Heuristics in the Loop—Evolving Smarter Agentic Workflows","ko":"EvoMAS: 휴리스틱 루프—보다 똑똑한 에이전틱 워크플로우 진화","authors":"Yangbo Wei, Zhen Huang, Ronghao Xu, Hong Wang, WEI XING","pos":"#2409","link":"https://openreview.net/forum?id=cdWnpXkcQo","abs":"

The rapid development of Large Language Models has driven Multi-Agent Systems (MAS) growth, but constructing efficient MAS still requires labor-intensive manual design. Current automation methods often generate templated agents, rely on monolithic optimization, and ignore task complexity gradients. This paper presents Evolutionary MAS (EvoMAS), a biologically inspired framework that addresses these limitations through three interconnected dimensions: (1) dynamic and diverse evolutionary strategies with six biologically inspired operators (3 exploration, 3 exploitation) and adaptive strategy selection; (2) role-level evolution that dynamically optimizes agent specialization and collaboration patterns; and (3) curriculum-guided evolution that partitions tasks by difficulty and evolves sequentially from simple to complex under cross-stage stability constraints. To bridge the inefficiency of pure evolutionary search and the rigidity of manual design, we introduce the Cyber Creator, a meta-control system that combines dynamic rule formulation with reflective updates. Experiments show EvoMAS consistently outperforms existing methods across multiple domains while remaining cost-efficient, with agent roles evolving from homogeneous actors to specialized reasoning ensembles.

","absKo":"

Large Language Models의 빠른 발전은 Multi-Agent Systems (MAS)의 성장을 이끌었지만, 효율적인 MAS를 구축하는 일은 여전히 노동 집약적인 수동 설계를 필요로 한다. 기존 자동화 방법은 대개 템플릿화된 agent를 생성하고, 단일체적 최적화에 의존하며, task complexity gradient를 무시한다. 이 논문은 이러한 한계를 세 가지 상호 연결된 차원으로 해결하는 생물학적 영감의 framework인 Evolutionary MAS (EvoMAS)를 제안한다. (1) 여섯 개의 생물학적 영감 operator(탐색 3개, 활용 3개)와 적응적 strategy selection을 갖춘 동적이고 다양한 evolutionary strategy, (2) agent specialization과 collaboration pattern을 동적으로 최적화하는 role-level evolution, (3) 난이도에 따라 task를 분할하고 stage 간 안정성 제약 하에서 단순한 것에서 복잡한 것으로 순차적으로 진화시키는 curriculum-guided evolution이 그것이다. 순수한 evolutionary search의 비효율성과 수동 설계의 경직성을 연결하기 위해, 우리는 동적 규칙 형성과 reflective update를 결합한 meta-control system인 Cyber Creator를 도입한다. 실험 결과 EvoMAS는 여러 도메인에서 기존 방법들을 일관되게 능가하면서도 비용 효율성을 유지하며, agent role이 동질적 actor에서 특화된 reasoning ensemble로 진화함을 보여준다.

"},{"id":"61535","en":"AdaNav: Adaptive Reasoning with Uncertainty for Vision-Language Navigation","ko":"AdaNav: Vision-Language Navigation을 위한 불확실성 기반 적응형 추론","authors":"Xin Ding, Jianyu Wei, Yifan Yang, Shiqi Jiang, QIANXI ZHANG, Hao Wu, Fucheng Jia, Liang Mi, Yuxuan Yan, weijun wang, Yunxin Liu, Zhibo Chen, Ting Cao","pos":"#303","link":"https://openreview.net/forum?id=p8vZ7TuyKm","abs":"

Vision-Language Navigation (VLN) requires agents to follow natural language instructions by grounding them in sequential visual observations over long horizons. Explicit reasoning could enhance temporal consistency and perception–action alignment, but reasoning at fixed steps often leads to suboptimal performance and unnecessary computation. To address this, we propose AdaNav, an uncertainty-based adaptive reasoning framework for VLN. At its core is the Uncertainty-Adaptive Reasoning Block (UAR), a lightweight plugin that dynamically triggers reasoning. We introduce Action Entropy as a policy prior for UAR and progressively refine it through a Heuristics-to-RL training method, enabling agents to learn difficulty-aware reasoning policies under the strict data limitations of embodied tasks. Results show that with only 6K training samples, AdaNav achieves substantial gains over closed-source models trained on million-scale data, improving success rate by 20% on R2R val-unseen, 11.7% on RxR-CE, and 11.4% in real-world scenes.

","absKo":"

Vision-Language Navigation (VLN)은 agent가 긴 시간 범위에 걸친 순차적 시각 관찰에 자연어 지시를 grounding하여 따라가야 하는 과제를 요구한다. 명시적 reasoning은 temporal consistency와 perception–action alignment를 향상시킬 수 있지만, 고정된 시점에서의 reasoning은 종종 최적 이하의 성능과 불필요한 계산을 초래한다. 이를 해결하기 위해, 우리는 VLN을 위한 uncertainty 기반 adaptive reasoning framework인 AdaNav를 제안한다. 그 핵심에는 reasoning을 동적으로 트리거하는 가벼운 plugin인 Uncertainty-Adaptive Reasoning Block (UAR)이 있다. 우리는 UAR의 policy prior로 Action Entropy를 도입하고, Heuristics-to-RL training method를 통해 이를 점진적으로 정교화하여, embodied task의 엄격한 data 제한 하에서도 difficulty-aware reasoning policy를 학습할 수 있도록 한다. 결과는 6K training sample만으로도 AdaNav가 million-scale data로 학습된 closed-source model을 크게 능가하며, R2R val-unseen에서 success rate를 20%, RxR-CE에서 11.7%, 실제 환경 장면에서 11.4% 향상시킴을 보여준다.

"},{"id":"62253","en":"Speedup Patch: Learning a Plug-and-Play Policy to Accelerate Embodied Manipulation","ko":"Speedup Patch: Embodied Manipulation 가속을 위한 Plug-and-Play Policy 학습","authors":"Zhichao Wu, Junyin Ye, Zhilong Zhang, Yihao Sun, Haoxin Lin, Jiaheng Luo, Haoxiang Ren, lei yuan, Yang Yu","pos":"#311","link":"https://openreview.net/forum?id=iJNYi8dwr4","abs":"While current embodied policies exhibit remarkable manipulation skills, their execution remains unsatisfactorily slow as they inherit the tardy pacing of human demonstrations. Existing acceleration methods typically require policy retraining or costly online interactions, limiting their scalability for large-scale foundation models. In this paper, we propose **S**peed**u**p **P**atch (**SuP**), a lightweight, policy-agnostic framework that enables **plug-and-play acceleration** using solely offline data. SuP introduces an external scheduler that adaptively downsamples action chunks provided by embodied policies to eliminate redundancies. Specifically, we formalize the optimization of our scheduler as a Constrained Markov Decision Process (CMDP) aimed at maximizing efficiency without compromising task performance. Since direct success evaluation is infeasible in offline settings, SuP introduces **World Model based state deviation** as a surrogate metric to enforce safety constraints. By leveraging a learned world model as a virtual evaluator to predict counterfactual trajectories, the scheduler can be optimized via offline reinforcement learning. Empirical results on simulation benchmarks (Libero, Bigym) and real-world tasks validate that SuP achieves an overall $1.8\\times$ execution speedup for diverse policies while maintaining their original success rates.","absKo":"현재 embodied policy들은 뛰어난 조작 능력을 보이지만, 인간 시연의 느린 속도를 그대로 물려받아 실행이 만족스럽지 못하게 느리다. 기존 가속 방법은 대개 policy 재학습이나 비용이 큰 online interaction을 요구해, 대규모 foundation model에 대한 확장성을 제한한다. 본 논문에서는 오직 offline data만을 사용해 plug-and-play 가속을 가능하게 하는 경량 policy-agnostic 프레임워크 **S**peed**u**p **P**atch (**SuP**)를 제안한다. SuP는 embodied policy가 제공하는 action chunk를 적응적으로 downsample하여 중복을 제거하는 external scheduler를 도입한다. 구체적으로, 우리는 scheduler의 최적화를 task performance를 해치지 않으면서 효율을 최대화하는 Constrained Markov Decision Process (CMDP)로 정식화한다. offline setting에서는 직접적인 성공 평가가 불가능하므로, SuP는 안전 제약을 강제하기 위한 대리 지표로 **World Model based state deviation**을 도입한다. 학습된 world model을 반사실적 trajectory를 예측하는 virtual evaluator로 활용함으로써, scheduler는 offline reinforcement learning으로 최적화될 수 있다. simulation benchmark(Libero, Bigym)와 실제 task에서의 실험 결과는 SuP가 다양한 policy에 대해 원래의 성공률을 유지하면서 전체적으로 $1.8\\times$의 실행 속도 향상을 달성함을 검증한다."},{"id":"64310","en":"A Mechanistic Understanding of Sim-and-Real Co-Training in Generative Policies","ko":"생성적 정책에서 시뮬레이션-현실 공동 학습에 대한 기전적 이해","authors":"Yu Lei, Minghuan Liu, Abhiram Maddukuri, Zhenyu Jiang, Yuke Zhu","pos":"#3707","link":"https://openreview.net/forum?id=OtKDK3ZtLJ","abs":"

Co-training, which combines limited in-domain real-world data with abundant surrogate data such as simulation or cross-embodiment demonstrations, has been widely adopted for training generative visuomotor robot policies. Despite its empirical success, the mechanisms underlying when and why co-training works remain poorly understood. Starting from theoretical analysis and a toy example, we identify these two key intrinsic factors for end-to-end co-training systems: balanced mixing ratio\" andstructured representation alignment\". We propose an explanation that when simulation and real-world data are combined with a balanced mixing ratio, co-training naturally learns representations that are aligned across domains while remaining domain-distinguishable, enabling effective knowledge transfer without sacrificing real-world adaptation, which we refer to as structured representation alignment. We validate the hypothesis with comprehensive sim-and-sim and sim-and-real robotic experiments, showing that structured representation alignment reliably emerges under balanced mixing ratios and largely determines downstream performance. Benchmarking several recent co-training methods further supports this explanation. Guided by our analysis, we propose a simple combination of co-training techniques that jointly promote alignment and domain discernibility, achieving substantial improvements over prior approaches.

","absKo":"

Co-training은 제한된 도메인 내 실제 데이터와 simulation이나 cross-embodiment demonstration과 같은 풍부한 surrogate data를 결합하는 방식으로, generative visuomotor robot policy 학습에 널리 사용되어 왔다. 실증적으로는 성공적이지만, co-training이 언제 그리고 왜 효과적인지에 대한 메커니즘은 여전히 충분히 이해되지 않고 있다. 이론 분석과 toy example에서 출발하여, 우리는 end-to-end co-training system의 두 가지 핵심 내재 요인인 balanced mixing ratio\" andstructured representation alignment\"를 식별한다. 우리는 simulation 데이터와 real-world 데이터를 balanced mixing ratio로 결합하면, co-training이 자연스럽게 도메인 간에 정렬되면서도 도메인 구분성은 유지하는 representation을 학습하게 되어, real-world adaptation을 희생하지 않으면서 효과적인 지식 전이가 가능해진다는 설명을 제안한다. 이를 structured representation alignment이라고 부른다. 우리는 comprehensive sim-and-sim 및 sim-and-real robotic experiment로 이 가설을 검증했으며, balanced mixing ratio 하에서 structured representation alignment가 안정적으로 나타나고 downstream performance를 크게 결정함을 보였다. 여러 최근 co-training 방법에 대한 benchmark 또한 이 설명을 뒷받침한다. 우리의 분석에 따라, 우리는 alignment와 domain discernibility를 함께 촉진하는 co-training 기법의 간단한 조합을 제안하며, prior approach를 크게 상회하는 성능을 달성한다.

"},{"id":"64525","en":"SpecPrune-VLA: Accelerating Vision-Language-Action Models via Action-Aware Self-Speculative Pruning","ko":"SpecPrune-VLA: 행동 인식 자기 추측적 가지치기를 통한 Vision-Language-Action 모델 가속","authors":"Hanzhen Wang, Jiaming Xu, Yushun Xiang, Jiayi Pan, Yongkang Zhou, Yong-Lu Li, Guohao Dai","pos":"#3708","link":"https://openreview.net/forum?id=MjE62ZnpRH","abs":"Pruning is a typical acceleration technique for compute-bound models by removing computation on unimportant values. Recently, it has been applied to accelerate Vision-Language-Action (VLA) model inference. However, existing acceleration methods focus on local information from the current action step and ignore the global context, leading to $>$20\\% success rate drop and limited speedup in some scenarios. In this paper, we point out \\textbf{spatial-temporal consistency} in VLA tasks: input images in consecutive steps exhibit high similarity, and propose the key insight that token selection should combine local information with global context of the model. Based on this, we propose SpecPrune-VLA, a training-free, two-level pruning method with heuristic control. **(1) Action-level static pruning.** We leverage global history and local attention to statically reduce visual tokens per action. **(2) Layer-level dynamic pruning.** We prune tokens adaptively per layer based on layer-wise importance. **(3) Lightweight action-aware controller:** We classify actions as coarse- or fine-grained by the speed of the end effector and adjust pruning aggressiveness accordingly. Extensive experiments show that SpecPrune-VLA achieves up to 1.57× speedup in LIBERO simulation and 1.70× on real-world tasks, with negligible success rate degradation.","absKo":"프루닝은 중요하지 않은 값에 대한 연산을 제거함으로써 compute-bound model을 가속하는 전형적인 기법이다. 최근에는 이를 Vision-Language-Action (VLA) model inference 가속에 적용하고 있다. 그러나 기존 가속 방법은 현재 action step의 local information에만 초점을 맞추고 global context를 무시하여, 일부 상황에서 성공률이 $>$20\\% 감소하고 가속 효과도 제한적이다. 본 논문에서는 VLA task에서 \\textbf{spatial-temporal consistency}를 지적한다. 즉, 연속된 step의 입력 이미지들은 높은 유사성을 보이며, token selection이 local information과 model의 global context를 결합해야 한다는 핵심 통찰을 제안한다. 이를 바탕으로 우리는 heuristic control을 갖춘 training-free, two-level pruning 방법인 SpecPrune-VLA를 제안한다. **(1) Action-level static pruning.** 우리는 global history와 local attention을 활용하여 action마다 visual token을 정적으로 줄인다. **(2) Layer-level dynamic pruning.** 우리는 layer-wise importance에 기반해 layer별로 token을 적응적으로 pruning한다. **(3) Lightweight action-aware controller:** end effector의 속도에 따라 action을 coarse-grained 또는 fine-grained로 분류하고, 이에 맞게 pruning aggressiveness를 조정한다. 광범위한 실험은 SpecPrune-VLA가 LIBERO 시뮬레이션에서 최대 1.57×, 실제 task에서 1.70×의 speedup을 달성하면서도 성공률 저하는 거의 없음을 보여준다."},{"id":"66657","en":"Fisher-Preserving Guidance: Training-Free Manifold Constraints for Safe Diffusion Control","ko":"Fisher-Preserving Guidance: 안전한 Diffusion 제어를 위한 학습 불필요 매니폴드 제약","authors":"Hao Ren, Zetong Bi, Yiming Zeng, Le Zheng, Zhi Li, Zhaoliang Wan, Lu Qi, Hui Cheng","pos":"#3800","link":"https://openreview.net/forum?id=1fDAyaKHMv","abs":"

Diffusion models are effective for waypoint prediction in visual navigation, but standard sampling and test time guidance can produce unsafe or inefficient trajectories when updates drift off the training manifold. We propose Fisher Preserving Guidance with Outer Product Span Projection, a training-free inference method that avoids large Fisher drift associated with off-distribution actions while optimizing a task objective. Our method computes the Fisher-preserving update via a low-rank Jacobian factorization, requiring only a single backward pass per step and enabling real-time use. We further introduce Truncated Fisher Denoising Sensitivity as an uncertainty signal and use it for robust multi-sample action blending. Experiments on toy and realistic navigation benchmarks, including Maze2D with TSDF-based guidance, PushT with official Diffusion Policy weights, and visual navigation in simulation and on real robots, demonstrate consistent improvements in performance over strong diffusion-policy baselines without additional training.

","absKo":"

Diffusion model은 visual navigation에서 waypoint prediction에 효과적이지만, standard sampling과 test-time guidance는 update가 training manifold에서 벗어날 경우 안전하지 않거나 비효율적인 trajectory를 생성할 수 있다. 우리는 task objective를 최적화하면서 분포 밖 행동과 관련된 큰 Fisher drift를 피하는 training-free inference 방법인 Fisher Preserving Guidance with Outer Product Span Projection을 제안한다. 본 방법은 low-rank Jacobian factorization을 통해 Fisher-preserving update를 계산하며, step당 단 한 번의 backward pass만 필요로 하므로 실시간 사용이 가능하다. 또한 uncertainty signal로서 Truncated Fisher Denoising Sensitivity를 도입하고, 이를 robust multi-sample action blending에 사용한다. TSDF-based guidance를 사용하는 Maze2D, official Diffusion Policy weight를 사용하는 PushT, 시뮬레이션 및 실제 로봇에서의 visual navigation을 포함한 toy 및 realistic navigation benchmark 실험은 추가 학습 없이도 강력한 diffusion-policy baseline에 비해 성능이 일관되게 향상됨을 보여준다.

"},{"id":"65548","en":"Predicting What Matters: Robust Generalist Robot Policy Learning via Future Semantic Mask","ko":"중요한 것 예측하기: 미래 의미 마스크를 통한 강건한 범용 로봇 정책 학습","authors":"Yunfan Lou, Xiaowei Chi, Xiaojie Zhang, Zezhong Qian, Chengxuan Li, Rongyu Zhang, yaoxu lyu, Guoyu Song, Chuyao Fu, Haoxuan Xu, Pengwei Wang, Shanghang Zhang","pos":"#3811","link":"https://openreview.net/forum?id=CWerqtOXif","abs":"

World models derived from large-scale video generative pre-training have emerged as a promising paradigm for generalist robot policy learning. However, standard approaches often focus on high-fidelity RGB video prediction, but this can result in overfitting to irrelevant factors, such as dynamic backgrounds and illumination changes. These distractions reduce the model's ability to generalize, ultimately leading to unreliable and fragile control policies. To address this, we introduce the Mask World Model (MWM), that leverages video diffusion architectures to predict the evolution of semantic masks instead of pixels. This shift imposes a geometric information bottleneck, forcing the model to capture essential physical dynamics and contact relations while filtering out visual noise. We seamlessly integrate this mask dynamics backbone with a diffusion-based policy head to enable robust end-to-end control. Extensive evaluations demonstrate the superiority of MWM on the LIBERO and RLBench simulation benchmarks, which significantly outperforming the state-of-the-arts RGB-based world models. Furthermore, real-world experiments and robustness evaluation (via random token pruning) reveal that MWM exhibits superior generalization capabilities with robust resilience to texture information loss.

","absKo":"

대규모 video generative pre-training으로부터 도출된 world model은 generalist robot policy learning을 위한 유망한 패러다임으로 부상했다. 그러나 표준 접근법은 고충실도 RGB video prediction에 초점을 두는 경우가 많고, 이는 dynamic background와 illumination change 같은 관련 없는 요인에 대한 overfitting으로 이어질 수 있다. 이러한 교란은 모델의 generalization 능력을 저하시켜, 궁극적으로 불안정하고 취약한 control policy를 초래한다. 이를 해결하기 위해 우리는 pixel 대신 semantic mask의 변화를 예측하도록 video diffusion architecture를 활용하는 Mask World Model (MWM)을 소개한다. 이 전환은 geometric information bottleneck을 부과하여, 모델이 visual noise를 걸러내면서 필수적인 physical dynamics와 contact relation을 포착하도록 강제한다. 우리는 이 mask dynamics backbone을 diffusion 기반 policy head와 매끄럽게 통합하여 robust end-to-end control을 가능하게 한다. 광범위한 평가 결과, MWM은 LIBERO와 RLBench simulation benchmark에서 우수성을 입증했으며, state-of-the-art RGB-based world model을 크게 능가한다. 더 나아가 실제 환경 실험과 robustness 평가(random token pruning)를 통해, MWM이 texture information loss에 대해 강한 resilience를 보이며 우수한 generalization 능력을 갖고 있음을 보여준다.

"},{"id":"65198","en":"EvoCF: Multi-Agent Collaboration via Agentic Memory-Driven Evolutionary Counterfactual Planning","ko":"EvoCF: 에이전트 메모리 기반 진화적 반사실 계획을 통한 멀티 에이전트 협업","authors":"Haotian Chi, Zeyu Feng, Xingrui Yu, Linbo Luo, Yew Soon ONG, Ivor Tsang, Hechang Chen, Yi Chang, Haiyan Yin","pos":"#3814","link":"https://openreview.net/forum?id=FsSirPJy9U","abs":"

Planning collaboration strategies for multi-agent embodied systems remains a core challenge for LLM-based planners, which often fail to capture the physical and coordination constraints of realworld environments. To address this, we present EvoCF, an agentic memory-driven evolutionary counterfactual planning framework for discovering improved multi-agent collaboration strategies through counterfactual plan generation and evaluation. First, we propose a symbolic constraint inductor that induces reusable symbolic constraints from failures, forming an evolving rule library. Then, we propose an evolutionary counterfactual plan generator that systematically explores semantically consistent plan variants through rule-conditioned mutations, enabling robust collaboration strategies beyond short-sighted one-shot LLM plans. Finally, we design an agentic memory-grounded evaluator that ranks candidate plans using retrieval-augmented evidence, producing interpretable, constraint-aware selections. Across multi-agent embodied simulation benchmarks, EvoCF consistently discovers more robust and executable plans compared to baseline approaches. Our results demonstrate that grounding multi-agent planning in agentic memory and counterfactual reasoning significantly enhances both effectiveness and robustness.

","absKo":"

Multi-agent embodied systems에서 협업 전략을 계획하는 일은 여전히 핵심적인 도전 과제이며, LLM-based planner는 실제 세계 환경의 물리적 제약과 coordination 제약을 포착하지 못하는 경우가 많다. 이를 해결하기 위해 우리는 counterfactual plan generation과 evaluation을 통해 개선된 multi-agent collaboration strategy를 발견하는 agentic memory-driven evolutionary counterfactual planning framework인 EvoCF를 제안한다. 먼저, 실패로부터 재사용 가능한 symbolic constraints를 유도하여 진화하는 rule library를 형성하는 symbolic constraint inductor를 제안한다. 다음으로, rule-conditioned mutation을 통해 의미론적으로 일관된 plan variant를 체계적으로 탐색하여, 단기적 one-shot LLM plan을 넘어서는 강건한 collaboration strategy를 가능하게 하는 evolutionary counterfactual plan generator를 제안한다. 마지막으로, retrieval-augmented evidence를 사용해 candidate plan을 순위화하는 agentic memory-grounded evaluator를 설계하여, 해석 가능하고 constraint-aware한 선택을 생성한다. 다양한 multi-agent embodied simulation benchmark 전반에서 EvoCF는 baseline 접근법과 비교해 더 강건하고 실행 가능한 plan을 일관되게 발견한다. 우리의 결과는 agentic memory와 counterfactual reasoning에 multi-agent planning을 grounding하는 것이 effectiveness와 robustness를 모두 크게 향상시킨다는 점을 보여준다.

"},{"id":"61279","en":"General Covariant Action Modeling: Constructing Generalized Manifolds via Spatio-Temporal Decoupling","ko":"일반 공변 행동 모델링: 시공간 분리를 통한 일반화 다양체 구성","authors":"Huaihai Lyu, Chaofan Chen, Mingyu Cao, Yuheng Ji, Changsheng Xu","pos":"#4011","link":"https://openreview.net/forum?id=rso6Nr7Noz","abs":"

Achieving robust generalization from limited data is a central challenge in embodied intelligence. Prevailing methods fail by regressing absolute coordinates, which violates the principle of general covariance. Theoretically, this conflates the intrinsic task geometry with rigid execution patterns, binding policies to specific motion styles and fixed speeds. To resolve this, we propose the Generalized Action Manifold (GAM) framework that enforces general covariance through structural disentanglement. Specifically, GAM constructs the manifold by enforcing invariance across two orthogonal dimensions: (1) Temporal Invariance, utilizing an Arc-Length Parameterizer to orthogonalize the spatial path geometry from temporal dynamics, ensuring robustness to velocity variations; (2) Geometric Invariance, where a Schema-Affine-Factorization mechanism maps trajectories to canonical “world lines” in the Lie-algebraic tangent space. This distinguishes invariant topological schemas from affine modulations, ensuring spatial generalizability. By integrating GAM within a structured Vision-Language-Action (VLA) architecture, we expand sparse training data into a continuous, valid action manifold. Empirical results demonstrate that GAM enables superior transfer and robustness capabilities, significantly outperforming geometry-agnostic baselines.

","absKo":"

제한된 data로부터 robust한 generalization을 달성하는 것은 embodied intelligence의 핵심 과제이다. 기존 방법들은 절대 좌표를 회귀함으로써 실패하는데, 이는 general covariance의 원칙을 위반하기 때문이다. 이론적으로 이는 task의 고유한 geometry와 rigid한 execution pattern을 혼동하여, policy를 특정 motion style과 고정된 speed에 결박시킨다. 이를 해결하기 위해 우리는 구조적 disentanglement를 통해 general covariance를 강제하는 Generalized Action Manifold (GAM) framework를 제안한다. 구체적으로, GAM은 두 개의 직교하는 차원에 걸쳐 invariance를 강제함으로써 manifold를 구성한다: (1) Temporal Invariance는 Arc-Length Parameterizer를 활용하여 spatial path geometry와 temporal dynamics를 직교화함으로써 velocity 변화에 대한 robustness를 보장한다; (2) Geometric Invariance는 Schema-Affine-Factorization 메커니즘을 통해 trajectory를 Lie-algebraic tangent space 내의 canonical한 “world lines”로 매핑한다. 이는 불변의 topological schema와 affine modulation을 구분하여 spatial generalizability를 보장한다. 구조화된 Vision-Language-Action (VLA) architecture 안에 GAM을 통합함으로써, 우리는 희소한 training data를 연속적이고 유효한 action manifold로 확장한다. 실험 결과는 GAM이 geometry-agnostic baseline을 크게 능가하면서 우수한 transfer 및 robustness 능력을 가능하게 함을 보여준다.

"},{"id":"63561","en":"Pretrained Vision-Language-Action Models are Surprisingly Resistant to Forgetting in Continual Learning","ko":"사전학습된 Vision-Language-Action 모델은 연속 학습에서 망각에 놀라울 만큼 강인하다","authors":"Huihan Liu, Changyeon Kim, Bo Liu, Minghuan Liu, Yuke Zhu","pos":"#408","link":"https://openreview.net/forum?id=VzdSHEab4G","abs":"

Continual learning is a long-standing challenge in robot policy learning, where a policy must acquire new skills over time without catastrophically forgetting previously learned ones. While prior work has extensively studied continual learning in relatively small behavior cloning (BC) policy models trained from scratch, its behavior in modern large-scale pretrained Vision-Language-Action (VLA) models remains underexplored. In this work, we find that pretrained VLAs are remarkably resistant to forgetting compared with smaller policy models trained from scratch. Simple Experience Replay (ER) works surprisingly well on VLAs, sometimes achieving zero forgetting even with a small replay data size. Our analysis reveals that pretraining plays a critical role in downstream continual learning performance: large pretrained models mitigate forgetting with a small replay buffer size while maintaining strong forward learning capabilities. Furthermore, we find that VLAs can retain relevant knowledge from prior tasks despite performance degradation during learning new tasks. This knowledge retention enables rapid recovery of seemingly forgotten skills through finetuning. Together, these insights imply that large-scale pretraining fundamentally changes the dynamics of continual learning, enabling models to continually acquire new skills over time with simple replay.

","absKo":"Continual learning은 robot policy learning에서 오랫동안 풀리지 않은 과제로, policy가 이전에 학습한 skill을 catastrophic forgetting 없이 시간이 지나며 새로운 skill을 습득해야 한다. 기존 연구는 비교적 작은 behavior cloning (BC) policy model을 scratch부터 학습하는 상황에서 continual learning을 광범위하게 연구해 왔지만, 현대의 대규모 pretrained Vision-Language-Action (VLA) model에서의 동작은 아직 충분히 탐구되지 않았다. 본 연구에서는 pretrained VLA가 scratch부터 학습한 더 작은 policy model에 비해 forgetting에 현저히 강하다는 것을 발견했다. Simple Experience Replay (ER)는 VLA에서 놀라울 정도로 잘 작동하며, 때로는 작은 replay data size에서도 zero forgetting을 달성한다. 우리의 분석은 pretraining이 downstream continual learning 성능에서 결정적 역할을 한다는 점을 보여준다. 즉, 대규모 pretrained model은 강한 forward learning capability를 유지하면서 작은 replay buffer size로 forgetting을 완화한다. 또한 우리는 VLA가 새로운 task를 학습하는 동안 성능이 저하되더라도 이전 task의 관련 지식을 보존할 수 있음을 발견했다. 이러한 지식 보존은 finetuning을 통해 잊힌 것처럼 보이는 skill을 빠르게 복구할 수 있게 한다. 종합하면, 이러한 통찰은 대규모 pretraining이 continual learning의 dynamics를 근본적으로 바꾸며, 단순한 replay만으로도 model이 시간이 지나며 새로운 skill을 지속적으로 습득할 수 있게 함을 시사한다.

"},{"id":"62938","en":"From Noise to Control: Parameterized Diffusion Policies","ko":"노이즈에서 제어로: Parameterized Diffusion Policies","authors":"Renhao Zhang, Haotian Fu, Mingxi Jia, George Konidaris, Yilun Du, Bruno C. da Silva","pos":"#411","link":"https://openreview.net/forum?id=bjoDtIUSvx","abs":"

We propose Parameterized Diffusion Policy (PDP), a framework that learns a diffusion policy parameterized in a smooth continuous space. By structuring a latent manifold such that distances between latents' values reflect the semantic similarity of physical trajectories, we transform diffusion from a mechanism of stochastic diversity into a precise tool for behavior steering. Our approach also enables smooth interpolation between known strategies and efficient generalization to novel constraints without the need to update policy weights. We demonstrate that PDP significantly improves adaptation performance on complex multimodal benchmarks in both simulation and real-robot hardware compared to regular diffusion policy, particularly in scenarios requiring the discovery of novel behaviors.

","absKo":"

우리는 Parameterized Diffusion Policy(PDP)를 제안한다. 이는 매끄러운 연속 공간에서 parameterized된 diffusion policy를 학습하는 프레임워크이다. latent manifold를 구성하여 latent value 간 거리가 물리적 trajectory의 semantic similarity를 반영하도록 함으로써, 우리는 diffusion을 stochastic diversity를 위한 메커니즘에서 행동을 정밀하게 조정하는 도구로 변환한다. 이 접근법은 또한 알려진 전략들 사이의 매끄러운 interpolation과 policy weight를 갱신할 필요 없는 새로운 제약에 대한 효율적인 generalization을 가능하게 한다. 우리는 PDP가 regular diffusion policy와 비교해 simulation과 실제 robot hardware 모두에서 복잡한 multimodal benchmark에 대한 적응 성능을 크게 향상시키며, 특히 새로운 행동의 발견이 필요한 상황에서 두드러짐을 보인다.

"},{"id":"63603","en":"When Replanning Becomes the Bottleneck: Budgeted Replanning for Embodied Agents","ko":"체계적 에이전트에서 재계획이 병목이 될 때: 예산 기반 재계획","authors":"Shuaijun Liu, Feiyang You, Xingwei Chen, Ningxin Su","pos":"#600","link":"https://openreview.net/forum?id=VZAd3E3D6w","abs":"

Embodied agents replan frequently to recover from execution drift, partial observability, and coordination hazards. In many LLM-based planners, each replanning call consumes an accumulated textual context that grows over time and across agents (history, failures, summaries, and messages). Once this context becomes large, replanning latency develops heavy tails and can miss real-time deadlines even when task success remains high---a failure mode that is hard to detect from average latency or success alone. We present BRACE, a controller that formulates replanning for embodied agents as a budgeted control loop. At each replanning trigger, BRACE decides whether to replan, selects a replanning mode, and allocates an explicit token budget together with a latency service-level objective (SLO), while accounting for the overhead of optional efficiency modules. As a reusable component, we introduce E-RECAP, a cost-aware progressive token pruning method that predicts token utility and prunes replanning contexts across transformer layers while preserving critical head and tail tokens. On Habitat-Lab navigation with growing multi-agent context, E-RECAP reduces tokens per replanning call by 71-76% and end-to-end replanning latency by 2.1-2.6x with minimal impact on success or SPL. In Meta Habitat, BRACE combined with E-RECAP reduces SLO violation rates from 85.5% to 4.7% without degrading task success. Results across three embodied platforms demonstrate that tail-aware, per-call budgeting is an effective and practical design principle for replanning systems.

","absKo":"

Embodied agent는 execution drift, partial observability, coordination hazard에서 회복하기 위해 자주 replanning을 수행한다. 많은 LLM-based planner에서 각 replanning 호출은 시간이 지남에 따라, 그리고 agent 간에 누적되는 텍스트 context(history, failure, summary, message)를 소비한다. 이 context가 충분히 커지면, task success는 여전히 높더라도 replanning latency에 heavy tail이 생기며 real-time deadline을 놓칠 수 있다. 이는 평균 latency나 성공률만으로는 발견하기 어려운 실패 모드이다. 우리는 embodied agent의 replanning을 budgeted control loop로 정식화하는 controller인 BRACE를 제시한다. 각 replanning trigger에서 BRACE는 replanning 여부를 결정하고, replanning mode를 선택하며, 명시적인 token budget과 latency service-level objective (SLO)를 할당하는 동시에, 선택적 efficiency module의 overhead를 고려한다. 재사용 가능한 component로서, 우리는 token utility를 예측하고 critical head token과 tail token을 보존하면서 transformer layer 전반에서 replanning context를 pruning하는 cost-aware progressive token pruning method인 E-RECAP를 도입한다. 점점 커지는 multi-agent context를 가진 Habitat-Lab navigation에서 E-RECAP은 replanning 호출당 token 수를 71-76% 줄이고 end-to-end replanning latency를 2.1-2.6배 감소시키면서 success나 SPL에는 최소한의 영향을 준다. Meta Habitat에서는 BRACE와 E-RECAP을 결합할 경우 task success를 저하시키지 않으면서 SLO violation rate를 85.5%에서 4.7%로 줄인다. 세 개의 embodied platform에 걸친 결과는 tail-aware per-call budgeting이 replanning system의 효과적이고 실용적인 설계 원칙임을 보여 준다.

"},{"id":"64465","en":"VLANeXt: Recipes for Building Strong VLA Models","ko":"VLANeXt: 강력한 VLA 모델 구축을 위한 레시피","authors":"Xiao-Ming Wu, Bin Fan, Kang Liao, Jian-Jian Jiang, Runze Yang, Yihang Luo, Zhonghua Wu, Wei-Shi Zheng, Chen Change Loy","pos":"#702","link":"https://openreview.net/forum?id=NKpnzoFqFu","abs":"

Following the rise of large foundation models, Vision–Language–Action models (VLAs) emerged, leveraging strong visual and language understanding for general-purpose policy learning. Yet, the current VLA landscape remains fragmented and exploratory. Although many groups have proposed their own VLA models, inconsistencies in training protocols and evaluation settings make it difficult to identify which design choices truly matter. To bring structure to this evolving space, we reexamine the VLA design space under a unified framework and evaluation setup. Starting from a simple VLA baseline similar to RT-2 and OpenVLA, we systematically dissect design choices along three dimensions: foundational components, perception essentials, and action modelling perspectives. From this study, we distill 12 key findings that together form a practical recipe for building strong VLA models. The outcome of this exploration is a simple yet effective model, VLANeXt. VLANeXt outperforms prior state-of-the-art methods on the LIBERO and LIBERO-plus benchmarks and demonstrates strong generalization in real-world experiments. We will release a unified, easy-to-use codebase that serves as a common platform for the community to reproduce our findings, explore the design space, and build new VLA variants on top of a shared foundation.

","absKo":"

large foundation model의 부상 이후, Vision–Language–Action model(VLA)이 등장하여 강력한 visual 및 language understanding을 general-purpose policy learning에 활용하고 있다. 그러나 현재 VLA landscape는 여전히 파편화되어 있고 탐색 단계에 머물러 있다. 많은 그룹이 각자의 VLA model을 제안했지만, training protocol과 evaluation setting의 불일치 때문에 어떤 design choice가 진정으로 중요한지 파악하기 어렵다. 이 진화하는 공간에 구조를 부여하기 위해, 우리는 unified framework와 evaluation setup 아래에서 VLA design space를 재검토한다. RT-2와 OpenVLA와 유사한 단순한 VLA baseline에서 출발하여, foundational component, perception essential, action modelling perspective라는 세 가지 차원에 걸쳐 design choice를 체계적으로 분해한다. 이 연구를 통해 우리는 강력한 VLA model을 구축하기 위한 실용적 recipe를 이루는 12개의 핵심 발견을 도출한다. 이 탐구의 결과물은 단순하면서도 효과적인 model인 VLANeXt이다. VLANeXt는 LIBERO 및 LIBERO-plus benchmark에서 이전 state-of-the-art method를 능가하고, 실제 환경 실험에서도 강한 generalization을 보여준다. 우리는 커뮤니티가 우리의 발견을 재현하고, design space를 탐색하며, shared foundation 위에 새로운 VLA variant를 구축할 수 있는 공통 platform 역할을 하는, 통합적이고 사용하기 쉬운 codebase를 공개할 예정이다.

"},{"id":"61157","en":"VLA-ATTC: Adaptive Test-Time Compute for VLA Models with Relative Action Critic Model","ko":"VLA-ATTC: Relative Action Critic Model을 활용한 VLA Model의 Adaptive Test-Time Compute","authors":"Wenhao Li, Xiu Su, Yichao Cao, Hongyan Xu, Xiaobo Xia, Shan You, Yi Chen, Chang Xu","pos":"#703","link":"https://openreview.net/forum?id=tAqHcMkhy7","abs":"

Vision-Language-Action (VLA) models have demonstrated remarkable capabilities and generalization in embodied manipulation. However, their decision-making relies on a fast, instinctive process that lacks deliberation. This strategy often leads to suboptimal or catastrophic actions when facing complex or ambiguous scenarios that require greater consideration. In this paper, we introduce \\textbf{VLA-ATTC}, a framework that endows VLA models with adaptive test-time compute (TTC). VLA-ATTC employs an uncertainty-based ``cognitive clutch'' to dynamically transition from reflexive execution to a TTC deliberation phase when necessary. During TTC phase, a novel \\textbf{Relative Action Critic} (RAC) model identifies the optimal action from generated candidates via pairwise comparisons. This relative mechanism replaces unstable absolute value estimation, significantly simplifying the learning objective. Furthermore, we introduce an efficient sampling strategy to amortize computational costs and an automated data pipeline that curates preference pairs without manual annotation. On the LIBERO-LONG benchmark, VLA-ATTC reduces the failure rate of the SOTA model PI0.5 by over 50\\%.

","absKo":"

Vision-Language-Action (VLA) 모델은 embodied manipulation에서 놀라운 성능과 generalization을 보여왔다. 그러나 이들의 의사결정은 숙고가 부족한 빠르고 직관적인 과정에 의존한다. 이러한 전략은 더 큰 고려가 필요한 복잡하거나 모호한 상황에 직면했을 때 종종 최적 이하의 행동이나 치명적인 행동으로 이어진다. 본 논문에서는 VLA 모델에 adaptive test-time compute (TTC)를 부여하는 \\textbf{VLA-ATTC} 프레임워크를 제안한다. VLA-ATTC는 uncertainty 기반의 ``cognitive clutch''를 사용해 필요할 때 반사적 실행에서 TTC 숙고 단계로 동적으로 전환한다. TTC 단계에서는 새로운 \\textbf{Relative Action Critic} (RAC) 모델이 pairwise comparison을 통해 생성된 후보들 중 최적의 행동을 식별한다. 이 relative mechanism은 불안정한 absolute value estimation을 대체하여 학습 목표를 크게 단순화한다. 또한 계산 비용을 amortize하기 위한 효율적인 sampling 전략과, 수작업 annotation 없이 preference pair를 선별하는 자동화된 data pipeline을 도입한다. LIBERO-LONG benchmark에서 VLA-ATTC는 SOTA 모델 PI0.5의 failure rate를 50\\% 이상 감소시킨다.

"},{"id":"66262","en":"UAV$^2$: A Unified and Adaptive Scheduling Framework for UAV Autopilot System with Reinforcement Learning","ko":"UAV$^2$: 강화 학습을 이용한 UAV 자율비행 시스템을 위한 통합 적응형 스케줄링 프레임워크","authors":"Zeying Li, shuai zhao, Chaowen Wu, Boyang Li, Kai Huang","pos":"#704","link":"https://openreview.net/forum?id=5JkuqtBpNB","abs":"Unmanned aerial vehicle (UAV) autopilot systems typically comprise navigation and flight-control modules, and their effective scheduling is critical to achieving high flight performance. However, most existing UAV platforms adopt a split architecture in which navigation and flight control are deployed on separate hardware devices. This separation restricts system-wide observability and prevents holistic scheduling and optimization across the entire autopilot pipeline. Moreover, autonomous flight performance emerges from implicit, cross-coupled, and accumulated interactions among multiple factors, rendering traditional model-based or heuristic scheduling approaches ineffective. To address these challenges, we propose UAV$^2$, a unified and adaptive scheduling framework for UAV autopilot systems with reinforcement learning, targeting flight performance optimization. UAV$^2$ integrates navigation and flight control onto a single onboard computing platform and operating system, formulates the scheduling problem as a partially observable Markov decision process, and learns scheduling policies from runtime execution feedback. The proposed approach is trained and evaluated in a hardware-in-the-loop simulation environment. Experimental results demonstrate that the learned scheduling policy consistently outperforms fixed-rate scheduling strategies in terms of flight robustness and tracking performance.","absKo":"무인항공기(UAV) autopilot 시스템은 일반적으로 navigation과 flight-control 모듈로 구성되며, 이들의 효과적인 scheduling은 높은 비행 성능을 달성하는 데 매우 중요하다. 그러나 기존 UAV 플랫폼의 대부분은 navigation과 flight control이 서로 분리된 hardware device에 배치되는 split architecture를 채택한다. 이러한 분리는 system-wide observability를 제한하고, autopilot pipeline 전체에 걸친 holistic scheduling 및 optimization을 가로막는다. 또한 autonomous flight performance는 여러 요인 간의 implicit하고 cross-coupled되며 누적되는 상호작용에서 나타나므로, 전통적인 model-based 또는 heuristic scheduling 접근법은 효과적이지 않다. 이러한 문제를 해결하기 위해, 우리는 flight performance optimization을 목표로 reinforcement learning을 활용한 UAV autopilot system용 통합적이고 adaptive한 scheduling framework인 UAV$^2$를 제안한다. UAV$^2$는 navigation과 flight control을 단일 onboard computing platform 및 operating system 위에 통합하고, scheduling 문제를 partially observable Markov decision process로 정식화하며, runtime execution feedback으로부터 scheduling policy를 학습한다. 제안하는 접근법은 hardware-in-the-loop simulation 환경에서 학습 및 평가된다. 실험 결과, 학습된 scheduling policy는 flight robustness와 tracking performance 측면에서 고정 주기 scheduling 전략보다 일관되게 우수함을 보인다."},{"id":"61138","en":"Sample from What You See: Visuomotor Policy Learning via Diffusion Bridge with Observation-Embedded Stochastic Differential Equation","ko":"보이는 것에서 샘플링하기: Observation-Embedded Stochastic Differential Equation을 사용한 Diffusion Bridge 기반 Visuomotor Policy Learning","authors":"Zhaoyang Liu, Mokai Pan, Zhongyi Wang, Kaizhen Zhu, Haotao Lu, Haipeng Zhang, Jingya Wang, Ye Shi","pos":"#710","link":"https://openreview.net/forum?id=tL2zkbmg8t","abs":"

Imitation learning with diffusion models has advanced robotic control by capturing the multi-modal action distributions. However, existing methods typically treat observations only as high-level conditions to the denoising network, rather than integrating them into the stochastic dynamics of the diffusion process itself. As a result, the sampling is forced to begin from random noise, weakening the coupling between perception and control and often yielding suboptimal performance. We propose BridgePolicy, a generative visuomotor policy that directly integrates observations into the stochastic dynamics via a diffusion-bridge formulation. By constructing an observation-informed trajectory, BridgePolicy enables sampling to start from a rich and informative prior rather than random noise, substantially improving precision and reliability in control. A key difficulty is that diffusion bridge normally connects distributions of matched dimensionality, while robotic observations are heterogeneous and not naturally aligned with actions. To overcome this, we introduce a multi-modal fusion module and a semantic aligner to unify the visual and state inputs and align the observations with action representations, making diffusion bridge applicable to heterogeneous robot data. Extensive experiments across 52 simulation tasks on three benchmarks and 5 real-world tasks demonstrate that BridgePolicy consistently outperforms state-of-the-art generative policies.

","absKo":"

Diffusion model을 활용한 imitation learning은 multi-modal action distribution을 포착함으로써 robotic control을 발전시켜 왔습니다. 그러나 기존 방법은 보통 observation을 diffusion process 자체의 stochastic dynamics에 통합하기보다는, denoising network의 high-level condition으로만 취급합니다. 그 결과 sampling이 무작위 noise에서 시작하도록 강제되어 perception과 control의 결합이 약해지고, 종종 최적 이하의 성능을 보이게 됩니다. 우리는 observation을 diffusion-bridge formulation을 통해 stochastic dynamics에 직접 통합하는 generative visuomotor policy BridgePolicy를 제안합니다. observation-informed trajectory를 구성함으로써 BridgePolicy는 sampling이 random noise가 아니라 풍부하고 정보성 높은 prior에서 시작되도록 하며, control의 정확도와 신뢰성을 크게 향상시킵니다. 핵심 난점은 diffusion bridge가 일반적으로 dimensionality가 맞는 distribution을 연결하는 반면, robotic observation은 이질적이며 action과 자연스럽게 정렬되어 있지 않다는 점입니다. 이를 극복하기 위해, 우리는 multi-modal fusion module과 semantic aligner를 도입하여 visual input과 state input을 통합하고 observation을 action representation과 정렬함으로써, diffusion bridge를 이질적인 robot data에 적용 가능하게 만듭니다. 세 개의 benchmark에서 52개의 simulation task와 5개의 real-world task에 걸친 광범위한 실험은 BridgePolicy가 일관되게 state-of-the-art generative policy를 능가함을 보여줍니다.

"},{"id":"64906","en":"SPUR: Scale-Partitioned Uncertainty Rectification for Robust UAV-on-UAV Interception","ko":"SPUR: 강건한 UAV-on-UAV 요격을 위한 Scale-Partitioned Uncertainty Rectification","authors":"Chenqi Yan, Zhaoyu Zeng, Yifeng Yang, Jundong Zhou, Zhuoyuan Ni, Junqi Wu, Qinying Gu, Xinbing Wang, Nanyang Ye","pos":"#712","link":"https://openreview.net/forum?id=IjQsbCrkMo","abs":"

Robust aerial target detection for autonomous UAV-on-UAV pursuit is severely hindered by continuous scale drift, long-tailed scale imbalance, and flight-induced visual noise, rendering standard empirical risk minimization strategies poorly aligned with real-world deployment. To address these challenges, we propose a scale-aware robust optimization framework that performs group-wise minimax optimization over scale-partitioned data, ensuring balanced robustness across long-, mid-, and close-range engagement regimes. We further introduce an uncertainty-rectified regression loss to suppress noise-driven errors without discarding informative hard examples, complemented by a control-aligned center accuracy penalty that prioritizes the localization precision required for stable flight control. Extensive experiments demonstrate that our method yields substantially improved robustness under visual degradation, with significantly slower decay in detection mAP and center-point accuracy compared to baselines. Validated through both photorealistic simulations and real-world flight tests, our system achieves real-time performance of 120 FPS on an embedded NVIDIA Orin NX platform, confirming its practical efficacy for high-speed interception.

","absKo":"

자율 UAV-on-UAV 추격을 위한 강건한 공중 표적 탐지는 지속적인 scale drift, long-tailed scale imbalance, 그리고 비행으로 유발되는 시각적 잡음 때문에 심각한 제약을 받으며, 표준 empirical risk minimization 전략은 실제 배포 환경과 잘 맞지 않는다. 이러한 문제를 해결하기 위해, 우리는 scale partitioned data에 대해 group-wise minimax optimization을 수행하는 scale-aware robust optimization framework를 제안하여, 장거리, 중거리, 근거리 교전 체제 전반에 걸쳐 균형 잡힌 강건성을 보장한다. 또한 잡음에 의해 유도된 오류를 유익한 hard example을 버리지 않고 억제하는 uncertainty-rectified regression loss를 도입하고, 안정적인 비행 제어에 필요한 위치 정밀도를 우선시하는 control-aligned center accuracy penalty를 결합한다. 광범위한 실험은 우리의 방법이 시각적 열화 조건에서 상당히 향상된 강건성을 보이며, baseline 대비 detection mAP과 center-point accuracy의 감소가 훨씬 더 느리게 나타남을 보여준다. photorealistic simulation과 실제 비행 시험 모두에서 검증된 본 시스템은 embedded NVIDIA Orin NX 플랫폼에서 real-time performance of 120 FPS를 달성하여, 고속 요격을 위한 실용적 효용을 확인한다.

"},{"id":"66289","en":"Focus-Then-Contact: Speeding Up Robotic Contact-Rich Task Learning with Affordance-Guided Real-World Residual Reinforcement Learning","ko":"집중 후 접촉: Affordance 유도 실세계 잔차 강화 학습을 통한 접촉 집약적 로봇 작업 학습 가속","authors":"Guanren Qiao, Ruixiang Ouyang, Sheng Xu, Ruixing Jin, Yueci Deng, Yunxin Tai, Kui Jia, Guiliang Liu","pos":"#801","link":"https://openreview.net/forum?id=536wLPu3Xp","abs":"

Real-World Reinforcement Learning (RL) has shown significant potential in robotic manipulation tasks. However, many methods still require substantial human-in-the-loop involvement to complete contact-rich tasks, especially when there are disruptions such as visual backgrounds or positional changes. To address this, we propose the Focus Then Contact (FTC), a lightweight and low-cost method to accelerate the convergence of human-in-the-loop real-world RL for contact-rich tasks. FTC leverages residual RL to provide base actions, helping the system quickly reach the target regions and improve sample efficiency. Additionally, FTC integrates an affordance-guided reward that drives the real-world RL system to quickly focus on key regions of interest, making it possible for the robotic arm to continuously engage with these goal areas through force-control feedback. At the same time, we optimize the human-in-the-loop implementation to prevent conflicts with RL over control of the robotic arm. We demonstrate the effectiveness of FTC on 6 contact-rich tasks, where it outperforms baseline methods in achieving high success rates and speeds up robotic contact-rich task learning under a real-world RL setting. Video materials can be seen in \\url{https://anonymous.4open.science/api/repo/FTC-website-BB5E/file/index.html}.

","absKo":"

Real-World Reinforcement Learning (RL)은 robotic manipulation task에서 상당한 잠재력을 보여왔다. 그러나 많은 방법은 여전히 contact-rich task를 완료하기 위해 상당한 human-in-the-loop 개입을 요구하며, 특히 visual background 변화나 위치 변화와 같은 방해가 있을 때 그러하다. 이를 해결하기 위해, 우리는 contact-rich task를 위한 human-in-the-loop real-world RL의 수렴을 가속하는 경량·저비용 방법인 Focus Then Contact (FTC)를 제안한다. FTC는 residual RL을 활용해 base action을 제공함으로써 시스템이 target region에 빠르게 도달하도록 돕고 sample efficiency를 향상시킨다. 또한 FTC는 affordance-guided reward를 통합하여 real-world RL 시스템이 핵심 관심 영역에 빠르게 집중하도록 유도하며, force-control feedback을 통해 robotic arm이 이러한 goal area와 지속적으로 상호작용할 수 있게 한다. 동시에, 우리는 human-in-the-loop 구현을 최적화하여 robotic arm 제어에 대한 RL과의 충돌을 방지한다. 우리는 6개의 contact-rich task에서 FTC의 효과를 입증하며, baseline method보다 높은 success rate를 달성하고 real-world RL 설정에서 robotic contact-rich task 학습을 가속함을 보인다. Video material은 \\url{https://anonymous.4open.science/api/repo/FTC-website-BB5E/file/index.html}에서 볼 수 있다.

"},{"id":"66596","en":"From Abstraction to Instantiation: Learning Behavioral Representation for Vision-Language-Action Model","ko":"추상화에서 구체화로: Vision-Language-Action 모델을 위한 행동 표현 학습","authors":"Bing Hu, Zaijing Li, Rui Shao, Junda Chen, April Hua Liu, Wei-Shi Zheng, Liqiang Nie","pos":"#802","link":"https://openreview.net/forum?id=2NRrsz4nUB","abs":"

Vision-Language-Action (VLA) models often suffer from performance degradation under distribution shifts, as they struggle to learn generalized behavior representations across varying environments. While existing approaches attempt to construct behavior representations through action-centric latent variables, they are often limited by short-horizon temporal fragmentation and static execution-alignment, leading to inconsistent behaviors in complex scenarios. To address these limitations, we propose \\textbf{BehaviorVLA}, a framework that facilitates robust manipulation through the learning of a temporally coherent behavioral representations. Our approach features two symmetric components: (1) the \\textbf{Visuomotor Behavior Encoder (VBE)}, which utilizes a causal Mamba-based architecture to aggregate long-horizon trajectory information into a unified behavior representation; and (2) the \\textbf{Phase-conditioned Behavior Decoder (PBD)}, which decodes this representation into precise actions by dynamically aligning task-level priors with real-time execution progress. Experiments on RoboTwin 2.0, LIBERO, and CALVIN demonstrate state-of-the-art success rates of 58\\%, 98\\%, and 4.36 (Avg. Len), respectively. Notably, in real-world sim-to-real transfer, BehaviorVLA matches the performance of OpenVLA-OFT using only 50\\% of the demonstration data, showcasing its superior data efficiency and generalization.

","absKo":"Vision-Language-Action(VLA) model은 다양한 환경 전반에 걸쳐 일반화된 behavior representation을 학습하는 데 어려움을 겪기 때문에 distribution shift 하에서 성능 저하를 자주 경험한다. 기존 접근법은 action-centric latent variable을 통해 behavior representation을 구성하려 하지만, short-horizon temporal fragmentation과 static execution-alignment의 제약을 받는 경우가 많아 복잡한 시나리오에서 일관성 없는 behavior를 초래한다. 이러한 한계를 해결하기 위해 우리는 시간적으로 일관된 behavioral representation 학습을 통해 강건한 조작을 가능하게 하는 프레임워크인 \\textbf{BehaviorVLA}를 제안한다. 우리의 접근은 두 개의 대칭적 구성요소를 특징으로 한다. (1) \\textbf{Visuomotor Behavior Encoder(VBE)}는 causal Mamba 기반 architecture를 활용하여 긴 horizon의 trajectory 정보를 단일 behavior representation으로 집계하고, (2) \\textbf{Phase-conditioned Behavior Decoder(PBD)}는 task-level prior를 실시간 실행 진행도와 동적으로 정렬하여 이 representation을 정밀한 action으로 디코딩한다. RoboTwin 2.0, LIBERO, CALVIN에서의 실험은 각각 58\\%, 98\\%, 4.36(Avg. Len)의 state-of-the-art success rate를 보여준다. 특히 실제 sim-to-real transfer에서 BehaviorVLA는 demonstration data의 50\\%만 사용하여 OpenVLA-OFT와 동등한 성능을 달성하며, 우수한 data efficiency와 generalization을 입증한다.

"},{"id":"66122","en":"GAE: Unleashing Physical Potential of VLM with Generalizable Action Expert","ko":"GAE: 일반화 가능한 행동 전문가로 VLM의 물리적 잠재력 발휘하기","authors":"Mingyu Liu, Zheng Huang, Xiaoyi Lin, Muzhi Zhu, Canyu Zhao, Yating Wang, Haoyi Zhu, Hao Chen, Chunhua Shen","pos":"#804","link":"https://openreview.net/forum?id=6lq2MGo42H","abs":"

Vision-language models demonstrate strong reasoning and planning abilities, yet grounding these predictions into precise robot actions remains a central challenge. Existing Vision-Language-Action methods typically entangle reasoning and action generation, leading to limited generalization and costly adaptation. We propose to learn a \\textbf{G}eneralizable \\textbf{A}ction \\textbf{E}xpert (\\textbf{GAE}), a task-agnostic model that converts sparse geometric plans into dense robot actions. Our approach introduces a sparse geometric interface: the VLM predicts sparse 3D waypoints representing high-level intention, while GAE maps these waypoints together with real-time point cloud observations to continuous action trajectories. GAE is pretrained on a large-scale pointcloud–trajectory dataset comprising \\textbf{150k} trajectories from both simulation and real-world robots. To further improve efficiency and generalization, we introduce an \\textbf{Action Pre-training, Pointcloud Fine-tuning (APPF)} scheme that decouples learning action dynamics from geometry grounding. After pretraining, GAE is frozen and reused across downstream tasks, requiring only lightweight fine-tuning of the VLM to produce the sparse interface. Extensive experiments show that our method achieves strong performance and generalization across diverse visual domains, camera viewpoints, and natural language instructions.

","absKo":"

Vision-language model은 강력한 reasoning과 planning 능력을 보여주지만, 이러한 예측을 정밀한 robot action으로 grounding하는 것은 여전히 핵심 과제입니다. 기존 Vision-Language-Action 방법은 일반적으로 reasoning과 action generation을 뒤섞어, generalization이 제한되고 adaptation 비용이 높습니다. 우리는 sparse geometric plan을 dense robot action으로 변환하는 task-agnostic model인 \\textbf{G}eneralizable \\textbf{A}ction \\textbf{E}xpert(\\textbf{GAE})를 학습하는 방법을 제안합니다. 우리의 접근법은 sparse geometric interface를 도입합니다. VLM은 고수준 의도를 나타내는 sparse 3D waypoint를 예측하고, GAE는 이러한 waypoint와 실시간 point cloud observation을 continuous action trajectory로 매핑합니다. GAE는 시뮬레이션과 실제 robot에서 얻은 \\textbf{150k} trajectories로 구성된 대규모 pointcloud–trajectory dataset으로 pretraining됩니다. 효율성과 generalization을 더 향상시키기 위해, 우리는 action dynamics 학습과 geometry grounding을 분리하는 \\textbf{Action Pre-training, Pointcloud Fine-tuning(APPF)} scheme을 도입합니다. pretraining 후 GAE는 frozen 상태로 유지되어 downstream task 전반에 재사용되며, sparse interface를 생성하기 위해 VLM만 lightweight fine-tuning하면 됩니다. 광범위한 실험은 제안 방법이 다양한 visual domain, camera viewpoint, natural language instruction 전반에서 강력한 성능과 generalization을 달성함을 보여줍니다.

"},{"id":"66451","en":"HDFlow: Hierarchical Diffusion-Flow Planning for Long-horizon Tasks","ko":"HDFlow: 장기 지평 작업을 위한 계층적 Diffusion-Flow 계획","authors":"Gireesh Nandiraju, Yuanliang(Avery) Ju, Chaoyi Xu, Weiheng Liu, Yuxuan Wan, He Wang","pos":"#807","link":"https://openreview.net/forum?id=3ZFvXHJwmB","abs":"Recent advances in generative models have shown promise in generating behavior plans for long-horizon, sparse reward tasks. While these approaches have achieved promising results, they often lack a principled framework for hierarchical decomposition and struggle with the computational demands of real-time execution, due to their iterative denoising process. In this work, we introduce $\\textbf{Hierarchical Diffusion-Flow}$ ($\\texttt{\\textbf{HDFlow}}$), a novel hierarchical planning framework that optimally leverages the strengths of $\\textit{diffusion}$ and $\\textit{rectified flow}$ models to overcome the limitations of single-paradigm generative planners. $\\texttt{\\textbf{HDFlow}}$ employs a high-level diffusion planner to generate sequences of strategic subgoals in a learned latent space, capitalizing on diffusion's powerful exploratory capabilities. These subgoals then guide a low-level rectified flow planner that generates smooth and dense trajectories, exploiting the speed and efficiency of ordinary differential equation (ODE)-based trajectory generation. We evaluate $\\texttt{\\textbf{HDFlow}}$ on four challenging furniture assembly tasks in both simulation and real-world, where it significantly outperforms state-of-the-art methods. Furthermore, we also showcase our method's generalizability on two long-horizon benchmarks comprising diverse locomotion and manipulation tasks. Project website: https://hdflow-page.github.io/","absKo":"최근 generative model의 발전은 long-horizon, sparse reward task를 위한 behavior plan 생성에서 가능성을 보여주었다. 이러한 접근은 유망한 결과를 달성했지만, 계층적 분해를 위한 원칙적인 framework가 부족하고, iterative denoising process 때문에 real-time execution의 계산 요구를 처리하는 데 어려움을 겪는다. 이 연구에서는 single-paradigm generative planner의 한계를 극복하기 위해 diffusion과 rectified flow model의 강점을 최적으로 활용하는 새로운 hierarchical planning framework인 $\\textbf{Hierarchical Diffusion-Flow}$ ($\\texttt{\\textbf{HDFlow}}$)를 제안한다. $\\texttt{\\textbf{HDFlow}}$는 상위 수준에서 diffusion planner를 사용해 학습된 latent space에서 전략적 subgoal의 sequence를 생성하며, diffusion의 강력한 탐색 능력을 활용한다. 이후 이 subgoal은 하위 수준 rectified flow planner를 안내하여, ordinary differential equation (ODE)-based trajectory generation의 속도와 효율성을 활용해 smooth하고 dense한 trajectory를 생성한다. 우리는 simulation과 real-world 모두에서 네 개의 어려운 furniture assembly task에 대해 $\\texttt{\\textbf{HDFlow}}$를 평가했으며, 그 결과 state-of-the-art method를 크게 능가했다. 또한 다양한 locomotion 및 manipulation task로 구성된 두 개의 long-horizon benchmark에서 본 방법의 generalizability도 보여준다. Project website: https://hdflow-page.github.io/"},{"id":"61357","en":"Hydra-Nav: Object Navigation via Adaptive Dual-Process Reasoning","ko":"Hydra-Nav: Adaptive Dual-Process Reasoning을 통한 객체 내비게이션","authors":"Zixuan Wang, Huang Fang, Shaoan Wang, Yuanfei Luo, Heng Dong, Wei Li, Yiming Gan","pos":"#808","link":"https://openreview.net/forum?id=r2RVWpvPk7","abs":"

While large vision-language models (VLMs) show promise for object goal navigation, current methods still struggle with low success rates and inefficient localization of unseen objects—failures primarily attributed to weak temporal-spatial reasoning. Meanwhile, recent attempts to inject reasoning into VLM-based agents improve success rates but incur substantial computational overhead. To address both the ineffectiveness and inefficiency of existing approaches, we introduce Hydra-Nav, a unified VLM architecture that adaptively switches between a deliberative \"slow system\" for analyzing exploration history and formulating high-level plans, and a reactive \"fast system\" for efficient execution. We train Hydra-Nav through a three-stage curriculum: (i) spatial-action alignment to strengthen trajectory planning, (ii) memory-reasoning integration to enhance temporal-spatial reasoning over long-horizon exploration, and (iii) iterative rejection fine-tuning to enable selective reasoning at critical decision points. Extensive experiments demonstrate that Hydra-Nav achieves state-of-the-art performance on the HM3D, MP3D, and OVON benchmarks, outperforming the second-best methods by 11.1\\%, 17.4\\%, and 21.2\\%, respectively. Furthermore, we introduce SOT (Success weighted by Operation Time), a new metric to measure search efficiency across VLMs with varying reasoning intensity. Results show that adaptive reasoning significantly enhances search efficiency over fixed-frequency baselines.

","absKo":"

대형 vision-language model (VLM)은 object goal navigation에 유망하지만, 현재 방법은 여전히 낮은 성공률과 unseen object의 비효율적 localization 문제를 겪는다. 이러한 실패는 주로 미약한 temporal-spatial reasoning에 기인한다. 한편, VLM 기반 agent에 reasoning을 주입하려는 최근 시도는 성공률을 향상시키지만 상당한 계산 오버헤드를 초래한다. 기존 접근의 비효율성과 비효과성을 동시에 해결하기 위해, 우리는 탐색 이력을 분석하고 고수준 계획을 수립하는 deliberative한 \"slow system\"과 효율적 실행을 위한 reactive한 \"fast system\" 사이를 적응적으로 전환하는 통합 VLM architecture인 Hydra-Nav를 제안한다. 우리는 세 단계 curriculum을 통해 Hydra-Nav를 학습한다: (i) trajectory planning을 강화하기 위한 spatial-action alignment, (ii) 장기 탐색에서 temporal-spatial reasoning을 향상시키는 memory-reasoning integration, (iii) 중요한 의사결정 지점에서 선택적 reasoning을 가능하게 하는 iterative rejection fine-tuning. 광범위한 실험은 Hydra-Nav가 HM3D, MP3D, OVON benchmark에서 state-of-the-art 성능을 달성하며, 각각 두 번째로 좋은 방법보다 11.1%, 17.4%, 21.2% 향상됨을 보여준다. 또한, 서로 다른 reasoning 강도를 갖는 VLM 전반의 탐색 효율성을 측정하기 위한 새로운 지표 SOT (Success weighted by Operation Time)를 제안한다. 결과는 적응적 reasoning이 고정 빈도 baseline보다 탐색 효율성을 유의미하게 향상시킴을 보인다.

"},{"id":"61826","en":"PSG-Nav: Probabilistic Scene Graph Navigation via Multiverse Decision Making","ko":"PSG-Nav: Multiverse Decision Making을 통한 확률적 Scene Graph Navigation","authors":"Rufeng Chen, Yue Chang, Xiaqiang Tang, Hechang Chen, Sihong Xie","pos":"#812","link":"https://openreview.net/forum?id=mcaZVJ7Apl","abs":"

Open-vocabulary navigation requires embodied agents to manage significant perception uncertainty stemming from semantic ambiguity and model errors. However, most existing works settle for local optimal deterministic approaches, depriving complex navigation decision-making over multiple composite possibilities that are critical for globally better solutions. In this paper, we propose Probabilistic Scene Graph Navigation (PSG-Nav), which constructs a 3D Probabilistic Scene Graph that uses full semantic categorical distributions to account for perception uncertainty. To efficiently use the local distributions to compose and reason about the optimal navigation landmarks, we propose Multiverse Decision to sample multiple most likely world settings from the joint distribution, and evaluate navigation landmarks based on the compatibility between landmarks and multiverses. To mitigate false positives due to epistemic uncertainty in open-vocabulary navigation, we introduce the Evidential Experience Calibrator, which enables online lifelong adaptation by cross-validating detections against memories of past successes and failures. Extensive experiments on widely-used benchmarks MP3D, HD3D, and HSSD demonstrate that PSG-Nav establishes new state-of-the-art results, achieving a Success Rate of 66.1%, 44.8%, and 67.9%, respectively.

","absKo":"Open-vocabulary navigation은 embodied agent가 semantic ambiguity와 model error에서 비롯되는 상당한 perception uncertainty를 다뤄야 한다. 그러나 기존 연구의 대부분은 local optimal deterministic 접근에 머물러, 전역적으로 더 나은 해에 중요한 여러 composite possibility를 아우르는 복잡한 navigation decision-making을 포기한다. 본 논문에서는 perception uncertainty를 반영하기 위해 full semantic categorical distribution을 사용하는 3D Probabilistic Scene Graph를 구성하는 Probabilistic Scene Graph Navigation (PSG-Nav)을 제안한다. 로컬 distribution을 효율적으로 사용해 최적의 navigation landmark를 구성하고 추론하기 위해, 우리는 joint distribution에서 가장 가능성 높은 여러 world setting을 샘플링하고, landmark와 multiverse 사이의 compatibility에 기반해 navigation landmark를 평가하는 Multiverse Decision을 제안한다. open-vocabulary navigation에서 epistemic uncertainty로 인한 false positive를 완화하기 위해, 과거 성공과 실패의 기억과 detection을 cross-validating하여 online lifelong adaptation을 가능하게 하는 Evidential Experience Calibrator를 도입한다. 널리 사용되는 benchmark인 MP3D, HD3D, HSSD에서 수행한 광범위한 실험은 PSG-Nav이 새로운 state-of-the-art 결과를 확립하며 각각 66.1%, 44.8%, 67.9%의 Success Rate를 달성함을 보여준다.

"},{"id":"63554","en":"Plan in Sandbox, Navigate in Open Worlds: Learning Physics-Grounded Abstracted Experience for Embodied Navigation","ko":"샌드박스에서 계획하고 오픈 월드에서 탐색하기: embodied navigation을 위한 물리 기반 추상 경험 학습","authors":"Zhixuan Shen, Jiawei Du, Ziyu Guo, Han Luo, Lilan Peng, Joey Tianyi Zhou, Haonan Luo, Tianrui Li","pos":"#813","link":"https://openreview.net/forum?id=W5e8c9nwNo","abs":"

Vision-Language Models (VLMs) have demonstrated exceptional general reasoning capabilities. However, their performance in embodied navigation remains hindered by a scarcity of aligned open-world vision and robot control data. Despite simulators providing a cost-effective alternative for data collection, the inherent reliance on photorealistic simulations often limits the transferability of learned policies. To this end, we propose \\textit{\\textbf{S}andbox-\\textbf{A}bstracted \\textbf{G}rounded \\textbf{E}xperience} (\\textbf{\\textit{SAGE}}), a framework that enables agents to learn within a physics-grounded semantic abstraction rather than a photorealistic simulation, mimicking the human capacity for mental simulation where plans are rehearsed in simplified physics abstractions before execution. \\textit{SAGE} system operates via three synergistic phases: (1) \\textit{Genesis}: constructing diverse, physics-constrained semantic environments to bootstrap experience; (2) \\textit{Evolution}: distilling experiences through Reinforcement Learning (RL), utilizing a novel asymmetric adaptive clipping mechanism to stabilize updates; (3) \\textit{Navigation}: bridging the abstract policy to real-world control. We demonstrate that \\textit{SAGE} significantly improves navigation performance, achieving a 53.21\\% LLM-Match Success Rate on A-EQA (+9.7\\% over baseline) and generalizing widely to real-world deployments.

","absKo":"

Vision-Language Models (VLMs)은 탁월한 일반 추론 능력을 보여왔다. 그러나 embodied navigation에서의 성능은 정렬된 open-world vision 및 robot control data의 부족으로 인해 여전히 제한된다. 시뮬레이터는 데이터 수집을 위한 비용 효율적인 대안을 제공하지만, photorealistic simulation에 대한 본질적 의존은 학습된 policy의 transferability를 종종 제한한다. 이를 위해 우리는 \\textit{\\textbf{S}andbox-\\textbf{A}bstracted \\textbf{G}rounded \\textbf{E}xperience} (\\textbf{\\textit{SAGE}})를 제안한다. 이는 agents가 photorealistic simulation이 아니라 physics-grounded semantic abstraction 내부에서 학습하도록 하는 framework로, 계획이 실행 전에 단순화된 physics abstraction에서 반복적으로 연습되는 인간의 mental simulation 능력을 모방한다. \\textit{SAGE} system은 세 가지 상호보완적 phase로 동작한다. (1) \\textit{Genesis}: 경험을 부트스트랩하기 위해 다양한 physics-constrained semantic environment를 구성한다. (2) \\textit{Evolution}: novel asymmetric adaptive clipping mechanism을 활용하여 update를 안정화하면서 Reinforcement Learning (RL)로 경험을 증류한다. (3) \\textit{Navigation}: abstract policy와 실제 world control을 연결한다. 우리는 \\textit{SAGE}가 navigation performance를 크게 향상시켜 A-EQA에서 53.21\\%의 LLM-Match Success Rate를 달성하고(+9.7\\% over baseline), 실제 배치로도 폭넓게 일반화됨을 보인다.

"},{"id":"61342","en":"Plug-and-Play Label Map Diffusion for Universal Goal-Oriented Navigation","ko":"범용 목표 지향 Navigation을 위한 Plug-and-Play Label Map Diffusion","authors":"Zhixuan Shen, Yijie Zeng, Shengxiang Luo, Tianrui Li, Haonan Luo","pos":"#814","link":"https://openreview.net/forum?id=r8vdDUZ4s3","abs":"

In embodied vision, Goal-Oriented Navigation (GON) requires robots to locate a specific goal within an unexplored environment. The primary challenge of GON arises from the need to construct a Bird's-Eye-View (BEV) map to understand the environment while simultaneously localizing an unobserved goal. Existing map-based methods typically employ self-centered semantic maps, often facing challenges such as reliance on complete maps or inconsistent semantic association. To this end, we propose Plug-and-Play Label Map Diffusion (PLMD), which defines a novel map completion diffusion model based on Denoising Diffusion Probabilistic Models (DDPM). PLMD generates obstacle and semantic labels for unobserved regions through a diffusion-based completion process, thereby enabling goal localization even in partially observed environments. Moreover, it mitigates inconsistent semantic association by leveraging structural consistency between known and unknown obstacle layouts and integrating obstacle priors into the semantic denoising process. By substituting predicted labels for unobserved regions, robots can accurately localize the specified objects. Extensive experiments demonstrate that PLMD \\textbf{(I)} effectively expands the region of unknown maps, \\textbf{(II)} integrates seamlessly into existing navigation strategies that rely on semantic maps, \\textbf{(III)} achieves state-of-the-art performance on three GON tasks.

","absKo":"

embodied vision에서 Goal-Oriented Navigation (GON)은 로봇이 탐색되지 않은 환경 안에서 특정 목표를 찾아야 하는 과제를 요구한다. GON의 핵심 도전은 환경을 이해하기 위한 Bird's-Eye-View (BEV) map을 구성하는 동시에, 관측되지 않은 목표를 localize해야 한다는 점에서 비롯된다. 기존의 map-based 방법은 일반적으로 self-centered semantic map을 사용하지만, 완전한 map에 대한 의존성이나 일관되지 않은 semantic association 같은 문제에 자주 직면한다. 이를 위해 우리는 Plug-and-Play Label Map Diffusion (PLMD)를 제안한다. PLMD는 Denoising Diffusion Probabilistic Models (DDPM) 기반의 새로운 map completion diffusion model을 정의한다. PLMD는 diffusion 기반 completion 과정을 통해 관측되지 않은 영역의 obstacle과 semantic label을 생성함으로써, 부분적으로만 관측된 환경에서도 goal localization을 가능하게 한다. 또한 known obstacle layout과 unknown obstacle layout 사이의 structural consistency를 활용하고 obstacle prior를 semantic denoising 과정에 통합하여 일관되지 않은 semantic association을 완화한다. 관측되지 않은 영역에 대해 예측된 label을 대체함으로써, 로봇은 지정된 객체를 정확하게 localize할 수 있다. 광범위한 실험은 PLMD가 \\textbf{(I)} unknown map 영역을 효과적으로 확장하고, \\textbf{(II)} semantic map에 의존하는 기존 navigation 전략에 매끄럽게 통합되며, \\textbf{(III)} 세 가지 GON task에서 state-of-the-art 성능을 달성함을 보여준다.

"},{"id":"62776","en":"Embodiment-Conditioned Mixture of Experts Increases the Evolvability of Robots","ko":"Embodiment-Conditioned Mixture of Experts가 로봇의 Evolvability를 증가시킴","authors":"Yibin Wang, Muhan Li, Zihan Guo, Sam Kriegman","pos":"#900","link":"https://openreview.net/forum?id=dTIxiQ0jhu","abs":"

In this paper, we introduce a model of evolution and learning in robots that co-optimizes a distribution of latent design vectors (genotypes) and a mixture of control experts (neural modules), which are gated by the latent coordinates of each decoded design (phenotype). This provides a scalable alternative to co-design algorithms that either train an individual policy for every robot, which is inefficient, or a monolithic universal controller for all robots, which results in overly conservative structures and behaviors. Our approach lies somewhere between these two extremes, preserving ancestral knowledge in a unified yet modular framework in which different body plans activate and deactivate different combinations of learned sensorimotor circuits for goal-directed behavior. This allows one part of the controller to be overhauled to better suit new species of designs as they emerge without disrupting the hard-earned knowledge contained within other expert modules. Pretrained expert policies can also be directly plugged into the mixture, providing a simple mechanism to indirectly steer evolution into areas of latent space containing desired morphological traits. We refer to this process as \"evolution by demo\" and use it to direct evolution toward the canonical form defined by the pretrained policy.

","absKo":"

이 논문에서는 latent design vector (genotype) 분포와 control expert (neural module) mixture를 co-optimize하는 로봇의 evolution과 learning 모델을 소개한다. 이때 expert는 각 decoded design (phenotype)의 latent coordinate에 의해 gating된다. 이는 로봇마다 개별 policy를 학습해야 하므로 비효율적인 co-design algorithm과, 모든 로봇을 위한 단일 universal controller를 학습해야 하므로 지나치게 보수적인 구조와 행동을 초래하는 접근 사이의 scalable alternative를 제공한다. 우리의 접근은 이 두 극단의 중간쯤에 위치하며, 서로 다른 body plan이 목표 지향 행동을 위해 학습된 sensorimotor circuit의 서로 다른 조합을 활성화하거나 비활성화하는 통합적이면서도 modular한 framework 안에서 ancestral knowledge를 보존한다. 이를 통해 컨트롤러의 일부를 새로 등장하는 design species에 더 잘 맞도록 대대적으로 개편하더라도, 다른 expert module에 축적된 힘들게 얻은 knowledge는 방해받지 않는다. pretrained expert policy를 mixture에 직접 plug-in할 수도 있으며, 이는 원하는 morphological trait를 포함한 latent space 영역으로 evolution을 간접적으로 유도하는 간단한 메커니즘을 제공한다. 우리는 이 과정을 \"evolution by demo\"라고 부르며, pretrained policy가 정의하는 canonical form을 향하도록 evolution을 유도하는 데 사용한다.

"},{"id":"65138","en":"Efficient Skill Grounding via Code Refactoring with Small Language Models","ko":"Small Language Model을 이용한 코드 리팩토링 기반 효율적 스킬 그라운딩","authors":"Sera Choi, Wonje Choi, Saehun Chun, Daehee Lee, Chaeun Lee, Jooyoung Kim, Honguk Woo","pos":"#902","link":"https://openreview.net/forum?id=GQ4zj5YiyC","abs":"

Effective skill grounding is essential for deploying reusable skills in embodied agents, as even minor embodiment or environmental differences can render an entire skill incompatible. This challenge is particularly pronounced in embodied settings, where agents must operate in dynamic, partially observable environments without access to large language models (LLMs). In this setting, reliance on LLMs is impractical, while small language models (sLMs) remain insufficient for the effective skill grounding required for reliable long-horizon control. We present RECENT, a refactoring-centric agent framework that enables efficient skill grounding with sLMs by decoupling skill semantics from embodiment- and environment-specific execution binding. By representing skills as executable code, RECENT preserves the semantic intent encoded in a skill’s control structure while grounding it by modifying only execution bindings through localized refactoring, rather than regenerating code from scratch. We evaluate RECENT across diverse skill grounding scenarios spanning multiple robot embodiments in dynamic environments, demonstrating robust long-horizon performance when deployed with an sLM. Across all scenarios, RECENT achieves the best performance among sLM-based Code-as-Policies (CaP) methods and matches the task performance of LLM-based CaP.

","absKo":"

재사용 가능한 skill을 embodied agent에 배포하려면 효과적인 skill grounding이 필수적이다. 작은 embodiment나 environment 차이만으로도 하나의 skill 전체가 호환되지 않게 될 수 있기 때문이다. 이 문제는 agent가 large language model(LLM) 없이 동적이고 부분적으로 관측 가능한 환경에서 동작해야 하는 embodied setting에서 특히 두드러진다. 이러한 설정에서는 LLM에 의존하는 것이 비현실적이며, small language model(sLM)은 신뢰할 수 있는 long-horizon control에 필요한 효과적인 skill grounding에는 여전히 부족하다. 우리는 skill semantics를 embodiment- 및 environment-specific execution binding과 분리함으로써 sLM을 이용한 효율적인 skill grounding을 가능하게 하는 refactoring-centric agent framework인 RECENT를 제시한다. skill을 executable code로 표현함으로써, RECENT는 skill의 control structure에 인코딩된 semantic intent를 보존하는 동시에, 코드를 처음부터 재생성하는 대신 localized refactoring을 통해 execution binding만 수정하여 이를 grounding한다. 우리는 동적 환경에서 여러 robot embodiment를 아우르는 다양한 skill grounding 시나리오 전반에서 RECENT를 평가했으며, sLM과 함께 배포했을 때 견고한 long-horizon 성능을 보였다. 모든 시나리오에서 RECENT는 sLM 기반 Code-as-Policies(CaP) 방법들 중 최고의 성능을 달성했고, LLM 기반 CaP의 task 성능과도 일치했다.

"},{"id":"66030","en":"EcoVLA: Environment-Aware Adaptive Pruning with Interleaved Inference Orchestration for Vision-Language-Action Models","ko":"EcoVLA: Vision-Language-Action 모델을 위한 교차 추론 오케스트레이션을 갖춘 환경 인식 적응형 가지치기","authors":"Yuting Huang, Leilei Ding, Zhipeng Tang, Zenghuan Zhu, Jiajun Deng, Xinrui Lin, Shuo Liu, Haojie Ren, Jianmin Ji, Yanyong Zhang","pos":"#903","link":"https://openreview.net/forum?id=7kH5uphAes","abs":"While Vision-Language-Action (VLA) models hold promise in embodied intelligence, their large parameter counts lead to substantial inference latency that hinders real-time manipulation, motivating parameter sparsification. However, as the environment evolves during VLA execution, the optimal sparsity patterns change accordingly. Static pruning lacks the adaptability required for environment dynamics, whereas fixed-interval dynamic layer pruning suffers from coarse granularity and high retraining overheads. To bridge this gap, we propose **EcoVLA**, a training-free, plug-and-play adaptive pruning framework that supports orthogonal combination with existing VLA acceleration methods. EcoVLA comprises two components: **E**nvironment-aware **A**daptive **P**runing (**EAP**) and **I**nterleaved **I**nference **O**rchestration (**$I^2O$**). EAP is a lightweight adaptive channel pruning method that incorporates the temporal consistency of the physical environment to update sparsity patterns. $I^2O$ leverages the FLOPs bubbles inherent in VLA inference to schedule the pruning method in parallel, ensuring negligible impact on latency. Evaluated on diverse VLA models and benchmarks, EcoVLA delivers state-of-the-art performance, achieving up to 1.60$\\times$ speedup with only a 0.4% drop in success rate, and further reaches 2.18$\\times$ speedup with only a 0.5% degradation when combined with token pruning. We further validate the effectiveness of EcoVLA on real-world robots.","absKo":"Vision-Language-Action (VLA) model은 embodied intelligence에서 큰 잠재력을 보이지만, 큰 parameter count 때문에 inference latency가 상당히 커져 real-time manipulation을 저해하며, 이는 parameter sparsification을 필요하게 한다. 그러나 VLA execution 동안 environment가 진화함에 따라 최적 sparsity pattern도 그에 맞게 변한다. Static pruning은 environment dynamics에 필요한 적응성을 제공하지 못하고, fixed-interval dynamic layer pruning은 coarse granularity와 높은 retraining overhead를 겪는다. 이러한 간극을 메우기 위해, 우리는 기존 VLA acceleration methods와 orthogonal하게 결합 가능한 training-free, plug-and-play adaptive pruning framework인 **EcoVLA**를 제안한다. EcoVLA는 두 구성요소로 이루어진다: **E**nvironment-aware **A**daptive **P**runing (**EAP**)과 **I**nterleaved **I**nference **O**rchestration (**$I^2O$**). EAP는 물리적 environment의 temporal consistency를 활용해 sparsity pattern을 갱신하는 경량 adaptive channel pruning method이다. $I^2O$는 VLA inference에 내재된 FLOPs bubble을 활용해 pruning method를 병렬로 스케줄링하여 latency에 대한 영향이 거의 없도록 보장한다. 다양한 VLA model과 benchmark에서 평가한 결과, EcoVLA는 state-of-the-art performance를 달성했으며, success rate를 0.4%만 낮추면서 최대 1.60$\\times$ speedup을, token pruning과 결합하면 0.5%의 성능 저하만으로 2.18$\\times$ speedup을 달성했다. 우리는 또한 실제 robot에서 EcoVLA의 효과를 추가로 검증한다."},{"id":"63736","en":"DynVLA: Learning World Dynamics for Action Reasoning in Autonomous Driving","ko":"DynVLA: 자율주행에서 행동 추론을 위한 세계 동역학 학습","authors":"Shuyao Shang, Bing Zhan, Yunfei Yan, Yuqi Wang, Yingyan Li, Yasong An, Xiaoman Wang, Jierui Liu, Lu Hou, Lue Fan, Zhaoxiang Zhang, Tieniu Tan","pos":"#904","link":"https://openreview.net/forum?id=UB6M4BSvwJ","abs":"

We propose DynVLA, a driving VLA model that introduces a new CoT paradigm termed Dynamics CoT. DynVLA forecasts compact world dynamics before action generation, enabling more informed and physically grounded decision-making. To obtain compact dynamics representations, DynVLA introduces a Dynamics Tokenizer that compresses future evolution into a small set of dynamics tokens. Considering the rich environment dynamics in interaction-intensive driving scenarios, DynVLA decouples ego-centric and environment-centric dynamics, yielding more accurate world dynamics modeling. We then train DynVLA to generate dynamics tokens before actions through SFT and RFT, improving decision quality while maintaining latency-efficient inference. Compared to Textual CoT, which lacks fine-grained spatiotemporal understanding, and Visual CoT, which introduces substantial redundancy due to dense image prediction, Dynamics CoT captures the evolution of the world in a compact, interpretable, and efficient form. Extensive experiments on NAVSIM, Bench2Drive, and a large-scale in-house dataset demonstrate that DynVLA consistently outperforms Textual CoT and Visual CoT methods, validating the effectiveness and practical value of Dynamics CoT.

","absKo":"

우리는 DynVLA를 제안한다. 이는 action generation 전에 compact world dynamics를 예측하는 새로운 CoT 패러다임인 Dynamics CoT를 도입한 driving VLA model이다. DynVLA는 보다 정보가 풍부하고 물리적으로 타당한 의사결정을 가능하게 한다. compact한 dynamics representation을 얻기 위해, DynVLA는 미래 전개를 소수의 dynamics token으로 압축하는 Dynamics Tokenizer를 도입한다. 상호작용이 많은 driving scenario에서 풍부한 environment dynamics를 고려하여, DynVLA는 ego-centric dynamics와 environment-centric dynamics를 분리하여 더 정확한 world dynamics modeling을 달성한다. 이후 우리는 SFT와 RFT를 통해 action 전에 dynamics token을 생성하도록 DynVLA를 학습시켜, 지연시간 효율적인 inference를 유지하면서 decision quality를 향상시킨다. 미세한 spatiotemporal understanding이 부족한 Textual CoT와, 밀집된 image prediction으로 인해 상당한 redundancy를 유발하는 Visual CoT와 비교할 때, Dynamics CoT는 세계의 전개를 compact하고 해석 가능하며 효율적인 형태로 포착한다. NAVSIM, Bench2Drive, 그리고 대규모 in-house dataset에서 수행한 광범위한 실험은 DynVLA가 Textual CoT 및 Visual CoT 방법보다 일관되게 우수함을 보여주며, Dynamics CoT의 효과성과 실용적 가치를 검증한다.

"},{"id":"60681","en":"DyGRO-VLA: Cross-Task Scaling of Vision–Language–Action Models via Dynamic Grouped Residual Optimization","ko":"DyGRO-VLA: Dynamic Grouped Residual Optimization을 통한 Vision-Language-Action Model의 Cross-Task Scaling","authors":"Sixu Lin, Yunpeng Qing, Litao Liu, Ming Zhou, Ruixing Jin, Xiaoyi Fan, Guiliang Liu","pos":"#905","link":"https://openreview.net/forum?id=xg5k9pCYR7","abs":"

Recent progress in Reinforcement Learning (RL) provides a principled approach to optimizing Vision-Language-Action (VLA) models, facilitating a shift from trajectory imitation to active learning in the task environment. Despite improvements in control precision, most RL optimizers remain task-specific, which reduces VLA models from generalist controllers to policies that overfit to a narrow set of tasks. In this study, we conduct an in-depth analysis of this phenomenon and highlight the importance of cross-task feature representations for improving the generalizability of VLA models. Motivated by this finding, we introduce DyGRO-VLA, a two-stage optimization framework that 1) effectively captures cross-task latent representations based on information-theoretic principles, and 2) dynamically refines policy optimization via a mixture-of-RL-residuals. DyGRO-VLA enables the RL optimizer to exploit task-relevant latent information while strategically mitigating adverse interference on the learned representations throughout the optimization process. We evaluate our approach on LIBERO, RoboTwin2 benchmarks, and further validate it on real world, demonstrating consistent improvements over strong baselines under multi-task training and distribution shift.

","absKo":"

Reinforcement Learning (RL)의 최근 발전은 Vision-Language-Action (VLA) model을 최적화하는 데 있어 원칙적인 접근법을 제공하며, trajectory imitation에서 task environment 내 active learning으로의 전환을 촉진한다. control precision은 향상되었지만, 대부분의 RL optimizer는 여전히 task-specific이어서 VLA model을 generalist controller에서 좁은 task 집합에 과적합되는 policy로 축소시킨다. 본 연구에서는 이 현상을 심층적으로 분석하고, VLA model의 일반화 가능성을 높이기 위해 cross-task feature representation의 중요성을 강조한다. 이러한 발견에 동기부여되어, 우리는 1) information-theoretic principle에 기반해 cross-task latent representation을 효과적으로 포착하고, 2) mixture-of-RL-residuals를 통해 policy optimization을 동적으로 정제하는 두 단계 최적화 framework인 DyGRO-VLA를 제안한다. DyGRO-VLA는 RL optimizer가 task-related latent information을 활용하도록 하는 한편, optimization 과정 전반에 걸쳐 학습된 representation에 대한 adverse interference를 전략적으로 완화한다. 우리는 LIBERO, RoboTwin2 benchmark에서 우리의 접근을 평가하고, real world에서도 추가 검증하여, multi-task training과 distribution shift 하에서 강력한 baseline 대비 일관된 성능 향상을 보임을 입증한다.

"},{"id":"63713","en":"Distinguishing Imitation Error from Intrinsic Motion Learning Difficulty","ko":"모방 오류와 내재적 모션 학습 난이도 구분하기","authors":"Zhaorui Meng, Lu Yin, Xinrui Chen, Chengxu Zuo, Anjun Chen, Guo Shihui, Yipeng Qin","pos":"#907","link":"https://openreview.net/forum?id=UNWIi5Qjg1","abs":"

Physics-based motion imitation is central to humanoid control, yet current evaluation metrics(e.g., MPJPE) only quantify imitation outcomes, not their underlying causes. This conflation obscures a critical diagnostic question: when imitation error occurs, does it stem from policy limitations or the intrinsic learning difficulty of the target motion? To resolve this ambiguity, we propose the Torque Variation Score (TVS), a physics-grounded metric that quantifies the inherent learning difficulty of a motion independently of any policy's performance. TVS measures the magnitude of torque variation required to correct small pose perturbations, directly capturing how dynamical properties shape the reinforcement learning landscape. We establish that high-TV motions induce flat reward landscapes and vanishing policy gradients, explaining persistent imitation failures. Extensive experiments with state-of-the-art methods (UHC, PHC+) confirm TVS strongly correlates with imitation error and enables principled error attribution: high error on low-TV motions indicates policy deficiency, while high error on high-TV motions reflects fundamental learning constraints. Beyond error diagnosis, TVS facilitates three practical applications: Maximum Imitable Difficulty (MID) for policy capability assessment, Difficulty-Stratified Joint Error (DSJE) for granular performance profiling, and Flawed Motion Detection for identifying segments with abnormally high learning difficulty to support mocap data curation and quality control. TVS provides a rigorous lens to distinguish policy-induced errors from motion-inherent challenges and enhances motion dataset reliability.

","absKo":"Physics-based motion imitation은 humanoid control의 핵심이지만, 현재의 평가 지표(e.g., MPJPE)는 imitation의 결과만을 정량화할 뿐 그 근본 원인은 측정하지 못한다. 이러한 혼동은 중요한 진단 질문을 가린다: imitation error가 발생했을 때, 그것은 policy의 한계 때문인가, 아니면 목표 motion 자체의 내재적 학습 난이도 때문인가? 이 모호성을 해소하기 위해, 우리는 정책의 성능과 독립적으로 motion의 본질적 학습 난이도를 정량화하는 physics-grounded metric인 Torque Variation Score (TVS)를 제안한다. TVS는 작은 pose perturbation을 보정하는 데 필요한 torque variation의 크기를 측정하며, 동역학적 특성이 reinforcement learning landscape를 어떻게 형성하는지를 직접 포착한다. 우리는 높은 TV motion이 평탄한 reward landscape와 사라지는 policy gradient를 유도함을 보이며, 이는 지속적인 imitation failure를 설명한다. 최신 state-of-the-art method(UHC, PHC+)를 사용한 광범위한 실험은 TVS가 imitation error와 강하게 상관됨을 확인하고, 원리적인 error attribution을 가능하게 함을 보여준다: low-TV motion에서 높은 error는 policy deficiency를 의미하는 반면, high-TV motion에서 높은 error는 근본적인 학습 제약을 반영한다. 오류 진단을 넘어, TVS는 세 가지 실용적 응용을 가능하게 한다: policy capability assessment를 위한 Maximum Imitable Difficulty (MID), 세분화된 성능 프로파일링을 위한 Difficulty-Stratified Joint Error (DSJE), 그리고 mocap data curation과 quality control을 지원하기 위해 비정상적으로 높은 학습 난이도를 가진 segment를 식별하는 Flawed Motion Detection. TVS는 policy가 유발한 오류와 motion 고유의 도전을 구분하는 엄밀한 관점을 제공하며 motion dataset의 신뢰성을 높인다.

"},{"id":"62484","en":"DeepSight: Long-Horizon World Modeling via Latent States Prediction for End-to-End Autonomous Driving","ko":"DeepSight: 잠재 상태 예측 기반 장기 Horizon World Modeling for End-to-End 자율주행","authors":"Lingjun Zhang, Changjie Wu, Linzhe Shi, Jiangyang Li, Jiaxin Liu, Lei Yang, Hang Zhang, Mu Xu, Hong Wang","pos":"#909","link":"https://openreview.net/forum?id=gBq6eiSTE4","abs":"

End-to-end autonomous driving systems are increasingly integrating Vision-Language Model (VLM) architectures, incorporating text reasoning or visual reasoning to enhance the robustness and accuracy of driving decisions. However, the reasoning mechanisms employed in most methods are direct adaptations from general domains, lacking in-depth exploration tailored to autonomous driving scenarios, particularly within visual reasoning modules. In this paper, we propose a driving world model that performs parallel prediction of latent semantic features for consecutive future frames in the bird’s-eye-view (BEV) space, thereby enabling long-horizon modeling of future world states. We also introduce an efficient and adaptive text reasoning mechanism that utilizes additional social knowledge and reasoning capabilities to further improve driving performance in challenging long-tail scenarios. We present a novel, efficient, and effective approach that achieves state-of-the-art (SOTA) results on the closed-loop Bench2drive benchmark. Code will be released soon.

","absKo":"

종단간 자율주행 시스템은 점점 Vision-Language Model (VLM) 아키텍처를 통합하고 있으며, 주행 의사결정의 강건성과 정확도를 높이기 위해 텍스트 추론 또는 시각 추론을 포함하고 있다. 그러나 대부분의 방법에서 사용되는 추론 메커니즘은 일반 도메인에서 직접 가져온 변형에 불과하며, 특히 visual reasoning 모듈을 포함한 자율주행 시나리오에 맞춘 심층적인 탐구는 부족하다. 본 논문에서는 bird’s-eye-view (BEV) 공간에서 연속적인 미래 프레임의 latent semantic feature를 병렬로 예측하는 driving world model을 제안하여, 미래 세계 상태의 long-horizon modeling을 가능하게 한다. 또한 어려운 long-tail 시나리오에서 주행 성능을 더욱 향상시키기 위해 추가적인 social knowledge와 reasoning capability를 활용하는 효율적이고 적응적인 text reasoning mechanism도 도입한다. 우리는 closed-loop Bench2drive benchmark에서 state-of-the-art (SOTA) 결과를 달성하는 새롭고 효율적이며 효과적인 접근법을 제시한다. 코드는 곧 공개될 예정이다.

"},{"id":"64411","en":"AIR-VLA: Vision-Language-Action Systems for Aerial Manipulation","ko":"AIR-VLA: 공중 조작을 위한 Vision-Language-Action 시스템","authors":"Jianli Sun, Bin Tian, Qiyao Zhang, Chengxiang Li, Zihan Song, Zhiyong Cui, Yisheng Lv, Yonglin Tian","pos":"#913","link":"https://openreview.net/forum?id=NuR4lG4gKB","abs":"

While Vision-Language-Action (VLA) models have achieved remarkable success in ground-based embodied intelligence, their application to Aerial Manipulation Systems (AMS) remains a largely unexplored frontier. The inherent characteristics of AMS, including floating-base dynamics, strong coupling between the UAV and the manipulator, and the multi-step, long-horizon nature of operational tasks, pose severe challenges to existing VLA paradigms designed for static or 2D mobile bases. To bridge this gap, we propose AIR-VLA, the first VLA benchmark specifically tailored for aerial manipulation. We construct a physics-based simulation environment and release a high-quality multimodal dataset comprising 3000 manually teleoperated demonstrations, covering base manipulation, object & spatial understanding, semantic reasoning, and long-horizon planning. Leveraging this platform, we systematically evaluate mainstream VLA models and state-of-the-art VLM models. Our experiments not only validate the feasibility of transferring VLA paradigms to aerial systems but also, through multi-dimensional metrics tailored to aerial tasks, reveal the capabilities and boundaries of current models regarding UAV mobility, manipulator control, and high-level planning. AIR-VLA establishes a standardized testbed and data foundation for future research in general-purpose aerial robotics. The resource of AIR-VLA will be available at https://anonymous.4open.science/r/AIR-VLA-dataset-B5CC/.

","absKo":"

Vision-Language-Action(VLA) 모델은 지상 기반 embodied intelligence에서 놀라운 성과를 거두었지만, Aerial Manipulation Systems(AMS)에의 적용은 아직 크게 탐구되지 않은 영역이다. AMS 고유의 특성, 즉 floating-base dynamics, UAV와 manipulator 사이의 강한 결합, 그리고 다단계의 장기 지평(long-horizon) 작업 특성은 정적이거나 2D mobile base를 위해 설계된 기존 VLA 패러다임에 심각한 도전 과제를 제기한다. 이 격차를 메우기 위해 우리는 항공 조작에 특화된 최초의 VLA 벤치마크인 AIR-VLA를 제안한다. 우리는 물리 기반 시뮬레이션 환경을 구축하고, base manipulation, object & spatial understanding, semantic reasoning, 장기 계획을 포괄하는 3000개의 수동 teleoperated demonstration으로 구성된 고품질 multimodal dataset을 공개한다. 이 플랫폼을 활용하여, 우리는 주류 VLA 모델과 최신 VLM 모델을 체계적으로 평가한다. 우리의 실험은 VLA 패러다임을 항공 시스템으로 이전하는 것이 가능함을 검증할 뿐만 아니라, 항공 작업에 맞춘 다차원 지표를 통해 UAV 이동성, manipulator 제어, 고수준 계획에 관한 현재 모델의 능력과 한계를 드러낸다. AIR-VLA는 범용 항공 로보틱스의 향후 연구를 위한 표준화된 testbed와 데이터 기반을 확립한다. AIR-VLA의 자원은 https://anonymous.4open.science/r/AIR-VLA-dataset-B5CC/에서 제공될 예정이다.

"},{"id":"61922","en":"HALO: A Unified Vision-Language-Action Model for Embodied Multimodal Chain-of-Thought Reasoning","ko":"HALO: embodied 멀티모달 Chain-of-Thought 추론을 위한 통합 Vision-Language-Action 모델","authors":"Quanxin Shou, Fangqi Zhu, Shuang Chen, Puxin Yan, Zhengyang Yan, Yikun Miao, Xiaoyi Pang, Zicong Hong, Ruikai Shi, Hao HUANG, Jie ZHANG, Song Guo","pos":"#806","link":"https://openreview.net/forum?id=lduY9csXqw","abs":"Vision–Language–Action (VLA) models have shown strong performance in robotic manipulation, but often struggle in long-horizon or out-of-distribution scenarios due to the lack of explicit mechanisms for multimodal reasoning and anticipating how the world will evolve under action. Recent works introduce textual chain-of-thought or visual subgoal prediction within VLA models to reason, but still fail to offer a unified human-like reasoning framework for joint textual reasoning, visual foresight, and action prediction. To this end, we propose HALO, a unified VLA model that enables embodied multimodal chain-of-thought (EM-CoT) reasoning through a sequential process of textual task reasoning, visual subgoal prediction for fine-grained guidance, and EM-CoT-augmented action prediction. We instantiate HALO with a Mixture-of-Transformers (MoT) architecture that decouples semantic reasoning, visual foresight, and action prediction into specialized experts while allowing seamless cross-expert collaboration. To enable HALO learning at scale, we introduce an automated pipeline to synthesize EM-CoT training data along with a carefully crafted training recipe. Extensive experiments demonstrate that: (1) HALO achieves superior performance in both simulated and real-world environments, surpassing baseline policy $\\pi_0$ by 34.1\\% on RoboTwin benchmark; (2) all proposed components of the training recipe and EM-CoT design help improve task success rate; and (3) HALO exhibits strong generalization capabilities under aggressive unseen environmental randomization with our proposed EM-CoT reasoning.","absKo":"Vision–Language–Action (VLA) 모델은 로봇 조작에서 뛰어난 성능을 보여주었지만, multimodal reasoning과 행동에 따라 세계가 어떻게 변할지를 예측하는 명시적 메커니즘이 부족해 long-horizon이나 out-of-distribution 상황에서는 종종 어려움을 겪는다. 최근 연구들은 VLA 모델 내부에 textual chain-of-thought나 visual subgoal prediction을 도입해 추론을 수행하지만, 텍스트 추론, visual foresight, action prediction을 결합한 통합적이고 인간과 유사한 reasoning framework를 제공하는 데에는 여전히 실패하고 있다. 이를 위해 우리는 embodied multimodal chain-of-thought (EM-CoT) reasoning을 가능하게 하는 통합 VLA 모델 HALO를 제안한다. HALO는 sequential process를 통해 textual task reasoning, 세밀한 지도를 위한 visual subgoal prediction, 그리고 EM-CoT로 증강된 action prediction을 수행한다. 우리는 HALO를 Mixture-of-Transformers (MoT) architecture로 구현하여 semantic reasoning, visual foresight, action prediction을 특화된 experts로 분리하면서도 experts 간의 원활한 협력을 가능하게 한다. HALO의 대규모 학습을 위해, 우리는 EM-CoT training data를 자동으로 합성하는 pipeline과 이를 뒷받침하는 정교한 training recipe를 도입한다. 광범위한 실험은 다음을 보여준다. (1) HALO는 simulated 및 real-world 환경 모두에서 우수한 성능을 달성하며, RoboTwin benchmark에서 baseline policy $\\pi_0$를 34.1\\% 상회한다. (2) 제안한 training recipe와 EM-CoT design의 모든 구성요소가 task success rate 향상에 기여한다. (3) HALO는 우리가 제안한 EM-CoT reasoning을 통해 공격적인 unseen environmental randomization 하에서도 강한 generalization capability를 보인다."},{"id":"62343","en":"Learning Cardiac Latent Representations in Vectorcardiogram Space","ko":"Vectorcardiogram 공간의 심장 잠재표현 학습","authors":"Bosong Huang, Panzhen Zhao, Zengxiang Li, Patricia Lee, Wei Jin, Alan Liew, Ming Jin, Shirui Pan","pos":"#1012","link":"https://openreview.net/forum?id=hS6iw4PM8K","abs":"

Electrocardiography (ECG) is a cornerstone of cardiac assessment, making the learning of informative ECG representations fundamental to tasks ranging from disease diagnosis to clinical report generation. However, existing methods operate almost exclusively in the observable ECG signal space. In practice, the standard twelve-lead ECG represents multiple projections of the same underlying cardiac electrical activity from different spatial orientations. Therefore, representation learning in the ECG space inevitably introduces substantial redundancy, which may lead to spurious correlations and increased risk of overfitting. To address this and motivated by the Frank vectorcardiogram (VCG) model, we propose learning a unified latent representation of cardiac electrical activity directly in the VCG space. We introduce LVCG, the first general self-supervised representation learning framework designed to operate in this physically grounded latent space. By learning view-invariant latent VCG representations rather than lead-specific artifacts, VCG minimizes redundancy and improves generalization. LVCG generally outperforms ECG-space baselines across tasks, demonstrating enhanced robustness and generalization, especially in domain shift settings. Our code has been made available at https://anonymous.4open.science/r/LVCG-A0C4

","absKo":"

Electrocardiography(ECG)는 심장 평가의 핵심으로, 정보성 있는 ECG 표현을 학습하는 일은 질병 진단부터 임상 보고서 생성에 이르기까지 다양한 과제의 근간이 된다. 그러나 기존 방법들은 거의 전적으로 관측 가능한 ECG 신호 공간에서 동작한다. 실제로 표준 12-lead ECG는 동일한 심장 전기 활동을 서로 다른 공간적 방향에서 투영한 여러 관측치로 볼 수 있다. 따라서 ECG 공간에서의 representation learning은 필연적으로 상당한 중복성을 도입하며, 이는 허위 상관(spurious correlation)을 유발하고 overfitting 위험을 높일 수 있다. 이러한 문제를 해결하기 위해, 그리고 Frank vectorcardiogram(VCG) 모델에서 영감을 받아, 우리는 VCG 공간에서 심장 전기 활동의 통합 latent representation을 직접 학습하는 방법을 제안한다. 우리는 이 물리적으로 타당한 latent 공간에서 동작하도록 설계된 최초의 일반적 self-supervised representation learning framework인 LVCG를 소개한다. lead-specific artifact 대신 view-invariant한 latent VCG representation을 학습함으로써, VCG는 중복성을 최소화하고 generalization을 향상시킨다. LVCG는 전반적으로 다양한 과제에서 ECG-space baseline을 능가하며, 특히 domain shift 설정에서 향상된 robustness와 generalization을 보여준다. 코드는 https://anonymous.4open.science/r/LVCG-A0C4 에 공개되어 있다.

"},{"id":"63982","en":"From Text to Forecasts: Bridging Modality Gap with Temporal Evolution Semantic Space","ko":"텍스트에서 예측으로: Temporal Evolution Semantic Space로 모달리티 간 격차를 메우기","authors":"Lehui Li, Yuyao Wang, Jisheng Yan, Wei Zhang, Jinliang Deng, Haoliang Sun, Zhongyi Han, Yongshun Gong","pos":"#1606","link":"https://openreview.net/forum?id=S2Fd1GEyv6","abs":"

Incorporating textual information into time-series forecasting holds promise for addressing event-driven non-stationarity; however, a fundamental modality gap hinders effective fusion: textual descriptions express temporal impacts implicitly and qualitatively, whereas forecasting models rely on explicit and quantitative signals. Through controlled semi-synthetic experiments, we show that existing methods over-attend to redundant tokens and struggle to reliably translate textual semantics into usable numerical cues. To bridge this gap, we propose \\method{}, which introduces a Temporal Evolution Semantic Space as an intermediate bottleneck between modalities. This space consists of interpretable, numerically grounded temporal primitives—mean shift, volatility, shape, and lag—extracted from text by an LLM via structured prompting and filtered through confidence-aware gating. Experiments on four real-world datasets demonstrate up to a 29\\% reduction in forecasting error compared to state-of-the-art uni-modal and multimodal baselines. The code is available at https://anonymous.4open.science/r/MMTSF.

","absKo":"

텍스트 정보를 time-series forecasting에 통합하는 것은 event-driven non-stationarity를 다루는 데 유망하지만, 근본적인 modality gap이 효과적인 fusion을 가로막는다. 텍스트 설명은 temporal impact를 암시적이고 정성적으로 표현하는 반면, forecasting model은 명시적이고 정량적인 signal에 의존하기 때문이다. 통제된 semi-synthetic 실험을 통해, 기존 방법들이 중복 token에 과도하게 주의를 기울이고 텍스트 의미를 실제로 사용할 수 있는 수치적 단서로 안정적으로 변환하는 데 어려움을 겪는다는 것을 보인다. 이 간극을 메우기 위해, 우리는 \\method{}를 제안한다. 이는 modality 사이의 intermediate bottleneck으로 Temporal Evolution Semantic Space를 도입한다. 이 공간은 LLM이 structured prompting을 통해 text에서 추출하고 confidence-aware gating으로 필터링한, 해석 가능하고 수치적으로 grounded된 temporal primitive들, 즉 mean shift, volatility, shape, lag로 구성된다. 네 개의 실제 데이터셋에서의 실험은 state-of-the-art uni-modal 및 multimodal baseline과 비교해 forecasting error를 최대 29\\%까지 줄임을 보여준다. 코드는 https://anonymous.4open.science/r/MMTSF 에서 확인할 수 있다.

"},{"id":"65637","en":"Time series saliency maps: Explaining models across multiple domains","ko":"시계열 중요도 맵: 여러 도메인에 걸친 모델 설명","authors":"Christodoulos Kechris, Jonathan Dan, David Atienza","pos":"#3203","link":"https://openreview.net/forum?id=Bd0NNopzpC","abs":"

Traditional saliency map methods, popularized in computer vision, highlight individual points (pixels) of the input that contribute the most to the model's output. However, in time series, they offer limited insights, as semantically meaningful features are often found in other domains. We introduce Cross-domain Integrated Gradients, a generalization of Integrated Gradients. Our method enables feature attributions in any domain that can be formulated as an invertible, differentiable transformation of the time domain. Crucially, our derivation extends the original Integrated Gradients into the complex domain, enabling frequency-based attributions. We provide the necessary theoretical guarantees, namely, path independence and completeness. We validate our method via controlled experiments with mechanistic analysis, quantitative faithfulness tests, and real-world case studies. Our approach reveals interpretable, problem-specific attributions that time-domain methods cannot capture in three real-world tasks across a variety of model architectures, machine-learning tasks, and cross-domain transforms: frequency-based attribution for a regression task in wearable heart rate extraction, independent component analysis in a classification task for electroencephalography-based seizure detection, and seasonal-trend decomposition for a forecasting problem with a zero-shot time-series foundation model. We release an open-source TensorFlow/PyTorch library to enable plug-and-play cross-domain explainability for time-series models. These results demonstrate the ability of Cross-Domain Integrated Gradients to provide semantically meaningful insights into time-series models that are impossible to achieve with traditional saliency in the time domain.

","absKo":"

컴퓨터 비전에서 널리 사용되어 온 전통적인 saliency map 방법은 모델 출력에 가장 크게 기여하는 입력의 개별 점(pixels)을 강조한다. 그러나 time series에서는 의미적으로 중요한 feature가 다른 domain에 존재하는 경우가 많아, 이러한 방법이 제공하는 통찰은 제한적이다. 우리는 Integrated Gradients의 일반화인 Cross-domain Integrated Gradients를 제안한다. 이 방법은 time domain의 invertible하고 differentiable한 transformation으로 정식화할 수 있는 모든 domain에서 feature attribution을 가능하게 한다. 특히 우리의 유도는 원래의 Integrated Gradients를 complex domain으로 확장하여 frequency-based attribution을 가능하게 한다. 우리는 path independence와 completeness라는 필요한 이론적 보장을 제시한다. 또한 제어된 실험, mechanistic analysis, 정량적 faithfulness test, 실제 사례 연구를 통해 방법을 검증한다. 우리의 접근은 다양한 model architecture, machine-learning task, cross-domain transform에 걸친 세 가지 실제 과제에서 time-domain method로는 포착할 수 없는 해석 가능한 문제별 attribution을 드러낸다. 구체적으로는 wearable heart rate extraction의 regression task에서 frequency-based attribution, electroencephalography-based seizure detection의 classification task에서 independent component analysis, 그리고 zero-shot time-series foundation model을 사용한 forecasting 문제에서 seasonal-trend decomposition을 다룬다. 우리는 time-series model을 위한 plug-and-play cross-domain explainability를 가능하게 하는 open-source TensorFlow/PyTorch library를 공개한다. 이러한 결과는 Cross-Domain Integrated Gradients가 time domain의 traditional saliency로는 불가능한, time-series model에 대한 의미적으로 중요한 통찰을 제공할 수 있음을 보여준다.

"},{"id":"66052","en":"TimeSAE: Sparse Decoding for Faithful Explanations of Black-Box Time Series Models","ko":"TimeSAE: 블랙박스 시계열 모델의 충실한 설명을 위한 희소 디코딩","authors":"Khalid Oublal, Quentin Bouniot, Qi Gan, Stephan Clemencon, Zeynep Akata","pos":"#3204","link":"https://openreview.net/forum?id=7X5GcRLHhx","abs":"

As black box models and pretrained models gain traction in time series applications, understanding and explaining their predictions becomes increasingly vital, especially in high-stakes domains where interpretability and trust are essential. However, most of the existing methods involve only in-distribution explanation, and do not generalize outside the training support, which requires the learning capability of generalization. In this work, we aim to provide a framework to explain black-box models for time series data through the dual lenses of Sparse Autoencoders (SAEs) and causality. We show that many current explanation methods are sensitive to distributional shifts, limiting their effectiveness in real-world scenarios. Building on the concept of Sparse Autoencoder, we introduce TimeSAE, a framework for black-box model explanation. We conduct extensive evaluations of TimeSAE on both synthetic and real-world time series datasets, comparing it to leading baselines. The results, supported by both quantitative metrics and qualitative insights, show that TimeSAE provides more faithful and robust explanations. Our code is available in an easy-to-use library TimeSAE-Lib: https://anonymous.4open.science/w/TimeSAE-571D/

","absKo":"

black box model과 pretrained model이 time series application에서 주목받으면서, 특히 interpretability와 trust가 중요한 고위험 도메인에서 이들의 예측을 이해하고 설명하는 일은 점점 더 중요해지고 있다. 그러나 기존 방법의 대부분은 in-distribution explanation에만 국한되며, generalization 능력을 요구하는 training support 밖으로는 일반화되지 않는다. 본 연구에서는 Sparse Autoencoder(SAE)와 causality라는 두 관점에서 time series 데이터에 대한 black-box model 설명 프레임워크를 제안한다. 우리는 현재의 많은 explanation method가 distributional shift에 민감하여 실제 환경에서의 효과가 제한된다는 점을 보인다. Sparse Autoencoder 개념을 바탕으로, 우리는 black-box model explanation 프레임워크인 TimeSAE를 도입한다. 우리는 synthetic 및 real-world time series dataset 모두에서 TimeSAE를 광범위하게 평가하고, 선도적인 baseline과 비교한다. 정량적 지표와 정성적 통찰로 뒷받침된 결과는 TimeSAE가 더 faithful하고 robust한 설명을 제공함을 보여준다. 코드는 사용하기 쉬운 라이브러리 TimeSAE-Lib: https://anonymous.4open.science/w/TimeSAE-571D/ 에서 제공된다.

"},{"id":"64118","en":"FIPN: Forward Self-Organizing Interpretable Polynomial Networks for Time Series Forecasting","ko":"FIPN: 시계열 예측을 위한 순방향 자기조직화 해석 가능 다항식 네트워크","authors":"YiZhen Wang, Zheng Wang, EUN-HU KIM, Zunwei Fu","pos":"#3704","link":"https://openreview.net/forum?id=QlLPY2cWh4","abs":"

Most existing time series forecasting models are trained with backpropagation, which often brings high computational cost and limited transparency, so it can be hard to understand why a model makes a given prediction. This paper presents FIPN, a forward self-organizing interpretable polynomial network for time series forecasting. FIPN grows its architecture layer by layer and avoids backpropagation. Each neuron couples a fuzzy-rule antecedent with a Fourier-enhanced polynomial consequent: fuzzy clustering softly partitions the input space and produces interpretable rule weights for local regimes, while the consequent operates directly on the original features and uses Fourier functions to capture periodic and frequency-related structure. Forward growth can lead to redundancy, collinearity, and overfitting as depth increases, so FIPN introduces regularized node scoring, node-level dropout, and persistent access to raw inputs at every layer to stabilize closed-form estimation and improve generalization. Experiments on long-horizon forecasting benchmarks show that FIPN achieves competitive accuracy with a compact model size, and the learned fuzzy rules provide consistent, structure-based explanations. These results suggest that forward self-organizing polynomial networks offer a practical balance among accuracy, efficiency, and interpretability for long-term time series forecasting.

","absKo":"

기존의 대부분의 time series forecasting 모델은 backpropagation으로 학습되는데, 이는 종종 높은 계산 비용과 제한된 투명성을 초래하여 모델이 왜 특정 예측을 내렸는지 이해하기 어렵게 만든다. 이 논문은 time series forecasting을 위한 forward self-organizing interpretable polynomial network인 FIPN을 제시한다. FIPN은 architecture를 layer별로 성장시키며 backpropagation을 사용하지 않는다. 각 neuron은 fuzzy-rule antecedent와 Fourier-enhanced polynomial consequent를 결합한다. 즉, fuzzy clustering은 input space를 soft하게 분할하고 local regime에 대한 해석 가능한 rule weight를 생성하는 반면, consequent는 원본 feature에 직접 작동하며 Fourier function을 사용해 periodic 및 frequency-related structure를 포착한다. Forward growth는 depth가 증가할수록 redundancy, collinearity, overfitting을 초래할 수 있으므로, FIPN은 정규화된 node scoring, node-level dropout, 그리고 모든 layer에서 raw input에 대한 지속적인 접근을 도입하여 closed-form estimation을 안정화하고 generalization을 개선한다. long-horizon forecasting benchmark에서의 실험은 FIPN이 compact한 model size로 경쟁력 있는 정확도를 달성하며, 학습된 fuzzy rule이 일관된 structure-based explanation을 제공함을 보여준다. 이러한 결과는 forward self-organizing polynomial network가 장기 time series forecasting에서 정확도, 효율성, 해석가능성 사이의 실용적인 균형을 제공함을 시사한다.

"},{"id":"60982","en":"From geometry to dynamics: Learning overdamped Langevin dynamics from sparse observations with geometric constraints","ko":"Geometry에서 Dynamics로: Geometric Constraint가 있는 Sparse Observation으로부터 Overdamped Langevin Dynamics 학습","authors":"Dimitra Maoutsa","pos":"#3912","link":"https://openreview.net/forum?id=usJwxV4CQh","abs":"

How can we learn the laws underlying the dynamics of stochastic systems when their trajectories are sampled sparsely in time? Existing methods either require temporally resolved high-frequency observations, or rely on geometric arguments that apply only to conservative systems, limiting the range of dynamics they can recover. Here, we present a new framework that reconciles these two perspectives by reformulating inference as a stochastic control problem. Our method uses geometry-driven path augmentation, guided by structure in the system’s invariant density to reconstruct likely trajectories and infer the underlying dynamics without assuming specific parametric models. Applied to overdamped Langevin systems, our approach accurately recovers stochastic dynamics even from severely undersampled data, outperforming existing methods in synthetic benchmarks. This work demonstrates the effectiveness of incorporating geometric inductive biases into stochastic system identification methods, with broad applications across physics, biology, and control.

","absKo":"

stochastic system의 trajectory가 시간적으로 희소하게 샘플링될 때, 그 동역학을 지배하는 법칙을 어떻게 학습할 수 있을까? 기존 방법은 시간적으로 해상된 고주파 관측을 요구하거나, 보존계에만 적용되는 기하학적 논리에 의존하여 복원할 수 있는 동역학의 범위를 제한한다. 여기서는 inference를 stochastic control problem으로 재정식화하여 이 두 관점을 조화시키는 새로운 framework를 제시한다. 우리의 방법은 시스템의 invariant density에 존재하는 구조에 의해 안내되는 geometry-driven path augmentation을 사용하여, 특정 parametric model을 가정하지 않고도 가장 가능성 높은 trajectory를 재구성하고 underlying dynamics를 추론한다. overdamped Langevin system에 적용했을 때, 우리의 접근법은 심하게 undersampled된 데이터에서도 stochastic dynamics를 정확하게 복원하며 synthetic benchmark에서 기존 방법을 능가한다. 이 연구는 stochastic system identification method에 geometric inductive bias를 통합하는 것이 효과적임을 보여 주며, physics, biology, control 전반에 걸쳐 폭넓은 응용 가능성을 가진다.

"},{"id":"60845","en":"DualTimesField: Rethinking Time Series as Continuous-Time Trends and Events","ko":"DualTimesField: 시계열을 연속시간 추세와 이벤트로 다시 생각하기","authors":"Wencheng Zhang, Long Li, Huayi Qin, Zongjuan Wu, Jing Li, Wanghu Chen","pos":"#3913","link":"https://openreview.net/forum?id=w6pabBPG7D","abs":"

Effective time series representation is critical for revealing temporal dynamics in many fields. However, existing approaches encounter fundamental limitations. Discrete-time representations struggle with irregular sampling and the tradeoff of fidelity and efficiency, while traditional implicit neural representations suffer from spectral bias and frequency entanglement. To address these challenges, we conceptualize time series as the superposition of continuous trends and discrete events from a continuous-time perspective and propose DualTimesField, a framework that utilizes dual implicit neural fields. Its Continuous Time Field captures smooth trends through bandwidth-limited parameterization, while a Discrete Geometric Field models transient events using learnable Gabor atoms, gated sparsity, and coarse-to-fine scale annealing. This explicit field separation effectively overcomes both limitations. Experiments on nine real-world benchmarks demonstrate substantial improvements in representation fidelity, achieving 51.2% average MSE reduction over discrete-time baselines and competitive interpolation on irregular data. Code is available at https://anonymous.4open.science/r/DualTimesField-AF32.

","absKo":"

효과적인 time series representation은 여러 분야에서 temporal dynamics를 드러내는 데 핵심적이다. 그러나 기존 접근법에는 근본적인 한계가 있다. discrete-time representation은 불규칙한 sampling과 fidelity-효율성의 tradeoff를 처리하는 데 어려움을 겪는 반면, 전통적인 implicit neural representation은 spectral bias와 frequency entanglement의 문제를 가진다. 이러한 도전을 해결하기 위해 우리는 time series를 continuous-time 관점에서 continuous trend와 discrete event의 중첩으로 개념화하고, dual implicit neural field를 활용하는 DualTimesField를 제안한다. Continuous Time Field는 bandwidth-limited parameterization을 통해 부드러운 trend를 포착하고, Discrete Geometric Field는 학습 가능한 Gabor atom, gated sparsity, coarse-to-fine scale annealing을 사용하여 일시적 event를 모델링한다. 이러한 명시적 field 분리는 두 한계를 효과적으로 극복한다. 아홉 개의 실제 benchmark 실험은 representation fidelity의 상당한 향상을 보여 주며, discrete-time baseline 대비 평균 MSE를 51.2% 감소시키고 불규칙 데이터에서 경쟁력 있는 interpolation을 달성한다. 코드: https://anonymous.4open.science/r/DualTimesField-AF32.

"},{"id":"64395","en":"Time Series Reasoning via Process-Verifiable Thinking Data Synthesis and Scheduling for Tailored LLM Reasoning","ko":"맞춤형 LLM 추론을 위한 과정 검증 가능 사고 데이터 합성 및 스케줄링을 통한 시계열 추론","authors":"Jiahui Zhou, Dan Li, Boxin Li, Xiao Zhang, Erli Meng, Lin Li, Zhuomin Chen, Jian Lou, See-Kiong Ng","pos":"#401","link":"https://openreview.net/forum?id=O3tDdAGKkj","abs":"

Time series is a pervasive data type across various application domains, rendering the reasonable solving of diverse time series tasks a long-standing goal. Recent advances in large language models (LLMs), especially their reasoning abilities unlocked through reinforcement learning (RL), have opened new opportunities for tackling tasks with long Chain-of-Thought (CoT) reasoning. However, leveraging LLM reasoning for time series remains infancy, hindered by the absence of carefully curated time series CoT data for training, limited data efficiency caused by underexplored data scheduling, and the lack of RL algorithms tailored for exploiting such time series CoT data. In this paper, we introduce VeriTime, a framework that tailors LLMs for time series reasoning through data synthesis, data scheduling, and RL training. First, we propose a data synthesis pipeline that constructs a time series–text multimodal dataset with process-verifiable annotations. Second, we design a data scheduling mechanism that arranges training samples according to a principled hierarchy of difficulty and task taxonomy. Third, we develop a two-stage reinforcement finetuning featuring fine-grained, multi-objective rewards that leverages verifiable process-level CoT data. Extensive experiments show that VeriTime substantially boosts LLM performance across diverse time series reasoning tasks. Notably, it enables compact 3B–4B models to achieve reasoning capabilities on par with or exceeding those of larger proprietary LLMs.

","absKo":"시계열은 다양한 응용 도메인 전반에 걸쳐 널리 존재하는 데이터 형태이며, 여러 시계열 태스크를 합리적으로 해결하는 것은 오래된 목표이다. 최근 large language models (LLMs), 특히 reinforcement learning (RL)을 통해 활성화된 추론 능력은 긴 Chain-of-Thought (CoT) 추론이 필요한 태스크를 다룰 새로운 기회를 열었다. 그러나 시계열에 LLM 추론을 활용하는 연구는 아직 초기 단계에 머물러 있으며, 학습을 위한 신중하게 정제된 시계열 CoT 데이터의 부재, 미탐색된 data scheduling으로 인한 제한된 데이터 효율성, 그리고 이러한 시계열 CoT 데이터를 활용하도록 설계된 RL 알고리즘의 부재가 걸림돌로 작용한다. 본 논문에서는 데이터 합성, data scheduling, RL training을 통해 LLM을 시계열 추론에 맞게 조정하는 VeriTime을 제안한다. 첫째, process-verifiable annotation을 포함한 시계열-텍스트 multimodal dataset을 구성하는 data synthesis pipeline을 제안한다. 둘째, 학습 샘플을 원칙적인 난이도 계층과 task taxonomy에 따라 배치하는 data scheduling 메커니즘을 설계한다. 셋째, 검증 가능한 process-level CoT 데이터를 활용하는 세밀한 multi-objective reward를 특징으로 하는 2단계 reinforcement finetuning을 개발한다. 광범위한 실험은 VeriTime이 다양한 시계열 추론 태스크 전반에서 LLM 성능을 크게 향상시킴을 보여준다. 특히, 3B~4B 규모의 compact model이 더 큰 proprietary LLM에 필적하거나 이를 능가하는 추론 능력을 달성하도록 한다.

"},{"id":"61541","en":"TimeSeed: Effective Time Series Forecasting with Sparse Endogenous Variables","ko":"TimeSeed: 희소 내생 변수 기반 효과적인 Time Series Forecasting","authors":"Zhaowang Wu, Kaixin Deng, Hua Yan","pos":"#402","link":"https://openreview.net/forum?id=p6VlvdOky9","abs":"

Time series forecasting has long relied on dense endogenous observations, yet in many real-world scenarios, such data is scarce or even absent. Existing approaches attempt to compensate with exogenous variables, but their reliance on incomplete endogenous histories makes them brittle under data scarcity. In this work, we introduce sparse endogenous forecasting as a new setting, where exogenous sequences and only sparse endogenous observations are available. To tackle this problem, we propose TimeSeed, a lightweight architecture that redefines sparse forecasting as a context reconstruction task. By jointly exploiting the stability of exogenous sequences and the limited but informative endogenous signals, TimeSeed reconstructs robust historical representations and transforms forecasting into a tractable sequence-based prediction problem. Remarkably, TimeSeed achieves this with a purely linear architecture using only 0.19M parameters, consistently outperforming state-of-the-art deep models on seven real-world benchmarks, with an average improvement of 13.01\\% in MSE and 7.54\\% in MAE. These results establish sparse endogenous forecasting as a practical and promising paradigm, opening a new direction for time series analysis under extreme data scarcity. Code is available at this repository: \\url{https://anonymous.4open.science/r/Alistair-7}.

","absKo":"

Time series forecasting은 오랫동안 dense한 endogenous observation에 의존해 왔지만, 실제 많은 상황에서는 이러한 data가 부족하거나 아예 존재하지 않는다. 기존 접근은 exogenous variable로 이를 보완하려 하지만, 불완전한 endogenous history에 의존하기 때문에 data scarcity 하에서 취약하다. 본 연구에서는 exogenous sequence와 sparse endogenous observation만 이용 가능한 새로운 setting인 sparse endogenous forecasting을 제안한다. 이 문제를 해결하기 위해, 우리는 sparse forecasting을 context reconstruction task로 재정의하는 경량 architecture인 TimeSeed를 제안한다. TimeSeed는 exogenous sequence의 안정성과 제한적이지만 정보가 풍부한 endogenous signal을 jointly 활용하여 robust한 historical representation을 복원하고, forecasting을 다루기 쉬운 sequence-based prediction 문제로 변환한다. 놀랍게도 TimeSeed는 단 0.19M parameter만 사용하는 순수 linear architecture로 이를 달성하며, 7개의 real-world benchmark에서 state-of-the-art deep model을 일관되게 능가하여 MSE는 평균 13.01\\%, MAE는 7.54\\% 개선한다. 이 결과는 sparse endogenous forecasting을 실용적이고 유망한 paradigm으로 확립하며, 극심한 data scarcity 하에서 time series analysis의 새로운 방향을 연다. 코드는 이 repository에서 확인할 수 있다: \\url{https://anonymous.4open.science/r/Alistair-7}.

"},{"id":"65921","en":"What If We Let Forecasting Forget? A Sparse Bottleneck for Cross-Variable Dependencies","ko":"예측이 잊게 한다면? 교차 변수 의존성을 위한 희소 병목","authors":"Fan Zhang, Shiming Fan, Hua Wang","pos":"#403","link":"https://openreview.net/forum?id=8tv1WMj3X6","abs":"

Multivariate time series forecasting is critical in many real-world systems, and thus modeling cross-channel dependencies is essential. Although existing methods improve overall accuracy by enhancing representations and cross-channel interactions, it remains challenging to reliably capture inter-variable dependencies under specific conditions. We observe that dependencies in real data are often state-dependent and noisy; in such cases, dense interactions can amplify spurious correlations and lead to representation over-smoothing, which may yield unreliable predictions in certain scenarios. Motivated by this, we propose MS-FLOW, a sparse-bottleneck framework that explicitly models inter-variable interaction as capacity-limited information flow. Specifically, MS-FLOW replaces fully connected communication with selective sparse routing, retaining only a few critical dependency paths and injecting cross-variable signals under a strict communication budget, thereby suppressing redundant connections and spurious-correlation propagation. Extensive experiments demonstrate that MS-FLOW learns more reliable multivariate correlations, achieving state-of-the-art forecasting accuracy on 12 real-world benchmarks while producing fewer yet more reliable dependencies, shifting multivariate forecasting from “more interaction” to “more effective interaction”.

","absKo":"

Multivariate time series forecasting은 많은 실제 시스템에서 핵심적이며, 따라서 cross-channel dependency를 모델링하는 것이 필수적이다. 기존 방법들은 representation과 cross-channel interaction을 강화하여 전체 정확도를 향상시키지만, 특정 조건 하에서 변수 간 dependency를 신뢰성 있게 포착하는 것은 여전히 어렵다. 우리는 실제 데이터에서의 dependency가 종종 state-dependent이고 noisy하다는 점을 관찰한다. 이러한 경우 dense interaction은 spurious correlation을 증폭시키고 representation over-smoothing을 유발하여, 일부 상황에서 신뢰할 수 없는 예측으로 이어질 수 있다. 이를 바탕으로, 우리는 inter-variable interaction을 capacity-limited information flow로 명시적으로 모델링하는 sparse-bottleneck framework인 MS-FLOW를 제안한다. 구체적으로 MS-FLOW는 fully connected communication을 selective sparse routing으로 대체하여 몇 개의 핵심 dependency path만 보존하고, 엄격한 communication budget 하에서 cross-variable signal을 주입함으로써 redundant connection과 spurious-correlation propagation을 억제한다. 광범위한 실험은 MS-FLOW가 더 신뢰할 수 있는 multivariate correlation을 학습함을 보여주며, 12개의 실제 benchmark에서 state-of-the-art forecasting accuracy를 달성하는 동시에 더 적지만 더 신뢰할 수 있는 dependency를 생성하여, multivariate forecasting을 “더 많은 interaction”에서 “더 효과적인 interaction”으로 전환한다.

"},{"id":"65690","en":"What if Tomorrow is the World Cup Final? Counterfactual Time Series Forecasting with Textual Conditions","ko":"내일이 월드컵 결승이라면? 텍스트 조건을 활용한 반사실적 시계열 예측","authors":"Shuqi Gu, Yongxiang Zhao, Baoyu Jing, Kan Ren","pos":"#404","link":"https://openreview.net/forum?id=B8bAYc4COx","abs":"

Time series forecasting has become increasingly critical in real-world scenarios, where future sequences are influenced not only by historical patterns but also by forthcoming events. In this context, forecasting must dynamically adapt to complex and stochastic future conditions, which introduces fundamental challenges in both forecasting and evaluation. Traditional methods typically rely on historical data or factual future conditions, while overlooking counterfactual scenarios. Furthermore, many existing approaches are restricted to simple structured conditions, limiting their ability to generalize to the real-world complexities. To address these gaps, we introduce the task of counterfactual time series forecasting with textual conditions, enabling more flexible and condition-aware forecasting. We propose a comprehensive evaluation framework that encompasses both factual and counterfactual settings, even in the absence of ground truth time series. Additionally, we present a novel text-attribution mechanism that distinguishes mutable from immutable factors, thereby improving forecast accuracy under sophisticated and stochastic textual conditions.

","absKo":"

time series forecasting은 미래 sequence가 과거 패턴뿐 아니라 앞으로 발생할 사건의 영향도 받는 현실 세계 시나리오에서 점점 더 중요해지고 있다. 이러한 맥락에서 forecasting은 복잡하고 stochastic한 미래 조건에 동적으로 적응해야 하며, 이는 forecasting과 evaluation 모두에 근본적인 도전 과제를 제기한다. 전통적인 방법은 일반적으로 과거 데이터 또는 사실적인(factual) 미래 조건에 의존하는 반면, counterfactual 시나리오는 간과한다. 또한 기존 접근법의 상당수는 단순한 structured condition에 제한되어 있어, 현실 세계의 복잡성으로 일반화하는 능력이 제한된다. 이러한 간극을 해결하기 위해 우리는 text condition을 갖는 counterfactual time series forecasting task를 도입하여 더 유연하고 condition-aware한 forecasting을 가능하게 한다. 우리는 ground truth time series가 없더라도 factual 및 counterfactual setting을 모두 포괄하는 포괄적 evaluation framework를 제안한다. 추가로, 우리는 mutable factor와 immutable factor를 구분하는 새로운 text-attribution mechanism을 제시하여, 정교하고 stochastic한 text condition 하에서 forecast accuracy를 향상시킨다.

"},{"id":"66488","en":"TFRBench: A Reasoning Benchmark for Evaluating Forecasting Systems","ko":"TFRBench: 예측 시스템 평가를 위한 추론 벤치마크","authors":"Md Atik Ahamed, Mihir Parmar, Palash Goyal, Yiwen Song, Long T. Le, Qiang (Shaun) Cheng, Chun-Liang Li, Hamid Palangi, Jinsung Yoon, Tomas Pfister","pos":"#500","link":"https://openreview.net/forum?id=3ExTD9F0u1","abs":"We introduce TFRBench, the first benchmark designed to evaluate the reasoning capabilities of forecasting systems. Traditionally, time-series forecasting has been evaluated solely on numerical accuracy, treating foundation models as \"black boxes.\" Unlike existing benchmarks, TFRBench provides a protocol for evaluating the reasoning generated by forecasting systems--specifically their analysis of cross-channel dependencies, trends, and external events. To enable this, we propose a systematic multi-agent framework that utilizes an iterative verification loop to synthesize numerically grounded reasoning traces. Spanning ten datasets across five domains, our evaluation confirms that this reasoning is causally effective; useful for evaluation; and prompting LLMs with our generated traces significantly improves forecasting accuracy compared to direct numerical prediction (e.g., avg. $\\sim40.2$% $\\rightarrow$ $\\sim56.6$%), validating the quality of our reasoning. Conversely, benchmarking experiments reveal that off-the-shelf LLMs consistently struggle with both reasoning (lower LLM-as-a-Judge scores) and numerical forecasting, frequently failing to capture domain-specific dynamics. TFRBench thus establishes a new standard for interpretable, reasoning-based evaluation in time-series forecasting.","absKo":"우리는 forecasting system의 reasoning capability를 평가하기 위해 설계된 최초의 benchmark인 TFRBench를 소개한다. 전통적으로 time-series forecasting은 수치 정확도만으로 평가되어 왔으며, foundation model을 \"black box\"로 간주해 왔다. 기존 benchmark와 달리, TFRBench는 forecasting system이 생성한 reasoning, 특히 cross-channel dependency, trend, external event에 대한 분석을 평가하기 위한 protocol을 제공한다. 이를 위해 우리는 반복적 verification loop를 활용하여 수치적으로 근거가 있는 reasoning trace를 합성하는 체계적인 multi-agent framework를 제안한다. 5개 도메인에 걸친 10개 dataset을 포함하는 평가에서, 우리의 결과는 이러한 reasoning이 인과적으로 효과적이며 평가에 유용하고, 우리가 생성한 trace로 LLM에 prompting했을 때 direct numerical prediction보다 forecasting accuracy가 크게 향상됨을 보인다(예: avg. ~$\\sim40.2$% $\\rightarrow$ ~$\\sim56.6$%), 이는 우리의 reasoning quality를 검증한다. 반대로 benchmark 실험은 off-the-shelf LLM이 reasoning(낮은 LLM-as-a-Judge score)과 numerical forecasting 모두에서 일관되게 어려움을 겪으며, 도메인 특화 동역학을 포착하는 데 자주 실패함을 드러낸다. 따라서 TFRBench는 time-series forecasting에서 interpretable한 reasoning-based evaluation의 새로운 표준을 확립한다."},{"id":"63437","en":"TF-FACE: Time-Frequency Fusion Learning via Frequency-Domain Adaptive and Controllable Enhancement for Trajectory Prediction","ko":"TF-FACE: 궤적 예측을 위한 주파수 영역 적응형 제어 가능 향상을 통한 시간-주파수 융합 학습","authors":"Dongjian Song, Yunhao Meng, Songjun Huang, Jiayi Han","pos":"#501","link":"https://openreview.net/forum?id=XCqnmajjW5","abs":"

Accurately predicting the future trajectories of traffic participants is critical for safe, efficient, and human-friendly autonomous driving. Existing learning-based trajectory prediction methods are predominantly time-domain and insufficiently exploit latent frequency information, which limits their capability to capture low-frequency long-term dependencies and high-frequency short-term dynamics. To address this, we propose TF-FACE, a Time-Frequency learning framework with Frequency-domain Adaptive and Controllable Enhancement. TF-FACE introduces a fusion encoder with learnable gated frequency-domain attention that adaptively manipulates band-specific features for trajectory prediction. Building on the fused representation, we design a dual-stage decoder and a band-specific time–frequency dual-consistency loss to enable controllable decoupling and coupling across long- and short-term temporal scales, global and local scales, and then generate final multimodal predictions. Experiments on Argoverse 1 demonstrate that TF-FACE achieves state-of-the-art accuracy, while maintaining real-time inference for autonomous driving. Additional experiments are conducted on Argoverse 2, further validating the TF-FACE's performance and generalizability.

","absKo":"

교통 참여자의 미래 궤적을 정확히 예측하는 것은 안전하고 효율적이며 인간 친화적인 자율주행에 매우 중요하다. 기존의 학습 기반 궤적 예측 방법은 대부분 시간 영역에 치우쳐 있으며, 잠재적인 주파수 정보를 충분히 활용하지 못해 저주파 장기 의존성과 고주파 단기 동역학을 포착하는 능력이 제한된다. 이를 해결하기 위해, 우리는 Frequency-domain Adaptive and Controllable Enhancement를 갖춘 Time-Frequency learning framework인 TF-FACE를 제안한다. TF-FACE는 학습 가능한 gated frequency-domain attention을 갖는 fusion encoder를 도입하여, 궤적 예측을 위해 band-specific feature를 적응적으로 조절한다. 융합된 표현을 바탕으로, 우리는 장기 및 단기 시간 규모, 전역 및 국소 규모 전반에서 제어 가능한 분리와 결합을 가능하게 하고 최종 multimodal prediction을 생성하기 위해 dual-stage decoder와 band-specific time–frequency dual-consistency loss를 설계한다. Argoverse 1에서의 실험은 TF-FACE가 state-of-the-art 정확도를 달성하면서 자율주행을 위한 실시간 추론을 유지함을 보여준다. 추가로 Argoverse 2에서도 실험을 수행하여, TF-FACE의 성능과 일반화 가능성을 추가로 검증한다.

"},{"id":"62639","en":"ReNF: Rethinking the Principles of Neural Long-Term Time Series Forecasters","ko":"ReNF: Neural 장기 시계열 예측기의 원리 재고찰","authors":"Yihang Lu, Xianwei Meng, Enhong Chen","pos":"#503","link":"https://openreview.net/forum?id=emDhH9VFqm","abs":"

Neural Forecasters (NFs) have become a cornerstone of Long-term Time Series Forecasting (LTSF). However, recent progress has been hampered by an overemphasis on architectural complexity at the expense of fundamental forecasting principles. In this work, we revisit the principles of LTSF. We begin by formulating a Variance Reduction Hypothesis (VRH), positing that generating and combining multiple forecasts is essential to reducing the inherent uncertainty of NFs. Guided by this, we propose Boosted Direct Output (BDO), a streamlined paradigm that synergistically hybridizes the causal structure of Auto-Regressive (AR) with the stability of Direct Output (DO), while implicitly realizing the principle of forecast combination within a single network. Furthermore, we address the critical validation-test generalization gap by employing parameter smoothing to stabilize optimization. Extensive experiments demonstrate that these trivial yet principled improvements enable a direct temporal MLP to outperform recent, complex state-of-the-art models in nearly all benchmarks, without relying on intricate inductive biases. Finally, we empirically verify our hypothesis, establishing a dynamic performance bound that highlights promising directions for future research. The code for review is available at: \\url{https://anonymous.4open.science/r/ReNF-A151}.

","absKo":"Neural Forecaster (NF)는 Long-term Time Series Forecasting (LTSF)의 핵심이 되었다. 그러나 최근의 진전은 근본적인 forecasting principle보다 architecture complexity에 지나치게 집중한 탓에 저해되어 왔다. 본 연구에서는 LTSF의 원리를 다시 살펴본다. 먼저 Variance Reduction Hypothesis (VRH)를 정식화하는데, 이는 여러 forecast를 생성하고 결합하는 것이 NF의 본질적인 불확실성을 줄이는 데 필수적이라는 가설이다. 이를 바탕으로 우리는 Boosted Direct Output (BDO)를 제안한다. 이는 Auto-Regressive (AR)의 causal structure와 Direct Output (DO)의 안정성을 하나의 간결한 paradigm 안에서 시너지 있게 결합하며, 단일 network 내부에서 forecast combination 원리를 암묵적으로 실현한다. 더 나아가, parameter smoothing을 사용해 optimization을 안정화함으로써 validation-test generalization gap이라는 중요한 문제를 해결한다. 대규모 실험은 이러한 사소해 보이지만 원리에 충실한 개선만으로도 direct temporal MLP가 정교한 inductive bias에 의존하지 않고도 거의 모든 benchmark에서 최근의 복잡한 state-of-the-art model을 능가할 수 있음을 보여준다. 마지막으로, 우리는 가설을 실증적으로 검증하여, 향후 연구 방향을 부각하는 동적 performance bound를 확립한다. 검토용 코드는 다음에서 확인할 수 있다: \\url{https://anonymous.4open.science/r/ReNF-A151}."},{"id":"66190","en":"Ranking Time Series using a Time Warping Ideal Point Model","ko":"Time Warping Ideal Point 모델을 이용한 시계열 순위화","authors":"Lucas Zoroddu, Pierre Humbert, Laurent Oudre","pos":"#504","link":"https://openreview.net/forum?id=5wMBxVaTbC","abs":"

Expert-annotated time series datasets often suffer from low agreement, especially in medical applications where decisions rely on subjective criteria and inconsistent thresholds. Such variability degrades annotation quality and thus limits the reliability of supervised classification models. To address this, we propose to rely on a pairwise comparison-based approach, which provides a more robust alternative to individual annotation, since relative judgments are typically easier and yield higher consistency. The problem is thus transformed into a ranking problem and we introduce an ideal point model adapted to time series data using elastic similarity measures such as Dynamic Time Warping (DTW) and Time Warp Edit Distance (TWED). We prove Lipschitz continuity of these distances and demonstrate several convergence guarantees for this model. To facilitate gradient-based optimization, we also introduce a differentiable version of the TWED. Finally, we show through multiple experiments that our approach produces accurate and robust rankings under noisy annotation conditions.

","absKo":"

전문가 주석이 달린 time series 데이터셋은 특히 의사결정이 주관적 기준과 일관되지 않은 임계값에 의존하는 의료 응용에서 낮은 합의도 문제를 자주 겪는다. 이러한 변동성은 주석 품질을 떨어뜨리고, 그 결과 supervised classification models의 신뢰성을 제한한다. 이를 해결하기 위해, 우리는 개별 주석보다 더 강건한 대안인 pairwise comparison-based 접근을 사용할 것을 제안한다. 상대적 판단은 대체로 더 쉽고 일관성도 더 높게 나타나기 때문이다. 따라서 문제를 ranking problem으로 변환하고, Dynamic Time Warping (DTW)과 Time Warp Edit Distance (TWED) 같은 elastic similarity measures를 사용하여 time series data에 맞게 조정된 ideal point model을 도입한다. 우리는 이러한 거리의 Lipschitz continuity를 증명하고, 이 모델에 대한 여러 수렴 보장을 제시한다. gradient-based optimization을 용이하게 하기 위해, TWED의 differentiable 버전도 도입한다. 마지막으로, 여러 실험을 통해 우리의 접근법이 noisy annotation 조건에서도 정확하고 강건한 ranking을 생성함을 보인다.

"},{"id":"64647","en":"Parameter Decorrelation via Transition-Variance Alignment for Multivariate Time-series Forecasting","ko":"다변량 시계열 예측을 위한 Transition-Variance Alignment 기반 Parameter Decorrelation","authors":"Ji-Eun Choi, Jae-Hong Lee, Joon Hyuk Chang","pos":"#505","link":"https://openreview.net/forum?id=LIQfBYvHei","abs":"

Multivariate time-series forecasting (MTSF) learns from high-dimensional covariates with strong temporal dependence, periodic structure, and cross-variable correlations. While modern pipelines often mitigate non-stationarity through instance-wise normalization and decomposition, these interventions operate at the data level and do not directly control dependence that can emerge among the parameters during training. We study MTSF optimization from a parameter-decorrelation viewpoint. Modeling stochastic optimization as a Markov chain in parameter space and leveraging its stochastic differential equation interpretation, we use the per-step transition-variance induced by gradient noise as a tractable signal for optimization-induced dependence and update uncertainty. This signal can empirically inflate during training; we theoretically show that such inflation can degrade generalization diagnostics. Motivated by this mechanism, we propose transition-variance alignment (TVA), an architecture-agnostic procedure that regulates transition-variance by smoothly gating the step size based on the mismatch between an estimated noise scale and a chosen target. TVA maintains effective transition-variance near a prescribed scale without architectural changes, incurs negligible overhead, and integrates seamlessly with diverse methods. Across real-world multivariate benchmarks, TVA consistently improves forecasting accuracy.

","absKo":"

Multivariate time-series forecasting (MTSF)는 강한 temporal dependence, periodic structure, 그리고 cross-variable correlation을 지닌 고차원 covariate로부터 학습한다. 현대 파이프라인은 instance-wise normalization과 decomposition을 통해 non-stationarity를 완화하는 경우가 많지만, 이러한 개입은 데이터 수준에서 이루어지며 학습 중 파라미터 사이에 발생할 수 있는 dependence를 직접 제어하지는 못한다. 우리는 parameter-decorrelation 관점에서 MTSF optimization을 연구한다. stochastic optimization을 parameter space에서의 Markov chain으로 모델링하고, 그 stochastic differential equation 해석을 활용하여, gradient noise가 유도하는 per-step transition-variance를 optimization-induced dependence와 update uncertainty를 나타내는 다루기 쉬운 신호로 사용한다. 이 신호는 경험적으로 학습 중 증폭될 수 있으며, 우리는 이 증폭이 generalization diagnostics를 악화시킬 수 있음을 이론적으로 보인다. 이러한 메커니즘에 동기부여되어, 우리는 추정된 noise scale과 선택한 target 사이의 mismatch를 기준으로 step size를 부드럽게 gating하여 transition-variance를 조절하는 architecture-agnostic 절차인 transition-variance alignment (TVA)를 제안한다. TVA는 구조 변경 없이도 허용된 scale 근처에서 effective transition-variance를 유지하고, 오버헤드는 거의 없으며, 다양한 방법과 매끄럽게 통합된다. 실제 multivariate benchmark 전반에서 TVA는 예측 정확도를 일관되게 향상시킨다.

"},{"id":"64838","en":"PULSE: Generative Phase Evolution for Non-Stationary Time Series Forecasting","ko":"PULSE: 비정상 시계열 예측을 위한 Generative Phase Evolution","authors":"Yangyou Liu, Zezhi Shao, Xinyu Chen, Hu Chen, Fei Wang, Yuankai Wu","pos":"#506","link":"https://openreview.net/forum?id=JJIqZzujgE","abs":"

Time series forecasting under non-stationarity faces a fundamental tension between capturing stable representations and adapting to distribution shifts. Existing methods implicitly rely on static historical assumptions, leading to a critical failure mode we term Phase Amnesia, where models become blind to the evolving global context. To resolve this, we formalize non-stationary dynamics through three physical hypotheses: Wold decomposition, dynamical phase evolution, and heteroscedastic manifold generation. These principles inspire PULSE, a physics-informed, plug-and-play framework adopting a Disentangle--Evolve--Simulate design philosophy. Specifically, PULSE utilizes phase-anchored disentanglement to resolve optimization interference caused by dominant trends, employs a Phase Router to actively generate future trajectories, and introduces Statistic-Aware Mixup (SAM) to ensure robustness against out-of-distribution volatility. Empirically, PULSE enables a simple MLP backbone to consistently outperform state-of-the-art Transformers across 12 real-world benchmarks. This validates that a correct physics-informed inductive bias is far more critical than raw architectural complexity for non-stationary forecasting.

","absKo":"

non-stationarity 하에서의 time series forecasting은 안정적인 representation을 포착하는 것과 distribution shift에 적응하는 것 사이의 근본적 긴장에 직면한다. 기존 방법은 암묵적으로 정적인 과거 가정에 의존하여, 모델이 진화하는 global context를 인식하지 못하게 되는 Phase Amnesia라는 치명적 실패 양상을 초래한다. 이를 해결하기 위해 우리는 Wold decomposition, dynamical phase evolution, heteroscedastic manifold generation이라는 세 가지 물리적 가설을 통해 non-stationary dynamics를 정식화한다. 이러한 원리는 Disentangle--Evolve--Simulate 설계 철학을 채택한 physics-informed plug-and-play framework인 PULSE를 낳았다. 구체적으로, PULSE는 지배적인 trend로 인한 optimization interference를 해소하기 위해 phase-anchored disentanglement를 사용하고, future trajectory를 능동적으로 생성하기 위해 Phase Router를 적용하며, out-of-distribution volatility에 대한 견고성을 보장하기 위해 Statistic-Aware Mixup (SAM)을 도입한다. 실험적으로, PULSE는 단순한 MLP backbone이 12개의 실제 benchmark에서 state-of-the-art Transformer를 일관되게 능가하도록 한다. 이는 non-stationary forecasting에서 올바른 physics-informed inductive bias가 원시적인 architecture 복잡성보다 훨씬 더 중요함을 입증한다.

"},{"id":"64459","en":"Overcoming the Modality Gap in Context-Aided Forecasting","ko":"맥락 지원 예측에서 모달리티 격차 극복","authors":"Vincent Zheng, Étienne Marcotte, Arjun Ashok, Andrew Williams, Lijun Sun, Alexandre Drouin, Valentina Zantedeschi","pos":"#507","link":"https://openreview.net/forum?id=NOHlqCnYot","abs":"

Context-aided forecasting (CAF) holds promise for integrating domain knowledge and forward-looking information, enabling AI systems to surpass traditional statistical methods. However, recent empirical studies reveal a puzzling gap: multimodal models often fail to outperform their unimodal counterparts. We hypothesize that this underperformance stems from poor context quality in existing datasets, as verification is challenging. To address these limitations, we introduce a semi-synthetic data augmentation method that generates contexts both descriptive of temporal dynamics and verifiably complementary to numerical histories. This approach enables massive-scale dataset creation, resulting in CAF-7M, a corpus of 7 million context-augmented time series windows, including a rigorously verified test set. We demonstrate that semi-synthetic pre-training transfers effectively to real-world evaluation, and show clear evidence of context utilization. Our results suggest that dataset quality, rather than architectural limitations, has been the primary bottleneck in context-aided forecasting.

","absKo":"

Context-aided forecasting (CAF)은 domain knowledge와 forward-looking information을 통합하여 AI system이 전통적인 통계적 방법을 능가할 수 있게 해 줄 유망한 접근이다. 그러나 최근의 실증 연구는 puzzling gap을 드러낸다. multimodal model이 종종 unimodal counterpart를 능가하지 못하는 것이다. 우리는 이러한 성능 저하가 기존 dataset의 context quality가 낮고 검증이 어렵기 때문이라고 가정한다. 이러한 한계를 해결하기 위해, 우리는 temporal dynamics를 잘 설명하면서도 numerical history와 검증 가능하게 보완적인 context를 생성하는 semi-synthetic data augmentation 방법을 도입한다. 이 접근은 대규모 dataset 생성을 가능하게 하며, 엄격하게 검증된 test set을 포함하는 700만 개의 context-augmented time series window로 이루어진 CAF-7M 코퍼스를 만들어 낸다. 우리는 semi-synthetic pre-training이 실제 평가로 효과적으로 전이됨을 보이고, context 활용의 명확한 증거를 제시한다. 우리의 결과는 context-aided forecasting의 주요 병목이 architecture의 한계가 아니라 dataset quality였음을 시사한다.

"},{"id":"64890","en":"MemCast: Memory-Driven Time Series Forecasting with Experience-Conditioned Reasoning","ko":"MemCast: Experience-Conditioned 추론을 활용한 Memory-Driven 시계열 예측","authors":"Xiaoyu Tao, Mingyue Cheng, Ze Guo, Shuo Yu, Yaguo Liu, Qi Liu, Shijin Wang","pos":"#508","link":"https://openreview.net/forum?id=IoY0A6opfl","abs":"

Time series forecasting (TSF) plays a critical role in decision-making for many real-world applications. Recently, LLM-based forecasters have made promising advancements. Despite their effectiveness, existing methods often lack explicit experience accumulation and continual evolution. In this work, we propose MemCast, a learning-to-memory framework that reformulates TSF as an experience-conditioned reasoning task. Specifically, we learn experience from the training set and organize it into a hierarchical memory. This is achieved by summarizing prediction results into historical patterns, distilling inference trajectories into reasoning wisdom, and inducing extracted temporal features into general laws. Furthermore, during inference, we leverage historical patterns to guide the reasoning process and utilize reasoning wisdom to select better trajectories, while general laws serve as criteria for reflective iteration. Additionally, to enable continual evolution, we design a dynamic confidence adaptation strategy that updates the confidence of individual entries without leaking the test set distribution. Extensive experiments on multiple datasets demonstrate that MemCast consistently outperforms previous methods, validating the effectiveness of our approach. Our code is available at https://anonymous.4open.science/r/MemCast.

","absKo":"

Time series forecasting (TSF)은 다양한 실제 응용에서 의사결정에 핵심적인 역할을 한다. 최근 LLM-based forecaster들이 유망한 진전을 보였지만, 기존 방법은 explicit experience accumulation과 continual evolution이 부족한 경우가 많다. 본 연구에서는 TSF를 experience-conditioned reasoning task로 재구성하는 learning-to-memory framework인 MemCast를 제안한다. 구체적으로, training set에서 experience를 학습해 이를 hierarchical memory로 조직한다. 이는 예측 결과를 historical pattern으로 요약하고, inference trajectory를 reasoning wisdom으로 정제하며, 추출된 temporal feature를 general law로 유도하는 방식으로 이루어진다. 또한 inference 시에는 historical pattern을 활용해 reasoning process를 안내하고, reasoning wisdom을 사용해 더 나은 trajectory를 선택하며, general law는 reflective iteration의 기준으로 사용한다. 추가로 continual evolution을 가능하게 하기 위해, test set distribution을 누설하지 않으면서 개별 entry의 confidence를 갱신하는 dynamic confidence adaptation strategy를 설계했다. 여러 dataset에서 수행한 광범위한 실험은 MemCast가 이전 방법들을 일관되게 능가함을 보여주며, 우리의 접근법의 효과를 검증한다. 코드: https://anonymous.4open.science/r/MemCast.

"},{"id":"66176","en":"LagLLM: LLM-empowered lead–lag dependency learning for spatial-temporal time series forecasting","ko":"LagLLM: 공간-시간 시계열 예측을 위한 LLM 기반 선행-후행 의존성 학습","authors":"Binqing Wu, Jian Zhou, Zongjiang Shang, Ling Chen","pos":"#509","link":"https://openreview.net/forum?id=64LGgWZ47X","abs":"

Spatial–temporal time series forecasting is challenging due to complex lead–lag dependencies, which are often ignored or inadequately modeled by existing methods. Thus, we propose LagLLM, the first LLM-empowered framework that explicitly models lead–lag dependencies by unifying data-driven dynamics modeling and knowledge-driven semantic reasoning. Specifically, LagLLM constructs a lead–lag graph by integrating learnable embeddings, spatial proximity, and prompt-guided reasoning from a frozen LLM, which can capture lead-lag dependencies informed by underlying data structure and semantic knowledge. In addition, LagLLM introduces structural token sorting based on the graph, which can make a fine-turned LLM explicitly perceive directional and delayed interactions. Experiments on eight real-world datasets show that LagLLM achieves the state-of-the-art performance with improved accuracy, robustness, and interpretability. The code is available at https://anonymous.4open.science/r/LagLLM.

","absKo":"

Spatial–temporal time series forecasting은 복잡한 lead–lag dependency 때문에 어렵고, 이러한 dependency는 기존 방법에서 종종 무시되거나 충분히 모델링되지 못했다. 이에 우리는 data-driven dynamics modeling과 knowledge-driven semantic reasoning을 통합하여 lead–lag dependency를 명시적으로 모델링하는, 최초의 LLM-empowered framework인 LagLLM을 제안한다. 구체적으로 LagLLM은 learnable embedding, spatial proximity, 그리고 frozen LLM으로부터의 prompt-guided reasoning을 통합하여 lead–lag graph를 구성하며, 이를 통해 underlying data structure와 semantic knowledge에 의해 뒷받침되는 lead-lag dependency를 포착할 수 있다. 추가로 LagLLM은 graph에 기반한 structural token sorting을 도입하여, fine-tuned LLM이 directional하고 delayed된 interaction을 명시적으로 인식하도록 만든다. 8개의 실제 데이터셋에 대한 실험은 LagLLM이 향상된 accuracy, robustness, interpretability를 바탕으로 state-of-the-art performance를 달성함을 보여준다. 코드는 https://anonymous.4open.science/r/LagLLM 에서 확인할 수 있다.

"},{"id":"64743","en":"KUMA: A Novel Framework with Koopman Separation and Efficient Multilevel Extraction in Time Series Forecasting","ko":"KUMA: 시계열 예측에서 Koopman Separation과 효율적 다단계 추출을 갖춘 새로운 프레임워크","authors":"Sijie Xiong, Cheng Tang, Atsushi Shimada","pos":"#510","link":"https://openreview.net/forum?id=KIJq2riCoy","abs":"

Time series forecasting plays a crucial role in a wide range of real-world applications and has become increasingly complex with the growth of multivariate dimensions and extended historical observations, leading to the prosperity of deep forecasting models. Previous models are hindered by three major challenges: high computational complexity, inefficient token utilization caused by redundancy and scarcity, and temporal distribution shifts resulting from non-stationary dynamics. Inspired by Koopman theory and the success of multilevel encoder–decoder architectures with skip connections, we design an input-dependent Koopman module to decompose time series into Koopman dynamics and residual dynamics. Building upon this formulation, we propose a U-shaped Multilevel Attention module (UMA) that integrates element-wise attention filtering and linear attention, giving rise to KUMA. The input-dependent Koopman operator mitigates the issue of operator mixture and alleviates temporal distribution shifts, while UMA achieves a favorable balance between token redundancy and token scarcity with acceptable computational efficiency. Comprehensive evaluations across 12 benchmark datasets demonstrate that KUMA achieves superior performance compared to existing excellent approaches.

","absKo":"

시계열 예측은 다양한 실제 응용에서 핵심적인 역할을 하며, 다변량 차원과 확장된 과거 관측이 증가함에 따라 점점 더 복잡해져 deep forecasting model의 발전을 이끌어 왔다. 기존 모델은 세 가지 주요 도전에 의해 제약을 받는다. 높은 계산 복잡도, 중복과 희소성으로 인해 비효율적인 token 활용, 그리고 비정상(non-stationary) 동역학에서 비롯되는 시간적 분포 이동이다. Koopman theory와 skip connection을 갖는 multilevel encoder-decoder architecture의 성공에 착안하여, 우리는 시계열을 Koopman dynamics와 residual dynamics로 분해하는 input-dependent Koopman module을 설계한다. 이러한 formulation 위에 element-wise attention filtering과 linear attention을 통합하는 U-shaped Multilevel Attention module (UMA)을 제안하며, 이를 통해 KUMA가 탄생한다. input-dependent Koopman operator는 operator mixture 문제를 완화하고 temporal distribution shift를 줄이며, UMA는 적절한 계산 효율성 하에서 token redundancy와 token scarcity 사이의 균형을 잘 맞춘다. 12개의 benchmark dataset에 대한 포괄적 평가 결과, KUMA가 기존의 우수한 접근법들보다 더 뛰어난 성능을 달성함을 보였다.

"},{"id":"66595","en":"Information Geometry Loss for Time Series Forecasting","ko":"시계열 예측을 위한 정보 기하 Loss","authors":"Jiayu Fang, Xuande Liu, Sangsha Fang, Zhen Tian, Hongwei Ma, Zhiqi Shao, Junbin Gao","pos":"#511","link":"https://openreview.net/forum?id=2NbXZDl4nB","abs":"

Time series forecasting fundamentally involves learning probability distributions over future observations. However, existing loss functions rely on point-wise Euclidean metrics, neglecting the intrinsic geometric structure of probability distributions. This leads to suboptimal alignment between predicted and true distributions, particularly for uncertainty quantification. We propose InfoGeo Loss, a principled loss function grounded in information geometry that measures distributional discrepancies on statistical manifolds. Our approach comprises three key components: (1) a distribution parameterization module that models predictions with learnable sufficient statistics, (2) a Fisher information metric that quantifies intrinsic distributional distance, and (3) a Bregman divergence component that captures asymmetric prediction errors. We further introduce a natural gradient weighting strategy for efficient optimization on statistical manifolds. Theoretically, we prove statistical consistency and establish convergence guarantees. Extensive experiments on seven datasets with five architectures show that InfoGeo Loss consistently outperforms existing losses, achieving average improvements of 6.8% in MSE and 5.3% in MAE.

","absKo":"시계열 예측은 본질적으로 미래 관측값에 대한 확률 분포를 학습하는 문제를 포함한다. 그러나 기존 loss function은 점별 Euclidean metric에 의존하여 확률 분포의 내재적 기하 구조를 간과한다. 이로 인해 특히 uncertainty quantification에서 예측된 분포와 실제 분포의 정렬이 최적 이하가 된다. 우리는 statistical manifold 위에서 분포 간 차이를 측정하는 information geometry에 기반한 원리적 loss function인 InfoGeo Loss를 제안한다. 우리의 접근은 세 가지 핵심 구성요소로 이루어진다. (1) 학습 가능한 sufficient statistic으로 예측을 모델링하는 distribution parameterization module, (2) 내재적 분포 거리(intrinsic distributional distance)를 정량화하는 Fisher information metric, (3) 비대칭 예측 오차를 포착하는 Bregman divergence component. 또한 statistical manifold에서 효율적인 최적화를 위한 natural gradient weighting 전략을 도입한다. 이론적으로 우리는 statistical consistency를 증명하고 수렴 보장을 확립한다. 다섯 가지 architecture를 사용한 일곱 개 dataset에 대한 광범위한 실험에서 InfoGeo Loss는 기존 loss를 일관되게 능가하며, MSE 평균 6.8\\%, MAE 평균 5.3\\%의 향상을 달성한다.

"},{"id":"64046","en":"Hierarchical ODE: Learning Continuous-Time Physical Prototypes for Early Link Failure Detection","ko":"계층적 ODE: 조기 링크 장애 탐지를 위한 연속시간 물리 프로토타입 학습","authors":"Jiaen Lv, Leran Qi, Shaowei Wang","pos":"#512","link":"https://openreview.net/forum?id=ROje5jwFq1","abs":"

Time series prototype learning is fundamentally challenged by observational ambiguity. Discrete architectures fail to resolve this, as they lack the capacity to decouple stochastic noise from continuous dynamics. Furthermore, rigid closed-set assumptions fail to capture unseen diversity. To address these limitations, we propose a hierarchical ordinary differential equation clustering network, which utilizes neural ordinary differential equation to model latent state evolution as a continuous integral curve. This formulation enforces temporal continuity to effectively disentangle smooth feature trends from stochastic noise, while our adaptive hierarchical mechanism autonomously identifies the optimal number of prototypes without rigid prior constraints. Validated on the early link failure detection task with irregularly sampled time series, the proposed method effectively extracts underlying physical prototypes, thereby enabling robust failure detection.

","absKo":"시계열 prototype learning은 관측의 모호성이라는 근본적인 어려움에 직면한다. discrete architecture는 stochastic noise와 continuous dynamics를 분리해 내는 능력이 부족하기 때문에 이를 해소하지 못한다. 더 나아가, 경직된 closed-set 가정은 보지 못한 다양성을 포착하지 못한다. 이러한 한계를 해결하기 위해, 우리는 neural ordinary differential equation을 활용하여 latent state evolution을 continuous integral curve로 모델링하는 hierarchical ordinary differential equation clustering network를 제안한다. 이 formulation은 시간적 연속성을 강제하여 smooth한 feature trend와 stochastic noise를 효과적으로 분리하며, adaptive hierarchical mechanism은 경직된 사전 제약 없이 최적의 prototype 수를 자율적으로 식별한다. 불규칙하게 샘플링된 time series를 사용하는 초기 link failure detection task에서 검증한 결과, 제안 방법은 underlying physical prototype를 효과적으로 추출하여 강건한 failure detection을 가능하게 한다.

"},{"id":"65473","en":"Geometry-Aware Contrastive Learning for Few-Shot Automatic Modulation Recognition","ko":"Few-Shot 자동 변조 인식을 위한 기하 인지 대조 학습","authors":"guanqun zhao, Yitong Liu, Jiaxuan Fang, Hongwen Yang, Yufei Mao","pos":"#513","link":"https://openreview.net/forum?id=DGXfoyszbj","abs":"

Standard Self-Supervised Learning (SSL) for Automatic Modulation Recognition (AMR) struggles with ineffective isotropic augmentations, spectral instability, and semantic drift. To address these challenges, we propose Dynamic-Consistency Contrastive Learning (DyCo-CL), a geometry-aware framework that couples Virtual Adversarial Augmentation (VAA) with a semantic consistency loss. We provide a theoretical analysis indicating that this strategy acts as an implicit spectral regularizer for the encoder, enabling stable manifold exploration. Complementing this, our Signal-Adaptive Swin Backbone with fixed-window attention improves structural stability by constraining attention locality, while a Hybrid Knowledge Fusion module anchors representations with physical priors. Experiments on RML benchmarks show that DyCo-CL achieves a 6.27% accuracy gain in 1-shot settings over prior methods.

","absKo":"

Automatic Modulation Recognition(AMR)을 위한 표준 Self-Supervised Learning(SSL)은 비효율적인 isotropic augmentation, spectral instability, semantic drift로 어려움을 겪는다. 이러한 문제를 해결하기 위해 우리는 Virtual Adversarial Augmentation(VAA)과 semantic consistency loss를 결합한 geometry-aware framework인 Dynamic-Consistency Contrastive Learning(DyCo-CL)을 제안한다. 우리는 이 전략이 encoder에 대한 implicit spectral regularizer로 작용하여 안정적인 manifold exploration을 가능하게 한다는 이론적 분석을 제시한다. 여기에 더해, 고정 window attention을 갖는 Signal-Adaptive Swin Backbone은 attention locality를 제한함으로써 구조적 안정성을 향상시키고, Hybrid Knowledge Fusion module은 physical prior로 표현을 고정한다. RML benchmark 실험에서 DyCo-CL은 이전 방법 대비 1-shot 설정에서 6.27%의 정확도 향상을 달성한다.

"},{"id":"62910","en":"Dynamic Relational Priming Improves Transformer in Multivariate Time Series","ko":"동적 관계 Priming이 다변량 시계열의 Transformer를 개선하는 방법","authors":"Hunjae Lee, Corey Clark","pos":"#514","link":"https://openreview.net/forum?id=bz66grGgjI","abs":"

Standard attention mechanisms in transformers employ static token representations that remain unchanged across all pair-wise computations in each layer. This limits their representational alignment with the potentially diverse dynamics of each token-pair interaction. While they excel in domains with relatively homogeneous relationships, standard attention may be inadequate in capturing heterogeneous inter-channel dependencies of multivariate time series (MTS) data where different channel-pair interactions within a single system may be governed by entirely different physical laws or temporal dynamics. To better align the attention mechanism for such domain phenomena, we propose attention with dynamic relational priming (prime attention). Prime attention modulates token representations for each token-pair, optimizing each pair-wise interaction for that specific relationship. Our results demonstrate that prime attention consistently outperforms standard attention across benchmarks, achieving up to 6.5\\% improvement in forecasting accuracy. In addition, prime attention achieves comparable performance using up to 40\\% less sequence length compared to standard attention, demonstrating its superior relational modeling capabilities and potential for data efficiency.

","absKo":"

transformer의 표준 attention mechanism은 각 layer의 모든 pair-wise computation에서 변하지 않는 static token representation을 사용한다. 이는 각 token-pair interaction이 가질 수 있는 다양한 dynamics와의 representation alignment를 제한한다. 비교적 동질적인 관계를 다루는 domain에서는 강점을 보이지만, 표준 attention은 multivariate time series(MTS) 데이터의 heterogeneous inter-channel dependency를 포착하는 데는 부족할 수 있다. MTS에서는 하나의 시스템 내 서로 다른 channel-pair interaction이 전혀 다른 physical law나 temporal dynamics에 의해 지배될 수 있기 때문이다. 이러한 domain 현상에 attention mechanism을 더 잘 맞추기 위해, 우리는 dynamic relational priming(prime attention)을 사용한 attention을 제안한다. Prime attention은 각 token-pair에 대해 token representation을 조절하여, 해당 특정 관계에 맞게 pair-wise interaction을 최적화한다. 우리의 결과는 prime attention이 benchmark 전반에서 표준 attention을 일관되게 능가하며, forecasting accuracy를 최대 6.5\\% 향상시킨다는 점을 보여준다. 또한 prime attention은 표준 attention보다 sequence length를 최대 40\\% 줄이면서도 비슷한 성능을 달성해, 더 우수한 relational modeling capability와 데이터 효율성의 가능성을 보여준다.

"},{"id":"64969","en":"AnomSeer: Reinforcing Multimodal LLMs to Reason for Time-Series Anomaly Detection","ko":"AnomSeer: 시계열 이상 탐지를 위한 멀티모달 LLM 추론 강화","authors":"Junru Zhang, Lang Feng, Haoran Shi, Xu Guo, Han Yu, Yabo Dong, Duanqing Xu","pos":"#607","link":"https://openreview.net/forum?id=IGW4mVoDSs","abs":"

Time-series anomaly detection (TSAD) with multimodal large language models (MLLMs) is an emerging area, yet a persistent challenge remains: MLLMs rely on coarse time-series heuristics but struggle with multi-dimensional, detailed reasoning, which is vital for understanding complex time-series data. We present AnomSeer to address this by reinforcing the model to ground its reasoning in precise, structural details of time series, unifying anomaly classification, localization, and explanation. At its core, an expert chain-of-thought trace is generated to provide a verifiable, fine-grained reasoning from classical analyses (e.g., statistical measures, frequency transforms). Building on this, we propose a novel time-series grounded policy optimization (TimerPO) that incorporates two additional components beyond standard reinforcement learning: a time-series grounded advantage based on optimal transport and an orthogonal projection to ensure this auxiliary granular signal does not interfere with the primary detection objective. Across diverse anomaly scenarios, AnomSeer, with Qwen2.5-VL-3B/7B-Instruct, outperforms larger commercial baselines in classification and localization accuracy, particularly on point- and frequency-driven exceptions. Moreover, it produces plausible reasoning traces that support its conclusions.

","absKo":"

Multimodal large language model (MLLM)을 활용한 time-series anomaly detection (TSAD)은 떠오르는 연구 분야이지만, 지속적인 과제가 하나 남아 있다. MLLM은 거친 time-series heuristic에는 의존하지만, 복잡한 time-series 데이터를 이해하는 데 핵심적인 다차원적이고 세부적인 reasoning에는 어려움을 겪는다는 점이다. 우리는 이를 해결하기 위해 AnomSeer를 제안한다. AnomSeer는 모델이 정확하고 구조적인 time-series detail에 근거해 추론하도록 강화하며, anomaly classification, localization, explanation을 하나로 통합한다. 핵심에는 classical analysis(예: statistical measures, frequency transforms)로부터 검증 가능한 fine-grained reasoning을 제공하는 expert chain-of-thought trace가 생성된다. 이를 바탕으로 우리는 standard reinforcement learning을 넘어 두 가지 추가 구성요소를 통합한 novel time-series grounded policy optimization (TimerPO)을 제안한다. 하나는 optimal transport에 기반한 time-series grounded advantage이고, 다른 하나는 보조적인 granular signal이 primary detection objective와 간섭하지 않도록 보장하는 orthogonal projection이다. 다양한 anomaly scenario에서 Qwen2.5-VL-3B/7B-Instruct를 사용한 AnomSeer는 classification과 localization accuracy 모두에서 더 큰 commercial baseline을 능가하며, 특히 point- 및 frequency-driven exception에서 두드러진다. 나아가, 그 결론을 뒷받침하는 그럴듯한 reasoning trace도 생성한다.

"},{"id":"63442","en":"Beyond Extrapolation: Knowledge Utilization Paradigm with Bidirectional Inspiration for Time Series Forecasting","ko":"외삽을 넘어서: 시계열 예측을 위한 양방향 영감 기반 지식 활용 패러다임","authors":"Liu Chong, Yingjie Zhou, Hao Li, Pengyang Wang, Qingsong Wen, Ce Zhu","pos":"#608","link":"https://openreview.net/forum?id=XAyI76nJRt","abs":"

Time-series forecasting is critical in various scenarios, such as energy, transportation, and public health. However, most existing forecasters rely primarily on one-way inference, \\textit{i.e.}, mapping \\textbf{history} to \\textbf{target}, and overlook the structural information provided by a revised natural chain (``\\textbf{history} (model input) -- \\textbf{target} (ground-truth output) -- \\textbf{post-target continuation}''). The post-target continuation records how trajectories evolve after the target, which can help stabilize forecasting, but it is not observable at inference time. In this work, we aim to obtain an approximate proxy of the post-target continuation for the current input, providing structural knowledge for bidirectional forecasting. This idea is instantiated as KUP-BI (Knowledge Utilization Paradigm with Bidirectional Inspiration), a new time-series modeling paradigm that distills continuation-style knowledge (as an approximate post-target continuation proxy) from a \\emph{train-only} historical library and integrates it into standard forecasting backbones. The input stream and the continuation-proxy stream are fused via a lightweight feature-level gating module. This design does not introduce information beyond what is already contained in the training trajectories; instead, it provides a structured inductive bias that helps backbones exploit typical continuation patterns rather than relying solely on parametric extrapolation. Experimental results on six public datasets show that KUP-BI consistently improves the forecasting performance of state-of-the-art models, with small additional overhead.

","absKo":"

time-series forecasting은 energy, transportation, public health 등 다양한 상황에서 중요하다. 그러나 기존의 대부분 forecaster는 주로 단방향 추론, \\textit{i.e.}, \\textbf{history}를 \\textbf{target}으로 매핑하는 방식에 의존하며, 수정된 자연 연쇄 ``\\textbf{history} (model input) -- \\textbf{target} (ground-truth output) -- \\textbf{post-target continuation}''이 제공하는 구조적 정보를 간과한다. post-target continuation은 target 이후 trajectory가 어떻게 진화하는지를 기록하며, 예측을 안정화하는 데 도움이 될 수 있지만 inference 시점에는 관측할 수 없다. 본 연구에서는 현재 input에 대한 post-target continuation의 근사 proxy를 얻어 bidirectional forecasting을 위한 구조적 지식을 제공하고자 한다. 이 아이디어는 KUP-BI(Knowledge Utilization Paradigm with Bidirectional Inspiration)로 구체화되며, 이는 \\emph{train-only} historical library에서 continuation-style knowledge(근사 post-target continuation proxy)를 distill하여 standard forecasting backbone에 통합하는 새로운 time-series modeling paradigm이다. input stream과 continuation-proxy stream은 lightweight feature-level gating module을 통해 fusion된다. 이 설계는 이미 학습 trajectory에 포함된 정보를 넘어서는 추가 정보를 도입하지 않으며, 대신 backbone이 parametric extrapolation에만 의존하지 않고 전형적인 continuation pattern을 활용하도록 돕는 구조화된 inductive bias를 제공한다. 6개의 공개 데이터셋에서의 실험 결과는 KUP-BI가 작은 추가 오버헤드로 state-of-the-art model의 forecasting 성능을 일관되게 향상시킴을 보여준다.

"},{"id":"66155","en":"BioFormer: Rethinking Cross-Subject Generalization via Spectral Structural Alignment in Biomedical Time-Series","ko":"BioFormer: 생의학 시계열에서 스펙트럼-구조 정렬을 통한 교차 피험자 일반화 재고","authors":"Guikang Du, Haoran Li, Xinyu Liu, Zhibo Zhang, Xiaoli Gong, Jin Zhang","pos":"#609","link":"https://openreview.net/forum?id=6I7yMzQBts","abs":"

Cross-subject generalization in biomedical time-series (BTS) refers to training on data from some subjects and testing on unseen subjects. The key challenge is to suppress subject-specific variability in BTS representations. Most existing methods implicitly suppress the variability through model building or subject adversarial learning, but rarely model it explicitly. We introduce \\textit{\\textbf{spectral drift}} as a new perspective to characterize subject-specific variability. Specifically, BTS signals under the same label often share consistent oscillatory structure, yet exhibit subject-dependent magnitude or phase shifts in specific frequency components, which we interpret as subject-specific variability. Building on this insight, we propose \\textbf{BioFormer}. At its core is a Frequency-Band Alignment Module (FBAM) that generates band-wise modulation factors from the spectral distribution and adaptively adjusts amplitude and phase to align spectral structure, thereby mitigating variability. We further pair FBAM with Sample-Conditional Layer Normalization, which infers normalization parameters from intrinsic signal statistics rather than subject identity, stabilizing cross-subject representations. Extensive experiments on six datasets demonstrate that BioFormer outperforms 12 baselines, yielding absolute F1-score improvements of 6\\%.

","absKo":"

biomedical time-series (BTS)에서의 cross-subject generalization은 일부 subject의 데이터로 학습하고 보지 못한 subject에서 테스트하는 것을 의미한다. 핵심 과제는 BTS representation에서 subject-specific variability를 억제하는 것이다. 기존 방법들은 대부분 model building이나 subject adversarial learning을 통해 이러한 variability를 암묵적으로 억제하지만, 이를 명시적으로 모델링하는 경우는 드물다. 우리는 subject-specific variability를 특징짓는 새로운 관점으로 spectral drift를 도입한다. 구체적으로, 동일한 label 아래의 BTS signal은 종종 일관된 oscillatory structure를 공유하지만, 특정 frequency component에서는 subject에 따라 magnitude나 phase shift가 달라지며, 우리는 이를 subject-specific variability로 해석한다. 이 통찰을 바탕으로 우리는 BioFormer를 제안한다. 그 핵심은 spectral distribution으로부터 band-wise modulation factor를 생성하고 amplitude와 phase를 적응적으로 조정하여 spectral structure를 정렬함으로써 variability를 완화하는 Frequency-Band Alignment Module (FBAM)이다. 또한 우리는 FBAM을 Sample-Conditional Layer Normalization과 결합하여, subject identity가 아니라 신호 고유의 통계로부터 normalization parameter를 추론함으로써 cross-subject representation을 안정화한다. 여섯 개 dataset에 대한 광범위한 실험은 BioFormer가 12개의 baseline을 능가하며, absolute F1-score를 6\\% 향상시킴을 보여준다.

"},{"id":"65944","en":"Bootstrapped Exploration with Causal Reasoning: A Training Paradigm for Adaptive Forecasting Agent","ko":"인과 추론을 활용한 부트스트랩 탐색: 적응형 예측 에이전트를 위한 학습 패러다임","authors":"Qingwen Zeng, Dajun Guo, Zhaoge Bi, lining chen, Jushang Qiu, Yitian Yang, Carl Yang, Huaming Chen, Ling Chen","pos":"#610","link":"https://openreview.net/forum?id=8d2LLxwU1r","abs":"

Time series forecasting is critical in domains such as finance, energy, and healthcare, yet real-world datasets often exhibit non-stationarity, noise, missing values, and distribution shifts, posing severe challenges for generalization. In practice, industry solutions typically rely on customized forecasting frameworks that combine imputation, decomposition, and specialized models. However, such frameworks are costly to engineer and maintain. Moreover, we observe that many frameworks suffer from the impacts of distribution shifts, which degrade their respective performance. It motivates a paradigm that transfers reliably across heterogeneous datasets while accumulating reusable strategy knowledge for large-scale, dynamic environments. Although large language model-based agents have recently shown strong reasoning and tool-use capabilities, existing approaches do not consistently adapt forecasting workflows across diverse time series. We identify two primary factors, including limited strategy-level supervision and the inherent complexity of mapping dataset-specific meta-features to effective forecasting strategies. To address these challenges, we propose BECRA, a novel agent training paradigm that learns forecasting intelligence through contrast-aware exploration and agent-level causal lesson extraction, without human-annotated supervision. BECRA distills symbolic strategy lessons that support in-context planning on unseen datasets, enabling zero-shot training adaptation.

","absKo":"

Time series forecasting은 finance, energy, healthcare 같은 분야에서 매우 중요하지만, 실제 데이터셋은 비정상성, 잡음, 결측값, distribution shift를 자주 보이므로 일반화에 심각한 어려움을 초래합니다. 실무에서는 산업용 솔루션이 보통 imputation, decomposition, specialized model을 결합한 맞춤형 forecasting framework에 의존합니다. 그러나 이러한 framework는 설계와 유지보수 비용이 큽니다. 더 나아가, 많은 framework가 distribution shift의 영향에 취약해 각자의 성능이 저하되는 현상을 관찰했습니다. 이는 이질적인 데이터셋 전반에 안정적으로 전이되면서, 대규모 동적 환경을 위한 재사용 가능한 전략 지식을 축적하는 패러다임의 필요성을 시사합니다. 최근 large language model 기반 agent가 강한 reasoning과 tool-use 능력을 보여주고 있지만, 기존 접근법은 다양한 time series 전반에서 forecasting workflow를 일관되게 적응시키지 못합니다. 우리는 제한된 strategy-level supervision과 dataset-specific meta-feature를 효과적인 forecasting strategy에 매핑하는 본질적 복잡성을 포함한 두 가지 주요 요인을 식별합니다. 이러한 도전을 해결하기 위해, 우리는 human-annotated supervision 없이 contrast-aware exploration과 agent-level causal lesson extraction을 통해 forecasting intelligence를 학습하는 새로운 agent training paradigm인 BECRA를 제안합니다. BECRA는 보이지 않는 dataset에서 in-context planning을 지원하는 symbolic strategy lesson을 distill하여 zero-shot training adaptation을 가능하게 합니다.

"},{"id":"64366","en":"Channel Adapter for Time Series Foundation Models in Zero-Shot Multivariate Forecasting","ko":"제로샷 다변량 예측에서 시계열 파운데이션 모델을 위한 채널 어댑터","authors":"Dongyuan Li, Renhe Jiang, Shun Zheng, Zheng Dong, Haotian Gao, Ying Zhang, Jiang Bian","pos":"#611","link":"https://openreview.net/forum?id=OJriSoFuDq","abs":"

Time Series Foundation Models (TSFMs) have achieved strong performance in univariate time series forecasting. However, most TSFMs rely on channel-independent pre-training that models each variable separately, limiting their ability to leverage inter-channel information that is crucial in real-world multivariate systems. Motivated by this limitation, we propose Chada, a lightweight plug-and-play channel adapter that allows frozen TSFMs to leverage multivariate correlations in a zero-shot setting. Chada first builds a budgeted pre-training dataset to cover diverse heterogeneous inter-channel dependency patterns. It then uses data-derived domain descriptors to learn a dataset-conditioned inter-channel similarity measure that reduces cross-domain metric distortion. Finally, it injects sparse inter-channel information via gated refinement, leveraging multivariate information without degrading intra-channel temporal dynamics. Extensive experiments on nine benchmarks validate the effectiveness of Chada, demonstrating consistent zero-shot improvements over four best-performing TSFMs while maintaining scalable deployment. Code is available at https://anonymous.4open.science/r/CHADA-A6.

","absKo":"

Time Series Foundation Model(TSFM)은 단변량 time series forecasting에서 강한 성능을 달성했습니다. 그러나 대부분의 TSFM은 각 변수를 별도로 model링하는 channel-independent pre-training에 의존하여, 실제 multivariate system에서 중요한 inter-channel information을 활용하는 능력이 제한됩니다. 이러한 한계에서 출발하여, 우리는 frozen TSFM이 zero-shot setting에서 multivariate correlation을 활용할 수 있도록 하는 경량 plug-and-play channel adapter인 Chada를 제안합니다. Chada는 먼저 다양한 이질적 inter-channel dependency pattern을 포괄하도록 budgeted pre-training dataset을 구축합니다. 그다음 data-derived domain descriptor를 사용하여 dataset-conditioned inter-channel similarity measure를 학습함으로써 cross-domain metric distortion을 줄입니다. 마지막으로 gated refinement를 통해 sparse inter-channel information을 주입하여, intra-channel temporal dynamics를 저해하지 않으면서 multivariate information을 활용합니다. 9개 benchmark에 대한 광범위한 실험은 Chada의 효과를 검증하며, scalable deployment를 유지하면서도 네 개의 최고 성능 TSFM보다 일관된 zero-shot 향상을 보임을 입증합니다. 코드는 https://anonymous.4open.science/r/CHADA-A6 에서 확인할 수 있습니다.

"},{"id":"65509","en":"CombinationTS: A Modular Framework for Understanding Time-Series Forecasting Models","ko":"CombinationTS: 시계열 예측 모델 이해를 위한 모듈식 프레임워크","authors":"Xiaorui Wang, Fanda Fan, Chenxi Wang, Yuxuan Yang, Rui Tang, Kuoyu Gao, simiao pang, Yuanfeng Shang, Liu, Wanling Gao, Lei Wang, Jianfeng Zhan","pos":"#612","link":"https://openreview.net/forum?id=CwHRT46VmC","abs":"Recent progress in time-series forecasting has led to rapidly increasing architectural complexity, yet many reported State-of-the-Art gains are statistically fragile or misattributed. We argue that progress requires a shift from model selection to modular attribution, identifying which components truly drive performance. We propose CombinationTS, a self-contained probabilistic evaluation framework that decomposes forecasting models into orthogonal modules—Input Transformation, Embedding, Encoder, and Decoder—and evaluates them under a shared evaluation condition space. By quantifying each component via marginalized effectiveness ($\\mu$) and stability ($\\sigma^2$), CombinationTS enables robust attribution beyond fragile point estimates. Through large-scale paired evaluation, we uncover the Identity Paradox: once the data view is well-designed, a parameter-free Identity encoder often matches or outperforms complex backbones. We further show that explicit structural priors introduced via input transformations yield a more favorable effectiveness–stability trade-off than increasing encoder complexity, establishing a principled baseline for architectural necessity. The code is available at https://anonymous.4open.science/r/CombinationTS.","absKo":"시계열 예측의 최근 진전은 아키텍처 복잡성을 빠르게 증가시켰지만, 보고된 많은 State-of-the-Art 향상은 통계적으로 취약하거나 잘못 귀속되어 있다. 우리는 진정한 진전에는 모델 선택에서 모듈식 귀속으로의 전환이 필요하다고 주장하며, 어떤 구성 요소가 실제로 성능을 견인하는지 식별해야 한다고 본다. 이를 위해 우리는 CombinationTS를 제안한다. CombinationTS는 예측 모델을 Input Transformation, Embedding, Encoder, Decoder의 직교 모듈로 분해하고, 공유된 평가 조건 공간에서 이들을 평가하는 독립형 확률적 평가 프레임워크다. 각 구성 요소를 marginalized effectiveness ($\\mu$)와 stability ($\\sigma^2$)로 정량화함으로써, CombinationTS는 취약한 점 추정을 넘어선 견고한 귀속을 가능하게 한다. 대규모 pairwise 평가를 통해 우리는 Identity Paradox를 발견한다. 데이터 관점이 잘 설계되면, 파라미터가 없는 Identity encoder가 복잡한 backbone과 비슷하거나 더 나은 성능을 보이는 경우가 많다. 또한 입력 변환을 통해 도입된 explicit structural prior가 encoder 복잡도를 높이는 것보다 더 유리한 effectiveness–stability trade-off를 제공함을 보이며, 아키텍처적 필요성에 대한 원리적 baseline을 확립한다. 코드는 https://anonymous.4open.science/r/CombinationTS 에서 제공된다."},{"id":"61370","en":"ConFlux: Multivariate Time Series in Flux, One Unified Forecast in Confluence","ko":"ConFlux: Flux 속 Multivariate Time Series, Confluence 속 하나의 통합 Forecast","authors":"Shiyu Wang, Yuchen Fang, Juntong Ni, Ziyi Zhang, Baichuan Mo, Xinyue Zhong, Chengxin Wang, Zhou Ye, Yang Xiang","pos":"#613","link":"https://openreview.net/forum?id=qugRvYjaFx","abs":"

Real-world multivariate time series are inherently in flux: different variables evolve asynchronously and interact in complex, time-varying ways, yet accurate forecasting requires these dispersed signals to converge into a single unified prediction. This structural mismatch between dynamic, heterogeneous inputs and a unified forecasting objective poses a fundamental challenge for building general-purpose multivariate forecasting models, especially in zero-shot and large-scale settings. To this end, inspired by the idea that \"all rivers run into the sea\", we propose ConFlux, a general-purpose foundation model for multivariate time-series forecasting by learning to adaptively integrate cross-channel information under a unified forecasting objective. Specifically, ConFlux first reorders variables to reduce cross-variable entanglement, then aggregates adjacent variables into compact patches that can be processed by a Vision Transformer-style architecture. This design shortens the effective context, reduces attention complexity, and provides a unified token representation for pre-training and downstream tasks. Experiments on 25 public datasets show that ConFlux achieves state-of-the-art performance in zero-shot, fine-tuning, and from-scratch settings, while offering faster inference and lower memory usage.

","absKo":"

현실 세계의 multivariate time series는 본질적으로 유동적이다. 서로 다른 변수들은 비동기적으로 진화하고 시간에 따라 복잡하고 가변적인 방식으로 상호작용하지만, 정확한 예측을 위해서는 이러한 분산된 신호가 하나의 통합된 예측으로 수렴해야 한다. 동적으로 이질적인 입력과 단일화된 forecasting 목표 사이의 이러한 구조적 불일치는, 특히 zero-shot 및 대규모 설정에서 범용 multivariate forecasting 모델을 구축하는 데 근본적인 도전 과제를 제기한다. 이러한 목적을 위해, \"모든 강은 바다로 흘러든다\"는 아이디어에서 영감을 받아, 우리는 cross-channel 정보를 unified forecasting objective 아래에서 적응적으로 통합하도록 학습하는 방식으로 multivariate time-series forecasting을 위한 범용 foundation model인 ConFlux를 제안한다. 구체적으로, ConFlux는 먼저 변수들을 재정렬하여 변수 간 얽힘을 줄이고, 이어서 인접한 변수들을 Vision Transformer 스타일 아키텍처로 처리할 수 있는 컴팩트한 patch로 집계한다. 이 설계는 effective context를 단축하고, attention complexity를 줄이며, pre-training과 downstream task를 위한 통합 token representation을 제공한다. 25개의 공개 데이터셋에서 수행한 실험은 ConFlux가 zero-shot, fine-tuning, from-scratch 설정에서 state-of-the-art 성능을 달성하면서도 더 빠른 추론과 더 낮은 메모리 사용량을 제공함을 보여준다.

"},{"id":"62284","en":"DRL-STAF: A DRL Framework for State-aware Forecasting of Complex Multivariate Hidden Markov Process","ko":"DRL-STAF: 복잡한 다변량 Hidden Markov Process의 상태 인식 예측을 위한 DRL 프레임워크","authors":"Manrui Jiang, Jingru Huang, Yong Chen, Chen Zhang","pos":"#614","link":"https://openreview.net/forum?id=hyksWoNb7h","abs":"

Forecasting multivariate hidden Markov processes is challenging due to nonlinear and nonstationary observations, latent state transitions, and cross-sequence dependencies. While deep learning methods achieve strong predictive accuracy, they typically lack explicit state modeling, whereas Hidden Markov Models (HMMs) provide interpretable latent states but struggle with complex nonlinear emissions and scalability. To address these limitations, we propose DRL-STAF, a Deep Reinforcement Learning based STate-Aware Forecasting framework that jointly predicts next-step observations and estimates the corresponding hidden states for complex multivariate hidden Markov processes. Specifically, DRL-STAF models complex nonlinear emissions using deep neural networks and estimates hidden state transitions via reinforcement learning, avoiding predefined transition structures and enabling flexible adaptation to diverse and high-order dynamics. In particular, DRL-STAF remains effective when typical HMM-based methods suffer from state-space explosion. Extensive experiments demonstrate that DRL-STAF consistently outperforms HMM variants, standalone deep learning models, and existing DL–HMM hybrids in both forecasting accuracy and hidden state estimation.

","absKo":"

비선형 및 비정상성 관측, 잠재 상태 전이, 그리고 시퀀스 간 의존성 때문에 다변량 Hidden Markov process의 예측은 어렵습니다. 딥러닝 방법은 강력한 예측 정확도를 달성하지만, 일반적으로 명시적인 상태 모델링이 부족합니다. 반면 Hidden Markov Model(HMM)은 해석 가능한 잠재 상태를 제공하지만 복잡한 비선형 방출과 확장성에서 어려움을 겪습니다. 이러한 한계를 해결하기 위해, 우리는 DRL-STAF를 제안합니다. DRL-STAF는 Deep Reinforcement Learning 기반 STate-Aware Forecasting 프레임워크로, 복잡한 다변량 Hidden Markov process에 대해 다음 시점 관측값을 공동으로 예측하고 대응하는 hidden state를 추정합니다. 구체적으로, DRL-STAF는 deep neural network를 사용해 복잡한 비선형 방출을 모델링하고 reinforcement learning을 통해 hidden state 전이를 추정함으로써, 사전 정의된 전이 구조를 피하고 다양한 고차 동역학에 유연하게 적응할 수 있습니다. 특히 DRL-STAF는 일반적인 HMM 기반 방법이 state-space explosion을 겪는 경우에도 효과적으로 동작합니다. 광범위한 실험 결과, DRL-STAF는 예측 정확도와 hidden state 추정 모두에서 HMM 변형, 독립적인 딥러닝 모델, 그리고 기존 DL-HMM 하이브리드보다 일관되게 우수함을 보여줍니다.

"},{"id":"64391","en":"TiMi: Empower Time Series Transformers with Multimodal Mixture of Experts","ko":"TiMi: 멀티모달 Mixture of Experts로 시계열 Transformer 강화하기","authors":"Jiafeng Lin, Yuxuan Wang, HUAKUN LUO, Jianmin Wang, Zhongyi Pei","pos":"#400","link":"https://openreview.net/forum?id=O5wCHXOEug","abs":"

Multimodal time series forecasting has garnered significant attention for its potential to provide more robust and accurate predictions than traditional single-modality models by leveraging rich information inherent in other modalities. However, due to fundamental challenges in modality alignment, existing methods often struggle to effectively incorporate multimodal data into predictions, particularly textual information that has a causal influence on time series fluctuations, such as emergency reports and policy announcements. In this paper, we reflect on the role of textual information in numerical forecasting and propose Time series transformers with Multimodal Mixture-of-Experts, TiMi, to unleash the causal reasoning capabilities of LLMs. Concretely, TiMi utilizes language models to generate inferences on future developments, which then serve as guidance for time series forecasting. To seamlessly integrate both exogenous factors and time series into predictions, we introduce a Multimodal Mixture-of-Experts (MMoE) module as a lightweight plug-in to empower Transformer-based time series models for multimodal forecasting, eliminating the need for explicit representation-level alignment. Experimentally, our proposed TiMi demonstrates consistent state-of-the-art performance on sixteen real-world multimodal forecasting benchmarks, outperforming advanced baselines while offering strong adaptability and interpretability.

","absKo":"

멀티모달 시계열 예측은 다른 modality에 내재된 풍부한 정보를 활용함으로써 기존 단일 modality model보다 더 강건하고 정확한 예측을 제공할 잠재력 때문에 큰 주목을 받아왔다. 그러나 modality 정렬이라는 근본적 어려움 때문에, 기존 방법들은 특히 긴급 보고서나 정책 발표처럼 시계열 변동에 인과적 영향을 미치는 텍스트 정보를 예측에 효과적으로 통합하는 데 자주 어려움을 겪는다. 본 논문에서는 수치 예측에서 텍스트 정보의 역할을 재고하고, LLM의 인과 추론 능력을 발휘하기 위해 Time series transformers with Multimodal Mixture-of-Experts, TiMi를 제안한다. 구체적으로, TiMi는 language model을 활용해 미래 전개에 대한 추론을 생성하고, 이를 시계열 예측의 guidance로 사용한다. 외생 요인과 시계열을 예측에 매끄럽게 통합하기 위해, 우리는 Transformer 기반 시계열 model에 멀티모달 forecasting을 위한 가벼운 plug-in으로 Multimodal Mixture-of-Experts(MMoE) module을 도입하여, 명시적인 representation-level alignment의 필요성을 제거한다. 실험적으로, 제안한 TiMi는 16개의 실제 멀티모달 forecasting benchmark 전반에서 일관되게 state-of-the-art 성능을 보이며, 강력한 adaptability와 interpretability를 제공하면서도 고도화된 baseline을 능가한다.

"},{"id":"66058","en":"V-ABS: Action-Observer Driven Beam Search for Dynamic Visual Reasoning","ko":"V-ABS: 동적 시각 추론을 위한 행위자-관찰자 유도 빔 서치","authors":"Zhiwei Ning, Xuanang Gao, Jiaxi Cao, Gengming Zhang, Shengnan Ma, Wenwen Tong, Hanming Deng, JIE YANG, Wei Liu","pos":"#1710","link":"https://openreview.net/forum?id=7Q2jAzmDAR","abs":"

Multimodal large language models (MLLMs) have achieved remarkable success in general perception, yet complex multi-step visual reasoning remains a persistent challenge. Although recent agentic approaches incorporate tool use, they often neglect critical execution feedback. Consequently, they suffer from the imagination-action-observer (IAO) bias, a misalignment between prior imagination and observer feedback that undermines reasoning stability and optimality. To bridge this gap, we introduce V-ABS, an action-observer driven beam search framework that enables deliberate reasoning through thinker-actor-observer iterations. We also propose an entropy-based adaptive weighting algorithm to mitigate the IAO bias by dynamically balancing the confidence scores between the policy priors and the observational feedback. Moreover, we construct a large-scale supervised fine-tuning (SFT) dataset comprising over 80k samples to guide the model to assign higher prior confidence to correct action paths. Extensive experiments across eight diverse benchmarks show that V-ABS achieves state-of-the-art performance, delivering an average improvement of 19.7\\% on the Qwen3-VL-8B baseline and consistent gains across both open-source and proprietary models.

","absKo":"

Multimodal large language models(MLLMs)은 일반적 perception에서 놀라운 성과를 거두었지만, 복잡한 multi-step visual reasoning은 여전히 지속적인 도전 과제이다. 최근의 agentic approach는 tool use를 통합하지만, 중요한 execution feedback을 종종 간과한다. 그 결과, 이들은 imagination-action-observer(IAO) bias, 즉 prior imagination과 observer feedback 사이의 불일치로 인해 reasoning의 stability와 optimality가 저하되는 문제를 겪는다. 이러한 격차를 메우기 위해, 우리는 thinker-actor-observer 반복을 통해 신중한 reasoning을 가능하게 하는 action-observer driven beam search framework인 V-ABS를 제안한다. 또한 policy prior와 observational feedback 사이의 confidence score를 동적으로 균형 조정하여 IAO bias를 완화하는 entropy-based adaptive weighting algorithm을 제안한다. 더 나아가, 모델이 올바른 action path에 더 높은 prior confidence를 부여하도록 유도하기 위해 8만 개 이상의 sample로 구성된 대규모 supervised fine-tuning(SFT) dataset을 구축한다. 8개의 다양한 benchmark에 대한 광범위한 실험은 V-ABS가 state-of-the-art 성능을 달성하며, Qwen3-VL-8B baseline에서 평균 19.7\\%의 향상을 제공하고 open-source 및 proprietary model 모두에서 일관된 성능 향상을 보임을 보여준다.

"},{"id":"66376","en":"Discovering Symmetry Groups with Flow Matching","ko":"Flow Matching을 이용한 대칭 그룹 발견","authors":"Yuxuan Chen, Jung Yeon Park, Floor Eijkelboom, Jianke Yang, Jan-Willem van de Meent, Lawson Wong, Robin Walters","pos":"#3010","link":"https://openreview.net/forum?id=4FSkKnTnvI","abs":"Symmetry is fundamental to understanding physical systems and can improve performance and sample efficiency in machine learning. Both pursuits require knowledge of the underlying symmetries in data, yet discovering these symmetries automatically is challenging. We propose LieFlow, a novel framework that reframes symmetry discovery as a distribution learning problem on Lie groups. Instead of searching for the symmetry generators, our approach operates directly in group space, modeling a symmetry distribution over a large hypothesis group $G$. The support of the learned distribution reveals the underlying symmetry group $H \\subseteq G$. Unlike previous works, LieFlow can discover both continuous and discrete symmetries within a unified framework, without assuming a fixed Lie algebra basis or a specific distribution over the group elements. Experiments on synthetic 2D and 3D point clouds and ModelNet10 show that LieFlow accurately discovers continuous and discrete subgroups, significantly outperforming a state-of-the-art baseline, LieGAN, in identifying discrete symmetries.","absKo":"대칭성은 물리적 시스템을 이해하는 데 근본적이며, machine learning에서 성능과 sample efficiency를 향상시킬 수 있다. 두 목표 모두 데이터에 내재한 대칭성에 대한 지식을 필요로 하지만, 이를 자동으로 발견하는 것은 어렵다. 우리는 symmetry discovery를 Lie group 위의 distribution learning 문제로 재해석하는 새로운 framework인 LieFlow를 제안한다. symmetry generator를 탐색하는 대신, 우리의 접근법은 group space에서 직접 동작하며 큰 hypothesis group $G$ 위의 symmetry distribution을 모델링한다. 학습된 distribution의 support는 underlying symmetry group $H \\subseteq G$를 드러낸다. 이전 연구와 달리, LieFlow는 고정된 Lie algebra basis나 group element에 대한 특정 distribution을 가정하지 않고도 continuous symmetry와 discrete symmetry를 하나의 통합된 framework 안에서 모두 발견할 수 있다. synthetic 2D 및 3D point cloud와 ModelNet10에서의 실험은 LieFlow가 continuous 및 discrete subgroup을 정확하게 발견하며, discrete symmetry 식별에서 state-of-the-art baseline인 LieGAN을 크게 능가함을 보여준다."},{"id":"64220","en":"PGT: Procedurally Generated Tasks for improving fine-grained understanding in MLLMs","ko":"PGT: MLLM의 세밀한 이해 향상을 위한 절차적 생성 과제","authors":"Rim Assouel, Amir Bar, Michal Drozdzal, Adriana Romero-Soriano","pos":"#3012","link":"https://openreview.net/forum?id=PrM9JAYDF6","abs":"

Despite remarkable progress in Multimodal Large Language Models (MLLMs), these models still struggle with fine-grained understanding tasks. In this work, we propose Procedurally Generated Tasks (PGT) a simple data-driven framework that serves a dual purpose: inducing fine-grained visual understanding and acting as a low-cost diagnostic tool to identify the source of perception failures. By overlaying unambiguous geometric primitives on images, PGT generate additional dense supervision that disentangles visual grounding capability from semantic priors. Extensive experiments on relational, quantitative, and 3D/depth understanding benchmarks show that PGT yields remarkable gains across diverse architectures. Instruction tuning MLLMs on LLaVA-v1.5-Instruct augmented with PGT data results in improvements of up to +20\\% on the What’sUp benchmark and +13.3\\% on CV-Bench-2D, while maintaining general perception capabilities. Moreover, finetuning state-of-the-art MLLMs on PGT data leads to boosts of up to +5.5\\% on What’sUp and +8.3\\% on CV-Bench-2D. These findings demonstrate that PGT effectively address the bottleneck of fine-grained perception, revealing that many spatial reasoning deficits stem from inadequate supervision signals rather than inherent architectural or resolution limitations.

","absKo":"

Multimodal Large Language Models (MLLMs)에서 괄목할 만한 진전이 있었음에도, 이러한 모델은 여전히 세밀한 이해(fine-grained understanding) 과제에서 어려움을 겪는다. 본 연구에서는 Procedurally Generated Tasks (PGT)를 제안하는데, 이는 세밀한 시각 이해를 유도하는 동시에 지각 실패의 원인을 식별하는 저비용 진단 도구로도 기능하는 간단한 데이터 기반 프레임워크이다. 이미지 위에 모호함이 없는 기하학적 프리미티브를 오버레이함으로써, PGT는 시각적 grounding 능력과 semantic prior를 분리해 내는 추가적인 dense supervision을 생성한다. 관계적, 정량적, 3D/depth 이해 벤치마크 전반에 걸친 대규모 실험은 PGT가 다양한 아키텍처에서 현저한 성능 향상을 보임을 보여준다. LLaVA-v1.5-Instruct에 PGT 데이터를 추가하여 MLLM을 instruction tuning하면 What’sUp 벤치마크에서 최대 +20\\%, CV-Bench-2D에서 +13.3\\%까지 향상되면서도 일반적인 지각 능력은 유지된다. 또한, state-of-the-art MLLM을 PGT 데이터로 finetuning하면 What’sUp에서 최대 +5.5\\%, CV-Bench-2D에서 +8.3\\%의 향상을 얻는다. 이러한 결과는 PGT가 세밀한 지각의 병목을 효과적으로 해결함을 보여 주며, 많은 spatial reasoning 결함이 본질적인 아키텍처나 해상도 한계가 아니라 불충분한 supervision signal에서 비롯됨을 드러낸다.

"},{"id":"62267","en":"Diffusion Language Model Parallel Decoding via Product-of-Experts Bridge","ko":"Product-of-Experts Bridge로 Diffusion Language Model 병렬 디코딩","authors":"Juntong Shi, Brian Trippe, Jure Leskovec, Stefano Ermon, Minkai Xu","pos":"#2500","link":"https://openreview.net/forum?id=iA7hUqvsvz","abs":"Diffusion language models (DLMs) offer substantial speed advantages through parallel decoding, but the lack of token dependencies limits generation quality compared to autoregressive (AR) models. Recent progress attempts to bridge the gap via importance sampling, with DLM being the proposal and AR being the target. However, due to the huge gap between their probability space, the sampling requires a large number of particles and thus expensive computation. In this paper, we introduce PoE-Bridge, a novel decoding framework that drastically improves generation speed and accuracy by introducing an intermediate distribution to bridge the gap. The distribution is constructed as a Product-of-Experts (PoE) of the DLM proposal and the AR target. With the intermediate distribution, we first conduct multi-token sampling with the DLM and then apply rejection sampling using the PoE to retain only the verified tokens. The generated chunks are then evaluated by the AR target via importance sampling to produce the final faithful generation. We further propose several improved techniques, including mixed-temperature sampling for enhanced diversity and elastic rejection windows for reducing wasted verification. Empirically, PoE-Bridge achieves significantly improved accuracy with $5\\times$ speedup over the standard DLM decoding approach, and recovering at least 95% of the target AR model's performance, efficiently advancing most of the quality gap on challenging mathematical reasoning and coding tasks.","absKo":"Diffusion language models (DLMs)는 병렬 디코딩을 통해 상당한 속도 이점을 제공하지만, token dependency가 부족해 autoregressive (AR) 모델에 비해 생성 품질이 제한된다. 최근의 진전은 DLM을 proposal로, AR을 target으로 하는 importance sampling을 통해 이 격차를 좁히려 시도하고 있다. 그러나 두 모델의 probability space 사이의 격차가 매우 크기 때문에, sampling에는 많은 수의 particle이 필요하고 그만큼 계산 비용도 크다. 본 논문에서는 intermediate distribution을 도입해 이 격차를 메움으로써 생성 속도와 정확도를 극적으로 향상시키는 새로운 decoding framework인 PoE-Bridge를 소개한다. 이 distribution은 DLM proposal과 AR target의 Product-of-Experts (PoE)로 구성된다. intermediate distribution을 사용하여, 먼저 DLM으로 multi-token sampling을 수행한 뒤 PoE를 이용한 rejection sampling을 적용해 검증된 token만 남긴다. 이후 생성된 chunk는 importance sampling을 통해 AR target으로 평가되어 최종적으로 faithful generation을 생성한다. 또한 다양성 향상을 위한 mixed-temperature sampling과 검증 낭비를 줄이기 위한 elastic rejection windows를 포함한 여러 개선 기법도 제안한다. 실험적으로 PoE-Bridge는 표준 DLM decoding 접근법 대비 $5\\times$의 속도 향상을 보이면서도 정확도를 크게 개선하며, target AR model 성능의 최소 95%를 회복해, 어려운 수학적 추론 및 coding task에서 품질 격차의 대부분을 효율적으로 좁힌다."},{"id":"63762","en":"Where Rectified Flows Leak: Characterizing Membership Signals Along the Interpolation Path","ko":"Rectified Flow의 누출 지점: 보간 경로를 따라 Membership 신호 특성화","authors":"Thomas Sesmat, Gabriel Meseguer-Brocal, Geoffroy Peeters","pos":"#2706","link":"https://openreview.net/forum?id=Ty5X41WbJw","abs":"Understanding what generative models retain from training data remains challenging, with implications for copyright and privacy. This question becomes particularly relevant as Rectified Flows power increasingly deployed systems. We analyze the interpolation path $X_\\lambda = (1-\\lambda)X_0 + \\lambda X_1$ that defines Rectified Flow training. We show that train-test distinguishability follows a bell-shaped curve over $\\lambda$, with a maximum whose location we derive in closed form under Gaussian assumptions. This signal accumulates during training while validation metrics remain stable. We validate these predictions on both audio and images, and show that the bell-shaped structure is universal while the peak prediction holds when our assumptions are satisfied. As a proof of concept, we implement a membership inference attack achieving 0.91 AUC by exploiting this $\\lambda$-resolved structure.","absKo":"생성 모델이 학습 데이터에서 무엇을 보존하는지 이해하는 일은 여전히 어렵고, 이는 저작권과 프라이버시와도 연결된다. 이러한 문제는 Rectified Flows가 점점 더 널리 배치되는 시스템을 구동하는 지금 특히 중요해진다. 우리는 Rectified Flow 학습을 정의하는 보간 경로 $X_\\lambda = (1-\\lambda)X_0 + \\lambda X_1$를 분석한다. train-test 구분 가능성이 $\\lambda$에 대해 종 모양 곡선을 따른다는 것을 보이고, Gaussian 가정하에서 그 최대값의 위치를 닫힌형으로 유도한다. 이 신호는 학습이 진행되는 동안 누적되는 반면 validation metric은 안정적으로 유지된다. 우리는 이러한 예측을 audio와 image 모두에서 검증하고, 종 모양 구조는 보편적이며 peak 예측은 우리의 가정이 성립할 때 맞아떨어짐을 보인다. 개념 증명으로서, 우리는 이 $\\lambda$로 분해된 구조를 활용해 0.91 AUC를 달성하는 membership inference attack을 구현한다."},{"id":"65263","en":"When Sample Selection Bias Precipitates Model Collapse","ko":"샘플 선택 편향이 모델 붕괴를 촉발할 때","authors":"Xinbao Qiao, Xianglong Du, Wei Liu, Jingqi Zhang, Peihua Mai, Meng Zhang, Yan (James) Pang","pos":"#2707","link":"https://openreview.net/forum?id=FFXvnzM254","abs":"

The proliferation of recursive synthetic data training promises to alleviate data scarcity but introduces the existential risk of model collapse, wherein recursive training on synthetic data erodes distributional tails and homogenizes outputs. Current literature identifies data selection as a pivotal solution, employing verifiers to prune datasets in pursuit of synthetic samples that approximate the true data manifold. However, this approach hinges on the fragile and often unrealistic assumption that a perfect verifier possesses global distributional knowledge. In real-world scenarios characterized by data silos, such as fragmented healthcare consortia or proprietary financial institutions, this assumption is invalidated by the inherent fragmentation of knowledge. We theoretically prove that such siloed selection accelerates model collapse, driving diversity decay governed by a power law. To bridge this gap, we propose an automated filtering criterion that synergizes the sensitivity theorem with Wasserstein geometry. Specifically, multiple parties collaboratively compute geodesic interpolations and the Wasserstein Barycenter as proxy measures, without exchanging raw data. These proxies serve as a collective reference, enabling multiple parties to score synthetic data rather than relying on a single biased perspective in a data silo. Empirical results show the failure of the baseline on skewed distributions, whereas our methods effectively prevent collapse. Code available at Anonymous Github.

","absKo":"

Recursive synthetic data training의 확산은 data scarcity를 완화할 가능성을 제시하지만, synthetic data에 대한 재귀적 학습이 distributional tail을 침식하고 출력을 획일화하는 model collapse라는 존재론적 위험을 수반한다. 현재 문헌은 data selection을 핵심 해결책으로 보고 있으며, verifier를 사용해 dataset을 정제함으로써 true data manifold를 근사하는 synthetic sample을 찾고자 한다. 그러나 이러한 접근은 완벽한 verifier가 전역 분포 지식을 가진다는 취약하고 종종 비현실적인 가정에 의존한다. fragmented healthcare consortium이나 proprietary financial institution과 같은 data silo가 존재하는 현실에서는, 지식의 본질적 파편화로 인해 이 가정이 성립하지 않는다. 우리는 이러한 siloed selection이 model collapse를 가속화하고, power law에 의해 지배되는 diversity decay를 야기함을 이론적으로 증명한다. 이 간극을 메우기 위해, 우리는 sensitivity theorem과 Wasserstein geometry를 결합한 automated filtering criterion을 제안한다. 구체적으로, 여러 party가 raw data를 교환하지 않은 채 geodesic interpolation과 Wasserstein Barycenter를 proxy measure로 공동 계산한다. 이 proxy는 collective reference로 작동하여, data silo의 단일 편향된 관점에 의존하지 않고 여러 party가 synthetic data를 평가할 수 있게 한다. 실험 결과는 skewed distribution에서 baseline이 실패함을 보이는 반면, 우리의 방법은 collapse를 효과적으로 방지함을 보여준다. 코드는 Anonymous Github에서 제공된다.

"},{"id":"60617","en":"VectorWorld: Efficient Streaming World Model via Diffusion Flow on Vector Graphs","ko":"VectorWorld: 벡터 그래프 위 Diffusion Flow를 통한 효율적 스트리밍 월드 모델","authors":"Chaokang Jiang, Desen Zhou, Jiuming Liu, Li Sun","pos":"#2708","link":"https://openreview.net/forum?id=yTEEiE3YtD","abs":"Closed-loop evaluation of autonomous-driving policies requires interactive simulation beyond log replay. However, existing generative world models often degrade in closed loop due to (i) history-free initialization that mismatches policy inputs, (ii) multi-step sampling latency that violates real-time budgets, and (iii) compounding kinematic infeasibility over long horizons. We propose VectorWorld, a streaming world model that incrementally generates ego-centric $64 \\mathrm{m}\\times 64\\mathrm{m}$ lane--agent vector-graph tiles during rollout. VectorWorld aligns initialization with history-conditioned policies by producing a policy-compatible interaction state via a motion-aware gated VAE. It enables real-time outpainting via solver-free one-step masked completion with an edge-gated relational DiT trained with interval-conditioned MeanFlow and JVP-based large-step supervision. To stabilize long-horizon rollouts, we introduce $\\Delta$Sim, a physics-aligned non-ego (NPC) policy with hybrid discrete--continuous actions and differentiable kinematic logit shaping. On Waymo open motion and nuPlan, VectorWorld improves map-structure fidelity and initialization validity, and supports stable, real-time $1\\mathrm{km}+$ closed-loop rollouts.","absKo":"자율주행 policy의 closed-loop 평가는 log replay를 넘어서는 interactive simulation을 요구한다. 그러나 기존 generative world model은 (i) policy input과 맞지 않는 history-free initialization, (ii) real-time budget을 위반하는 multi-step sampling latency, (iii) 긴 horizon에서 누적되는 kinematic infeasibility로 인해 closed loop에서 성능이 저하되는 경우가 많다. 우리는 rollout 동안 ego-centric $64 \\mathrm{m}\\times 64\\mathrm{m}$ lane--agent vector-graph tile을 점진적으로 생성하는 streaming world model인 VectorWorld를 제안한다. VectorWorld는 motion-aware gated VAE를 통해 policy-compatible interaction state를 생성함으로써 history-conditioned policy와 initialization을 정렬한다. 또한 interval-conditioned MeanFlow와 JVP 기반 large-step supervision으로 학습된 edge-gated relational DiT를 사용해 solver-free one-step masked completion을 수행함으로써 real-time outpainting을 가능하게 한다. 긴 horizon rollout을 안정화하기 위해, 우리는 hybrid discrete--continuous action과 differentiable kinematic logit shaping을 갖는 physics-aligned non-ego(NPC) policy인 $\\Delta$Sim을 도입한다. Waymo open motion과 nuPlan에서 VectorWorld는 map-structure fidelity와 initialization validity를 향상시키며, 안정적이고 real-time인 $1\\mathrm{km}+$ closed-loop rollout을 지원한다."},{"id":"65534","en":"Spatiotemporal Imputation with Graph-Informed Flow Matching","ko":"그래프 정보 기반 Flow Matching을 통한 시공간 대치","authors":"Zepeng Zhang, Aref Einizade, Jhony H. Giraldo, Olga Fink","pos":"#2712","link":"https://openreview.net/forum?id=Cc5oIi4EeL","abs":"

Missing data is a common challenge in spatiotemporal systems, arising in applications such as air quality monitoring and urban traffic management. Traditional machine learning approaches, like recurrent and graph neural networks, rely on iterative propagation, which tends to accumulate errors over time and space. Recent diffusion-based methods mitigate error propagation but require iterative sampling and often depend on problem-agnostic Gaussian priors, limiting both efficiency and effectiveness. To address these limitations, we propose GiFlow, a Graph-Informed Flow Matching framework for spatiotemporal imputation. GiFlow replaces the typical Gaussian prior with a graph-informed prior constructed via spatiotemporal filtering of observable signals, which better aligns the source distribution to the target and thereby simplifies the generation trajectory. The flow field is parameterized by a hybrid vector field model that integrates spatial attention, temporal attention, and spatiotemporal propagation, enabling joint modeling of spatial and temporal dependencies. Unlike diffusion models, GiFlow is trained via direct regression and supports deterministic, few-step generation at inference. Extensive experiments on both synthetic and real-world datasets with different missing patterns and missing rates demonstrate that the proposed GiFlow outperforms the state-of-the-art approaches in spatiotemporal imputation.

","absKo":"

결측 데이터는 대기질 모니터링과 도시 교통 관리 같은 응용에서 나타나는, 시공간 시스템의 일반적인 과제이다. recurrent network와 graph neural network 같은 전통적인 machine learning 접근법은 반복적 propagation에 의존하며, 이는 시간과 공간 전반에서 오류를 누적시키는 경향이 있다. 최근 diffusion 기반 방법은 오류 전파를 완화하지만 반복적 sampling이 필요하고, 종종 문제에 무관한 Gaussian prior에 의존하여 효율성과 효과성 모두를 제한한다. 이러한 한계를 해결하기 위해, 우리는 시공간 imputation을 위한 Graph-Informed Flow Matching 프레임워크인 GiFlow를 제안한다. GiFlow는 일반적인 Gaussian prior를 관측 가능한 신호의 시공간 filtering을 통해 구성한 graph-informed prior로 대체하여 source distribution을 target과 더 잘 정렬하고, 그 결과 생성 궤적을 단순화한다. flow field는 spatial attention, temporal attention, 그리고 spatiotemporal propagation을 통합한 hybrid vector field model로 parameterize되며, 이를 통해 공간 및 시간 의존성을 공동으로 모델링할 수 있다. diffusion model과 달리, GiFlow는 직접 회귀 방식으로 학습되며 추론 시 deterministic한 few-step generation을 지원한다. 다양한 missing pattern과 missing rate를 가진 synthetic 및 real-world 데이터셋에 대한 광범위한 실험은 제안한 GiFlow가 spatiotemporal imputation에서 state-of-the-art 방법들을 능가함을 보여준다.

"},{"id":"61422","en":"Riemannian Diffusion Models on General Manifolds via Physics-Informed Neural Networks","ko":"Physics-Informed Neural Networks를 통한 General Manifolds 위 Riemannian Diffusion Models","authors":"Gyeonghoon Ko, Juho Lee","pos":"#2713","link":"https://openreview.net/forum?id=qK2yxjJJCh","abs":"Riemannian diffusion models generalize score-based generative modeling to manifold-supported data via stochastic diffusion equations on the manifold. However, training requires sampling from and differentiating the manifold heat kernel, which is rarely available in closed form beyond a few highly symmetric manifolds. We propose a general approach that approximates the heat kernel by directly solving the manifold heat equation with a physics-informed neural network (PINN). Given an explicit manifold specification, we choose a coordinate system, derive the corresponding heat (Fokker--Planck) equation and a short-time asymptotic approximation, and then train a PINN to learn the log heat kernel. The resulting surrogate enables both forward noising (heat-kernel sampling) and conditional-score evaluation for denoising score matching. We demonstrate the method on diverse manifolds including $S^2$, $SO(3)$, $\\mathrm{SPD}(n)$, and permutation-quotiented point clouds.","absKo":"Riemannian diffusion models는 manifold 위에 지지된 데이터에 대한 확률적 diffusion equation을 통해 score-based generative modeling을 일반화한다. 그러나 학습에는 manifold heat kernel로부터의 샘플링과 이를 미분하는 과정이 필요하며, 이는 몇몇 매우 대칭적인 manifold를 제외하면 닫힌 형태로 거의 제공되지 않는다. 우리는 physics-informed neural network(PINN)를 사용해 manifold heat equation을 직접 풀어 heat kernel을 근사하는 일반적 접근법을 제안한다. 명시적인 manifold 사양이 주어지면, 우리는 좌표계를 선택하고 대응하는 heat(Fokker--Planck) 방정식과 short-time asymptotic approximation을 유도한 뒤, PINN을 학습시켜 log heat kernel을 추정한다. 이렇게 얻은 surrogate는 forward noising(heat-kernel sampling)과 denoising score matching을 위한 conditional-score 평가를 모두 가능하게 한다. 우리는 이 방법을 $S^2$, $SO(3)$, $\\mathrm{SPD}(n)$, permutation-quotiented point clouds를 포함한 다양한 manifold에서 입증한다."},{"id":"65552","en":"Reasoning on the Manifold: Bidirectional Consistency for Self-Verification in Diffusion Language Models","ko":"다양체에서의 추론: Diffusion Language Model의 자기 검증을 위한 양방향 일관성","authors":"Jiaoyang Ruan, Xin Gao, Yinda Chen, Hengyu Zeng, Liang Du, Guanghao Li, Jie Fu, Jian Pu","pos":"#2715","link":"https://openreview.net/forum?id=CUVBdw2tKy","abs":"

While Diffusion Large Language Models (dLLMs) offer structural advantages for global planning, efficiently verifying that they arrive at correct answers via valid reasoning traces remains a critical challenge. In this work, we propose a geometric perspective: Reasoning on the Manifold. We hypothesize that valid generation trajectories reside as stable attractors on the high-density manifold of the learned distribution, whereas invalid paths exhibit off-manifold drift. To operationalize this, we introduce Bidirectional Manifold Consistency (BMC), a training-free, unsupervised metric that quantifies the stability of the generated sequence through a forward-masking and backward-reconstruction cycle. Empirically, we demonstrate BMC's versatility across the full reasoning lifecycle: (1) in Diagnosis, it serves as a robust discriminator of solution validity without ground truth answer; (2) in Inference, it enables rejection resampling to effectively concentrate computational resources on complex reasoning tasks; and (3) in Alignment, it functions as a dense geometric reward that transforms sparse outcome supervision into fine-grained guidance, empowering models to self-evolve beyond standard baselines. Our results establish intrinsic geometric stability as a robust indicator of correctness for dLLMs.

","absKo":"

Diffusion Large Language Models (dLLMs)는 전역 계획(global planning)에서 구조적 이점을 제공하지만, valid reasoning trace를 통해 올바른 답에 도달했음을 효율적으로 검증하는 것은 여전히 중요한 과제다. 본 연구에서는 기하학적 관점인 Reasoning on the Manifold를 제안한다. 우리는 valid generation trajectory가 학습된 분포의 high-density manifold 위에 안정적인 attractor로 존재하는 반면, invalid path는 manifold 밖으로 drift한다고 가정한다. 이를 실제로 구현하기 위해, 우리는 forward-masking과 backward-reconstruction cycle을 통해 생성된 sequence의 안정성을 정량화하는 training-free, unsupervised metric인 Bidirectional Manifold Consistency (BMC)를 도입한다. 실험적으로 우리는 BMC의 유용성이 전체 reasoning lifecycle 전반에 걸친다는 것을 보인다. (1) Diagnosis에서는 ground truth answer 없이도 solution validity를 판별하는 강건한 discriminator로 기능하고, (2) Inference에서는 rejection resampling을 가능하게 하여 계산 자원을 복잡한 reasoning task에 효과적으로 집중시키며, (3) Alignment에서는 희소한 outcome supervision을 세밀한 guidance로 전환하는 dense geometric reward로 작동하여 모델이 standard baseline을 넘어 self-evolve할 수 있게 한다. 우리의 결과는 intrinsic geometric stability가 dLLM의 정답성을 나타내는 강건한 지표임을 확립한다.

"},{"id":"62058","en":"Flexibility-Aware Geometric Latent Diffusion for Full-Atom Peptide Design","ko":"Full-Atom Peptide Design를 위한 Flexibility-Aware Geometric Latent Diffusion","authors":"Dongjiang Niu, Xiaofeng Wang, Zhiqiang Wei, Zhen Li","pos":"#2800","link":"https://openreview.net/forum?id=kH2KFnlxmU","abs":"

Although peptides are well suited for flexible and shallow binding interfaces, their intrinsic flexibility induces a strongly coupled sequence–structure relationship that current fixed-geometry latent models cannot simultaneously model with conformational diversity and physical feasibility, ultimately limiting design quality. To overcome this bottleneck, PepFGLD is proposed as a receptor-conditioned, flexibility-aware framework for full-atom peptide design. The framework is motivated by a systematic analysis of existing limitations: geometry shifts driven by interfacial flexibility are not well captured by standard equivariant encoders; the static combination of sequence information and 3D geometry cannot represent their dynamic interactions; and diffusion models without timely geometric feedback tend to drift away from physically reasonable energy landscapes. In PepFGLD, FlexEGNN is used to improve the sensitivity of geometric representations to local flexibility, a coherent and adaptable latent conformational manifold is formed through bidirectional sequence–structure interaction and nonlinear latent mapping, and a time-dependent energy-guided diffusion mechanism is incorporated to balance exploration and convergence during diffusion so that sampling trajectories are continuously guided toward physically feasible full-atom structures. PepFGLD yields improved binding affinity and design success across multiple peptide design tasks.

","absKo":"

peptide는 유연하고 shallow한 binding interface에 적합하지만, intrinsic flexibility가 sequence와 structure 사이에 강하게 결합된 관계를 유도하여, 현재의 fixed-geometry latent model로는 conformational diversity와 physical feasibility를 동시에 모델링하기 어렵고 결국 설계 품질을 제한한다. 이러한 병목을 극복하기 위해 PepFGLD를 receptor-conditioned, flexibility-aware full-atom peptide design 프레임워크로 제안한다. 이 프레임워크는 기존 한계에 대한 체계적 분석에 기반한다. 즉, interfacial flexibility에 의해 유발되는 geometry shift는 표준 equivariant encoder가 충분히 포착하지 못하고, sequence information과 3D geometry의 정적인 결합은 이들의 동적 상호작용을 표현하지 못하며, 적시에 geometric feedback이 없는 diffusion model은 물리적으로 합리적인 energy landscape에서 벗어나기 쉽다. PepFGLD에서는 FlexEGNN을 사용해 local flexibility에 대한 geometric representation의 민감도를 향상시키고, 양방향 sequence-structure interaction과 nonlinear latent mapping을 통해 일관되고 적응적인 latent conformational manifold를 형성하며, 시간 의존적 energy-guided diffusion mechanism을 도입해 diffusion 동안 exploration과 convergence의 균형을 맞춤으로써 sampling trajectory가 물리적으로 feasible한 full-atom structure를 향하도록 지속적으로 유도한다. PepFGLD는 여러 peptide design task에서 binding affinity와 design success를 향상시킨다.

"},{"id":"62460","en":"Insertion Based Sequence Generation with Learnable Order Dynamics","ko":"학습 가능한 Order Dynamics를 통한 Insertion 기반 시퀀스 생성","authors":"Dhruvesh Patel, Benjamin Rozonoyer, Gaurav Pandey, Tahira Naseem, Ramón Astudillo, Andrew McCallum","pos":"#2805","link":"https://openreview.net/forum?id=gPxxn38tAK","abs":"

In many domains generating variable length sequences through insertions provides greater flexibility over autoregressive models. However, the action space of insertion models is much larger than that of autoregressive models (ARMs) making the learning challenging. To address this, we incorporate trainable order dynamics into the target rates for discrete flow matching, and show that with suitable choices of parameterizations, joint training of the target order dynamics and the generator is tractable without the need for numerical simulation. As the generative insertion model, we use a variable length masked diffusion model, which generates by inserting and filling mask tokens. On graph traversal tasks for which a locally optimal insertion order is known, we explore the choices of parameterization empirically and demonstrate the trade-offs between flexibility, training stability and generation quality. On de novo small molecule generation, we find that the learned order dynamics leads to a significant increase in validity and quality of the generated molecules, when compared to uniform order dynamics.

","absKo":"

많은 domain에서 insertion을 통해 variable length sequence를 생성하는 방식은 autoregressive model보다 더 큰 유연성을 제공한다. 그러나 insertion model의 action space는 autoregressive model (ARM)의 것보다 훨씬 크기 때문에 학습이 어렵다. 이를 해결하기 위해 우리는 discrete flow matching의 target rate에 학습 가능한 order dynamics를 통합하고, 적절한 parameterization을 선택하면 수치 시뮬레이션 없이 target order dynamics와 generator의 joint training이 가능함을 보인다. generative insertion model로는 variable length masked diffusion model을 사용하며, 이 model은 mask token을 삽입하고 채워 넣는 방식으로 생성한다. 지역적으로 최적의 insertion order가 알려진 graph traversal task에서 우리는 parameterization 선택을 실험적으로 탐구하고, flexibility, training stability, generation quality 사이의 trade-off를 보여준다. de novo small molecule generation에서는 학습된 order dynamics가 uniform order dynamics와 비교할 때 생성된 molecule의 validity와 quality를 유의미하게 높인다는 것을 발견했다.

"},{"id":"62309","en":"Locally Coherent Parallel Decoding in Diffusion Language Models","ko":"Diffusion Language Models에서 국소 일관성 병렬 디코딩","authors":"Michael Hersche, Nicolas Menet, Ronan Tanios, Abbas Rahimi","pos":"#2807","link":"https://openreview.net/forum?id=hn4ndkaD6x","abs":"

Diffusion language models (DLMs) have emerged as a promising alternative to autoregressive (AR) models, offering sub-linear generation latency and bidirectional capabilities that are particularly appealing for code generation and editing. Achieving sub-linear latency in discrete DLMs requires predicting multiple tokens in parallel. However, standard DLMs sample tokens independently from conditional marginal distributions, failing to capture the joint dependencies among concurrently generated tokens. As a result, they often lead to syntactic inconsistencies and break multi-token structures. In this work, we introduce CoDiLA (Coherent Diffusion with Local Autoregression), a method that reconciles parallel sampling with local dependency modeling. Rather than forcing the DLM to resolve fine-grained syntax, CoDiLA delegates local decoding to a small, auxiliary AR model operating on the diffusion latents. This design allows for parallel block generation while ensuring sequential validity within each block and maintaining core DLM capabilities, including bidirectional modeling across blocks. We demonstrate that using a highly compact auxiliary AR model (e.g., 0.6B parameters) effectively eliminates coherence artifacts, establishing a new Pareto frontier for accuracy and speed in code generation benchmarks.

","absKo":"

Diffusion language model (DLM)은 autoregressive (AR) model의 유망한 대안으로 부상했으며, sub-linear generation latency와 bidirectional capability를 제공하여 특히 code generation과 editing에 매력적이다. discrete DLM에서 sub-linear latency를 달성하려면 여러 token을 병렬로 예측해야 한다. 그러나 표준 DLM은 조건부 marginal distribution에서 token을 독립적으로 sampling하므로, 동시에 생성되는 token 간의 joint dependency를 포착하지 못한다. 그 결과 종종 syntactic inconsistency를 유발하고 multi-token structure를 깨뜨린다. 본 연구에서는 병렬 sampling과 local dependency modeling을 조화시키는 CoDiLA (Coherent Diffusion with Local Autoregression)를 제안한다. CoDiLA는 DLM이 세밀한 syntax를 직접 해소하도록 강제하는 대신, diffusion latent에서 동작하는 작은 보조 AR model에게 local decoding을 맡긴다. 이 설계는 각 block 내부의 sequential validity를 보장하면서 block 간 bidirectional modeling을 포함한 핵심 DLM capability를 유지하는 동시에 병렬 block generation을 가능하게 한다. 우리는 매우 compact한 보조 AR model(예: 0.6B parameters)을 사용해 coherence artifact를 효과적으로 제거할 수 있음을 보이며, code generation benchmark에서 accuracy와 speed에 대한 새로운 Pareto frontier를 확립한다.

"},{"id":"65569","en":"MAD: Manifold Attracted Diffusion","ko":"MAD: 다양체 유인 Diffusion","authors":"Dennis Elbrächter, Giovanni S. Alberti, Matteo Santacesaria","pos":"#2808","link":"https://openreview.net/forum?id=CL6Byu4aIA","abs":"

Score-based diffusion models are a highly effective method for generating samples from a distribution of images. We consider scenarios where the training data comes from a noisy version of the target distribution, and present an efficiently implementable modification of the inference procedure to generate noiseless samples. Our approach is motivated by the manifold hypothesis, according to which meaningful data is concentrated around some low-dimensional manifold of a high-dimensional ambient space. The central idea is that noise manifests as low magnitude variation in off-manifold directions in contrast to the relevant variation of the desired distribution which is mostly confined to on-manifold directions. We introduce the notion of an extended score and show that, in a simplified setting, it can be used to reduce small variations to zero, while leaving large variations mostly unchanged. We describe how its approximation can be computed efficiently from an approximation to the standard score and demonstrate its efficacy on toy problems, synthetic data, and real data.

","absKo":"

Score-based diffusion model은 이미지 분포로부터 sample을 생성하는 데 매우 효과적인 방법이다. 우리는 training data가 target distribution의 noisy version에서 오는 상황을 고려하고, noiseless sample을 생성하기 위해 inference procedure를 효율적으로 구현 가능한 방식으로 수정한 방법을 제시한다. 우리의 접근은 manifold hypothesis에 의해 동기부여되는데, 이에 따르면 의미 있는 데이터는 고차원 ambient space의 저차원 manifold 주변에 집중되어 있다. 핵심 아이디어는 noise가 중요한 분포의 관련 변동이 주로 on-manifold direction에 국한되는 것과 대조적으로, off-manifold direction에서는 작은 크기의 변동으로 나타난다는 점이다. 우리는 extended score라는 개념을 도입하고, 단순화된 설정에서 이를 사용하면 큰 변동은 대부분 그대로 두면서 작은 변동을 0으로 줄일 수 있음을 보인다. 또한 standard score의 approximation으로부터 그 approximation을 효율적으로 계산하는 방법을 설명하고, toy problem, synthetic data, 실제 data에서 그 유효성을 입증한다.

"},{"id":"62907","en":"PADS-TAL: Padding-Annealed Diffusion Sampling in Text-Aware Latent Space for Robust and Diverse Text-to-Music Generation","ko":"PADS-TAL: Robust하고 다양한 Text-to-Music 생성을 위한 Text-Aware Latent Space의 Padding-Annealed Diffusion Sampling","authors":"Taekoan Yoo, Wonkyung Jung, Kyunghun Kim, Kyeongbo Kong","pos":"#2812","link":"https://openreview.net/forum?id=c0iisI5tJj","abs":"

Text-to-Music diffusion models are increasingly used in real-world applications, yet deployment remains challenging: generations can collapse to limited patterns even with diverse initial noise and prompts, and inference-time diversity control often harms text alignment and fidelity by distorting key prompt cues established in early denoising. To address this, we propose Padding-Annealed Diffusion Sampling, which perturbs only a padding-indexed subspace while keeping non-padding conditioning fixed, enabling controlled exploration with reduced semantic drift. However, in a text-unaware VAE latent space, such exploration is less likely to stay within genre-faithful neighborhoods, limiting genre-consistent diversity. We therefore introduce Text-Aware Latent space that aligns local neighborhoods with text-implied genre structure, promoting genre-consistent diversity. Together, the two techniques form a unified pipeline that, compared to prior methods that perturb the full conditioning, achieves a better text alignment--diversity trade-off: at comparable text alignment, it delivers 15.4\\% higher diversity with a relatively small fidelity drop, and further improves within-genre diversity by 71.6\\%. Generated samples are available at https://pads-tal.github.io/PADS-TAL.io

","absKo":"

Text-to-Music diffusion model은 실제 응용에서 점점 더 많이 사용되고 있지만, 배포는 여전히 어렵다. 다양한 초기 noise와 prompt가 주어져도 생성이 제한된 패턴으로 붕괴될 수 있으며, inference-time diversity control은 초반 denoising에서 형성된 핵심 prompt cue를 왜곡해 text alignment와 fidelity를 해치기 쉽다. 이를 해결하기 위해, 우리는 Padding-Annealed Diffusion Sampling을 제안한다. 이는 non-padding conditioning은 고정한 채 padding-indexed subspace만 perturb하여, semantic drift를 줄이면서 제어된 탐색을 가능하게 한다. 그러나 text-unaware VAE latent space에서는 이러한 탐색이 genre-faithful neighborhood 안에 머물 가능성이 낮아 genre-consistent diversity가 제한된다. 따라서 우리는 text가 암시하는 genre structure와 local neighborhood를 정렬하는 Text-Aware Latent space를 도입하여 genre-consistent diversity를 촉진한다. 이 두 기법을 결합한 unified pipeline은 full conditioning을 perturb하는 기존 방법들과 비교해 더 나은 text alignment-diversity trade-off를 달성한다. 유사한 text alignment에서 diversity를 15.4\\% 높이면서 fidelity 하락은 비교적 작고, genre 내부 diversity는 71.6\\% 추가 향상된다. 생성 샘플은 https://pads-tal.github.io/PADS-TAL.io 에서 확인할 수 있다

"},{"id":"61603","en":"Probing the Geometry of Diffusion Models with the String Method","ko":"String Method로 Diffusion Models의 Geometry 탐구","authors":"Elio Moreau, Florentin Coeurdoux, Grégoire Ferré, Eric Vanden-Eijnden","pos":"#2815","link":"https://openreview.net/forum?id=oWjx6SZVKn","abs":"

Understanding the geometry of learned distributions is fundamental to improving and interpreting diffusion models, yet systematic tools for exploring their landscape remain limited. Standard latent-space interpolations fail to respect the structure of the learned distribution, often traversing low-density regions. We introduce a framework based on the string method that computes continuous paths between samples by evolving curves under the learned score function. Operating on pretrained models without retraining, our approach interpolates between three regimes: pure generative transport, which yields continuous sample paths; gradient-dominated dynamics, which recover minimum energy paths (MEPs); and finite-temperature string dynamics, which compute principal curves---self-consistent paths that balance energy and entropy. We demonstrate that the choice of regime matters in practice. For image diffusion models, MEPs contain high-likelihood but unrealistic ``cartoon'' images, confirming prior observations that likelihood maxima appear unrealistic; principal curves instead yield realistic morphing sequences despite lower likelihood. For protein structure prediction, our method computes transition pathways between metastable conformers directly from models trained on static structures, yielding paths with physically plausible intermediates. Together, these results establish the string method as a principled tool for probing the modal structure of diffusion models---identifying modes, characterizing barriers, and mapping connectivity in complex learned distributions.

","absKo":"

학습된 분포의 기하를 이해하는 것은 diffusion model을 개선하고 해석하는 데 근본적이지만, 그 landscape를 탐색하기 위한 체계적인 도구는 여전히 제한적이다. 표준 latent-space interpolation은 학습된 분포의 구조를 존중하지 못해, 종종 low-density region을 가로지른다. 우리는 학습된 score function 아래에서 curve를 진화시켜 샘플 사이의 연속 경로를 계산하는 string method 기반 프레임워크를 제안한다. 재학습 없이 pretrained model에 적용되는 우리의 접근법은 세 가지 regime 사이를 보간한다: 연속적인 sample path를 산출하는 pure generative transport, minimum energy path (MEP)를 복원하는 gradient-dominated dynamics, 그리고 energy와 entropy의 균형을 이루는 self-consistent path인 principal curve를 계산하는 finite-temperature string dynamics이다. 우리는 선택한 regime가 실제로 중요함을 보인다. 이미지 diffusion model에서 MEP는 높은 likelihood를 가지지만 비현실적인 ``cartoon'' 이미지를 포함하여, likelihood 최대점이 비현실적으로 보인다는 기존 관찰을 확인한다. 반면 principal curve는 더 낮은 likelihood에도 불구하고 현실적인 morphing sequence를 산출한다. 단백질 구조 예측에서는, 우리의 방법이 정적 구조로 학습된 model에서 직접 metastable conformer 사이의 전이 경로를 계산하여, 물리적으로 타당한 intermediate를 갖는 경로를 생성한다. 종합하면, 이 결과들은 string method를 diffusion model의 modal 구조를 탐침하는 원리적 도구로 확립한다. 즉, mode를 식별하고, barrier를 특성화하며, 복잡한 학습 분포에서 연결성을 매핑할 수 있게 한다.

"},{"id":"64747","en":"Compositional Generative Modeling from Decentralized Data","ko":"분산 데이터로부터의 조합적 Generative Modeling","authors":"Mashrur M. Morshed, Vishnu Boddeti","pos":"#2906","link":"https://openreview.net/forum?id=KH2KhgpCcY","abs":"

The physical world is fundamentally compositional, yet empirical data are often fragmented across decentralized silos that cannot be aggregated due to privacy, legal, or economic constraints. Such scenarios pose a fundamental challenge for generative modeling: learning models that collectively cover the union of these sources while enabling compositional generalization when the factors required for composition are distributed across isolated data sources. We introduce Decentralized Compositional Flow Matching (DCFM), a framework for learning generative models from decentralized private data without exchanging raw samples. DCFM enforces structural constraints that induce conditional independence across the global set of generative factors. As a result, DCFM allows novel combinations to emerge through interactions across peers, even when no single data source contains sufficient information to support composition on its own. Empirically, DCFM substantially outperforms federated learning and mixture-of-experts baselines across conditional image generation, robotic spatial planning and medical attribute co-occurrence modeling.

","absKo":"

물리적 세계는 본질적으로 compositional하지만, 실제 empirical data는 privacy, 법적 제약, 경제적 제약 때문에 통합할 수 없는 분산된 silo에 종종 분절되어 있다. 이러한 상황은 generative modeling에 근본적인 도전을 제기한다. 즉, composition에 필요한 factor가 서로 고립된 data source에 분포되어 있을 때도, 이들의 합집합을 collectively cover하면서 compositional generalization을 가능하게 하는 model을 학습해야 한다. 우리는 원시 sample을 교환하지 않고 분산된 private data로부터 generative model을 학습하기 위한 framework인 Decentralized Compositional Flow Matching (DCFM)을 제안한다. DCFM은 global generative factor 집합 전반에 걸쳐 conditional independence를 유도하는 structural constraint를 강제한다. 그 결과, 어떤 단일 data source도 자기 혼자서는 composition을 지원할 충분한 정보를 담고 있지 않더라도, peer 간 상호작용을 통해 새로운 조합이 나타날 수 있다. 실험적으로 DCFM은 conditional image generation, robotic spatial planning, medical attribute co-occurrence modeling 전반에서 federated learning 및 mixture-of-experts baseline보다 상당히 뛰어난 성능을 보인다.

"},{"id":"61978","en":"Cascaded Flow Matching for Heterogeneous Tabular Data with Mixed-Type Features","ko":"혼합형 특성을 가진 이종 tabular 데이터의 cascaded flow matching","authors":"Markus Mueller, Kathrin Gruber, Dennis Fok","pos":"#2907","link":"https://openreview.net/forum?id=l2ywV9sV0L","abs":"

Advances in generative modeling have recently been adapted to tabular data containing discrete and continuous features. However, generating mixed-type features that combine discrete states with an otherwise continuous distribution in a single feature remains challenging. We advance the state-of-the-art in diffusion models for tabular data with a cascaded approach. We first generate a low-resolution version of a tabular data row, that is, the collection of the purely categorical features and a coarse categorical representation of numerical features. Next, this information is leveraged in the high-resolution flow matching model via a novel guided conditional probability path and data-dependent coupling. The low-resolution representation of numerical features explicitly accounts for discrete outcomes, such as missing or inflated values, and therewith enables a more faithful generation of mixed-type features. We formally prove that this cascade tightens the transport cost bound. The results indicate that our model generates significantly more realistic samples and captures distributional details more accurately, for example, the detection score increases by 40\\%.

","absKo":"

생성 모델링의 발전은 최근 discrete feature와 continuous feature를 모두 포함하는 tabular data에 적용되고 있다. 그러나 하나의 feature 안에서 discrete state와 그 외의 continuous distribution이 결합된 mixed-type feature를 생성하는 것은 여전히 어렵다. 우리는 cascaded approach를 통해 tabular data를 위한 diffusion model의 state-of-the-art를 발전시킨다. 먼저 tabular data row의 low-resolution 버전을 생성하는데, 이는 순수 categorical feature들의 집합과 numerical feature들의 거친 categorical representation이다. 다음으로, 이 정보는 novel guided conditional probability path와 data-dependent coupling을 통해 high-resolution flow matching model에서 활용된다. numerical feature의 low-resolution representation은 missing value나 inflated value와 같은 discrete outcome을 명시적으로 고려하며, 이를 통해 mixed-type feature를 보다 충실하게 생성할 수 있다. 우리는 이 cascade가 transport cost bound를 더 타이트하게 만든다는 것을 형식적으로 증명한다. 결과는 우리의 모델이 훨씬 더 현실적인 sample을 생성하고 분포의 세부를 더 정확히 포착함을 보여주며, 예를 들어 detection score가 40\\% 증가한다.

"},{"id":"63949","en":"Beyond Generative Priors: Minority Sampling with JEPA-Guided Diffusion","ko":"Generative prior를 넘어서: JEPA-guided diffusion을 이용한 소수 표본 추출","authors":"Sol Park, Soobin Um","pos":"#2908","link":"https://openreview.net/forum?id=SHLRiQAlMp","abs":"

Minority sampling aims to generate low-density instances on a data manifold and is of central importance in applications such as medical diagnosis, anomaly detection, and creative AI. Existing approaches, however, define minority samples relative to generative priors learned from training data, confining rarity to model-specific notions that may poorly reflect real-world semantics. In this work, we propose a world-centric perspective on minority sampling, which defines rarity with respect to real-world priors rather than generator-induced densities. To this end, we introduce JEPA guidance, a diffusion sampling framework guided by a Joint-Embedding Predictive Architecture (JEPA)—a class of world models that encode broad, semantically rich representations. JEPA guidance steers diffusion trajectories toward low-density regions under the implicit density induced by the JEPA, thereby aligning generated minorities with real-world semantic rarity. To make JEPA guidance computationally practical, we develop principled approximation strategies accompanied by theoretical error bounds, significantly reducing the overhead of guidance computation. Extensive experiments across unconditional, class-conditional, and text-to-image generation demonstrate that JEPA guidance consistently improves the fidelity and semantic validity of minority samples, outperforming generator-centric baselines in capturing real-world notions of rarity.

","absKo":"

Minority sampling은 data manifold 상에서 low-density instance를 생성하는 것을 목표로 하며, medical diagnosis, anomaly detection, creative AI 같은 응용에서 핵심적으로 중요하다. 그러나 기존 접근은 학습 데이터로부터 학습된 generative prior를 기준으로 minority sample을 정의하므로, 희소성을 모델 특유의 개념으로 제한하며 실제 세계의 의미론을 제대로 반영하지 못할 수 있다. 본 연구에서는 real-world prior를 기준으로 희소성을 정의하는 world-centric perspective의 minority sampling을 제안한다. 이를 위해 우리는 Joint-Embedding Predictive Architecture (JEPA)로 안내되는 diffusion sampling framework인 JEPA guidance를 도입한다. JEPA는 넓고 의미적으로 풍부한 representation을 인코딩하는 world model의 한 계열이다. JEPA guidance는 JEPA가 유도하는 implicit density 아래에서 diffusion trajectory를 low-density region으로 유도함으로써, 생성된 minority를 real-world semantic rarity와 정렬시킨다. JEPA guidance를 계산적으로 실용적으로 만들기 위해, 우리는 이론적 error bound를 동반한 원리 기반 approximation strategy를 개발하여 guidance 계산의 오버헤드를 크게 줄였다. unconditional, class-conditional, text-to-image generation 전반에 걸친 광범위한 실험은 JEPA guidance가 minority sample의 fidelity와 semantic validity를 일관되게 향상시키며, real-world의 희소성 개념을 포착하는 데 있어 generator-centric baseline보다 우수함을 보인다.

"},{"id":"60918","en":"Adapting Noise to Data: Generative Flows from Learned 1D Processes","ko":"데이터에 Noise 적응하기: 학습된 1D Process로부터의 Generative Flows","authors":"Jannis Chemseddine, Gregor Kornhardt, Richard Duong, Gabriele Steidl","pos":"#2911","link":"https://openreview.net/forum?id=vLQO6nrpYq","abs":"

The default Gaussian latent in flow-based generative models poses challenges when learning certain distributions such as heavy-tailed ones. We introduce a general framework for learning data-adaptive latent distributions using one-dimensional quantile functions, optimized via the Wasserstein distance between noise and data. The quantile-based parameterization naturally adapts to both heavy-tailed and compactly supported distributions and shortens transport paths. Numerical results confirm the method’s flexibility and effectiveness achieved with negligible computational overhead.

","absKo":"flow-based generative model에서 기본 Gaussian latent는 heavy-tailed distribution과 같은 특정 분포를 학습할 때 어려움을 야기한다. 우리는 noise와 data 사이의 Wasserstein distance를 통해 최적화되는 1차원 quantile function을 사용하여, data-adaptive latent distribution을 학습하는 일반적인 프레임워크를 제안한다. quantile 기반 parameterization은 heavy-tailed 분포와 compactly supported 분포 모두에 자연스럽게 적응하며 transport path를 단축한다. 수치 실험 결과는 이 방법이 거의 추가 계산 비용 없이 유연성과 효과성을 달성함을 확인해 준다."},{"id":"63710","en":"Manifold-Aware Perturbations for Constrained Generative Modeling","ko":"제약된 생성 모델링을 위한 Manifold 인지 Perturbation","authors":"Katherine Keegan, Lars Ruthotto","pos":"#407","link":"https://openreview.net/forum?id=UPiyC9W4ms","abs":"

Generative models have enjoyed widespread success in a variety of applications. However, they encounter inherent mathematical limitations in modeling distributions where samples are constrained by equalities, as is frequently the setting in scientific domains. In this work, we develop a computationally cheap, mathematically justified, and highly flexible distributional modification for combating known pitfalls in equality-constrained generative models. We propose perturbing the data distribution in a constraint-aware way such that the new distribution has support matching the ambient space dimension while still implicitly incorporating underlying manifold geometry. Through theoretical analyses and empirical evidence on several representative tasks, we illustrate that our approach consistently enables data distribution recovery and stable sampling with both diffusion models and normalizing flows.

","absKo":"생성 모델은 다양한 응용에서 널리 성공을 거두어 왔다. 그러나 scientific domain에서 자주 나타나듯이, sample이 equality로 제약되는 분포를 모델링할 때는 본질적인 수학적 한계에 직면한다. 본 연구에서는 equality-constrained generative model에서 알려진 함정을 완화하기 위해 계산 비용이 낮고, 수학적으로 정당화되며, 매우 유연한 distributional modification을 개발한다. 우리는 제약을 인지하는 방식으로 data distribution을 perturb하여, 새로운 분포의 support가 ambient space 차원과 일치하면서도 underlying manifold geometry를 암묵적으로 포함하도록 제안한다. 이론적 분석과 여러 대표적 task에 대한 실증적 증거를 통해, 우리의 접근이 diffusion model과 normalizing flow 모두에서 일관되게 data distribution recovery와 안정적인 sampling을 가능하게 함을 보인다.

"},{"id":"61913","en":"Navigating the Energy Landscape of Collaboration: Multi-Agent Communication Graph Generation via Score-Based Diffusion","ko":"협업의 에너지 풍경 탐색: Score-Based Diffusion 기반 Multi-Agent 통신 그래프 생성","authors":"GuanHao Zhao, Wenbo Lu, Cheng Cheng, Zhenya Huang, Wei Song, Zhiding Liu, Runze Wu, Enhong Chen","pos":"#2112","link":"https://openreview.net/forum?id=lhSg9cJUKY","abs":"

The collective intelligence of Large Language Model (LLM)-based Multi-Agent Systems (MAS) is fundamentally governed by the underlying communication graph. However, discovering task-adaptive structures within this combinatorial search space remains a significant challenge. Existing methods, ranging from heuristic pruning to autoregressive generation, often lack a unified theoretical framework to guide the self-organization of agents into efficient teams. In this paper, we bridge non-equilibrium thermodynamics and generative modeling to formalize multi-agent graph generation as an energy minimization process. Specifically, we frame the emergence of efficient collaboration as a thermodynamic \"cooling\" process, where initially stochastic interactions converge to a low-energy, structured equilibrium. To implement this, We propose MAGE (Multi-Agent Communication Graph Generation), a score-based diffusion framework that constructs communication graphs by navigating the energy landscape via iterative denoising and first-order gradient guidance. Extensive experiments on representative benchmarks demonstrate that MAGE achieves state-of-the-art performance. Furthermore, qualitative analysis reveals that the generated graphs mirroring the functional specialization of human organizations, validating our thermodynamic hypothesis.

","absKo":"

Large Language Model (LLM) 기반 Multi-Agent Systems (MAS)의 집합 지능은 근본적으로 기반이 되는 communication graph에 의해 지배된다. 그러나 이 조합적 탐색 공간에서 task-adaptive한 구조를 발견하는 일은 여전히 중요한 도전 과제다. 휴리스틱 pruning부터 autoregressive generation에 이르는 기존 방법들은 agents가 효율적인 팀으로 self-organization하도록 이끄는 통합된 이론적 틀을 대체로 갖추지 못한다. 이 논문에서는 non-equilibrium thermodynamics와 generative modeling을 연결하여 multi-agent graph generation을 energy minimization process로 정식화한다. 구체적으로, 효율적인 collaboration의 출현을 thermodynamic한 \"cooling\" process로 해석하며, 초기의 stochastic interaction이 낮은 에너지의 구조화된 equilibrium으로 수렴한다고 본다. 이를 구현하기 위해, 우리는 MAGE (Multi-Agent Communication Graph Generation)를 제안한다. 이는 iterative denoising과 first-order gradient guidance를 통해 energy landscape를 탐색하며 communication graph를 구성하는 score-based diffusion framework이다. 대표 benchmark들에서 수행한 광범위한 실험은 MAGE가 state-of-the-art 성능을 달성함을 보여준다. 더 나아가 정성적 분석은 생성된 graph가 인간 조직의 functional specialization을 반영함을 드러내어, 우리의 thermodynamic hypothesis를 검증한다.

"},{"id":"61719","en":"Structured Diffusion Bridges: Inductive Bias for Denoising Diffusion Bridges","ko":"Structured Diffusion Bridges: Denoising Diffusion Bridge를 위한 Inductive Bias","authors":"Eitan Kosman, Gabriele Serussi, Chaim Baskin","pos":"#2711","link":"https://openreview.net/forum?id=nWI6lV7BvA","abs":"

Modality translation is inherently under-constrained, as multiple cross-modal mappings may yield the same marginals. Recent work has shown that diffusion bridges are effective for this task. However, most existing approaches rely on fully paired datasets, thereby imposing a single data-driven constraint. We propose a diffusion-bridge framework that characterizes the space of admissible solutions and restricts it via alignment constraints, treating paired supervision as an optional heuristic rather than a prerequisite. We validate our method on synthetic and real modality translation benchmarks across unpaired, semi-paired, and paired regimes, showing consistent performance across supervision levels. Notably, it achieves near fully-paired quality with a substantial relaxation in pairing requirements, and remaining applicable in the unpaired regime. These results highlight diffusion bridges as a flexible foundation for modality translation beyond fully paired data.

","absKo":"

Modality translation은 여러 cross-modal mapping이 동일한 marginal을 만들어낼 수 있기 때문에 본질적으로 under-constrained이다. 최근 연구는 diffusion bridge가 이 작업에 효과적임을 보여주었다. 그러나 대부분의 기존 접근법은 완전히 paired된 dataset에 의존하여 단일한 data-driven constraint를 부과한다. 우리는 허용 가능한 해의 공간을 특성화하고 alignment constraint로 이를 제한하는 diffusion-bridge framework를 제안하며, paired supervision을 필수 조건이 아니라 선택적 heuristic으로 다룬다. 우리는 unpaired, semi-paired, paired regime에 걸친 synthetic 및 real modality translation benchmark에서 우리의 방법을 검증하였고, supervision level 전반에 걸쳐 일관된 성능을 보였다. 특히 pairing 요구를 크게 완화하면서도 fully-paired에 가까운 품질을 달성했으며, unpaired regime에서도 여전히 적용 가능하다. 이러한 결과는 diffusion bridge가 fully paired data를 넘어서는 modality translation의 유연한 기반임을 보여준다.

"},{"id":"68808","en":"G2D2: Gradient-Guided Discrete Diffusion for Inverse Problem Solving","ko":"G2D2: 역문제 해결을 위한 Gradient-Guided Discrete Diffusion","authors":"Naoki Murata, Chieh-Hsin Lai, Yuhta Takida, Toshimitsu Uesaka, Bac Nguyen, Stefano Ermon, Yuki Mitsufuji","pos":"#2801","abs":"

Recent literature has effectively leveraged diffusion models trained on continuous variables as priors for solving inverse problems. Notably, discrete diffusion models with discrete latent codes have shown strong performance, particularly in modalities suited for discrete compressed representations, such as image and motion generation. However, their discrete and non-differentiable nature has limited their application to inverse problems formulated in continuous spaces. This paper presents a novel method for addressing linear inverse problems by leveraging generative models based on discrete diffusion as priors. We overcome these limitations by approximating the true posterior distribution with a variational distribution constructed from categorical distributions and continuous relaxation techniques. Furthermore, we employ a star-shaped noise process to mitigate the drawbacks of traditional discrete diffusion models with absorbing states, demonstrating that our method performs comparably to continuous diffusion techniques with lower GPU memory consumption.

","absKo":"

최근 문헌은 연속 변수로 학습된 diffusion model을 inverse problem 해결을 위한 prior로 효과적으로 활용해 왔다. 특히 discrete latent code를 사용하는 discrete diffusion model은, image 및 motion generation처럼 이산 압축 표현에 적합한 모달리티에서 강한 성능을 보여 왔다. 그러나 그 이산적이고 non-differentiable한 성질 때문에, 연속 공간에서 정식화된 inverse problem에는 적용이 제한되어 왔다. 이 논문은 discrete diffusion 기반 generative model을 prior로 활용하여 linear inverse problem을 해결하는 새로운 방법을 제시한다. 우리는 categorical distribution과 continuous relaxation technique로 구성한 variational distribution을 사용해 true posterior distribution을 근사함으로써 이러한 한계를 극복한다. 더 나아가, absorbing state를 갖는 기존 discrete diffusion model의 단점을 완화하기 위해 star-shaped noise process를 사용하며, 제안 방법이 더 낮은 GPU memory consumption으로 continuous diffusion 기법에 비견되는 성능을 보임을 입증한다.

"},{"id":"62959","en":"IDLM: Inverse-distilled Diffusion Language Models","ko":"IDLM: 역증류 Diffusion 언어 모델","authors":"David Li, Nikita Gushchin, Dmitry Abulkhanov, Eric Moulines, Ivan Oseledets, Maxim Panov, Aleksandr Korotin","pos":"#2802","link":"https://openreview.net/forum?id=bS8qU2wJxu","abs":"Diffusion Language Models (DLMs) have recently achieved strong results in text generation. However, their multi-step sampling leads to slow inference, limiting practical use. To address this, we extend Inverse Distillation, a technique originally developed to accelerate continuous diffusion models, to the discrete setting. Nonetheless, this extension introduces both theoretical and practical challenges. From a theoretical perspective, the inverse distillation objective lacks uniqueness guarantees, which may lead to suboptimal solutions. From a practical standpoint, backpropagation in the discrete space is non-trivial and often unstable. To overcome these challenges, we first provide a theoretical result demonstrating that our inverse formulation admits a unique solution, thereby ensuring valid optimization. We then introduce gradient-stable relaxations to support effective training. As a result, experiments on multiple DLMs show that our method, *Inverse-distilled Diffusion Language Models (IDLM)*, reduces the number of inference steps by $4 \\times$-$64 \\times$, while preserving the teacher model’s entropy and generative perplexity.","absKo":"Diffusion Language Models (DLMs)는 최근 text generation에서 강한 성능을 달성했다. 그러나 다단계 sampling 때문에 inference가 느려 실제 사용이 제한된다. 이를 해결하기 위해 우리는 원래 continuous diffusion model의 가속을 위해 개발된 Inverse Distillation을 discrete setting으로 확장한다. 하지만 이 확장은 이론적, 실용적 도전을 모두 야기한다. 이론적으로는 inverse distillation objective에 유일성 보장이 없어 최적이 아닌 해로 수렴할 수 있다. 실용적으로는 discrete space에서의 backpropagation이 비자명하고 종종 불안정하다. 이러한 문제를 극복하기 위해, 우리는 먼저 우리의 inverse formulation이 유일한 해를 갖는다는 이론적 결과를 제시하여 올바른 optimization을 보장한다. 그다음 효과적인 학습을 지원하기 위해 gradient-stable relaxation을 도입한다. 그 결과 여러 DLM에 대한 실험에서 우리의 방법인 *Inverse-distilled Diffusion Language Models (IDLM)*은 teacher model의 entropy와 generative perplexity를 보존하면서 inference step 수를 $4 \\times$-$64 \\times$ 줄였다."},{"id":"61598","en":"Is Your Diffusion Sampler Actually Correct? A Sampler-Centric Evaluation of Discrete Diffusion Language Models","ko":"당신의 Diffusion Sampler는 정말 올바른가? Discrete Diffusion Language Models의 Sampler-Centric Evaluation","authors":"Luhan Tang, Longxuan Yu, Shaorong Zhang, Greg Ver Steeg","pos":"#2806","link":"https://openreview.net/forum?id=oYCCOl7ahf","abs":"

Discrete diffusion language models (dLLMs) offer a fast and flexible alternative to autoregressive models (ARMs) for discrete sequence generation by performing iterative denoising with parallel updates. Despite these advantages, dLLMs are commonly evaluated using metrics developed for ARMs. Such evaluations rely on metrics computed from final generated samples and conflate model approximation error in the learned denoiser and sampler-induced error from the sampling dynamics. We introduce a sampler-centric oracle evaluation framework that replaces learned denoisers with an oracle Hidden Markov Model posterior derived from a ground-truth Markov chain, enabling isolation of sampler-induced error under controlled and method-consistent settings. We show that few-step discrete diffusion samplers are not distributionally correct, even under an exact oracle denoiser, with substantial distributional mismatch at the transition level persisting at small step counts and vanishing only when the number of diffusion steps approaches the sequence length. We also find that current metrics for evaluating dLLMs are insufficient: improvements in negative log-likelihood, generative perplexity, or MAUVE do not imply correct sampling.

","absKo":"

Discrete diffusion language model (dLLM)은 반복적 denoising과 parallel update를 통해 discrete sequence generation을 수행함으로써, autoregressive model (ARM)에 대한 빠르고 유연한 대안을 제공한다. 이러한 장점에도 불구하고 dLLM은 흔히 ARM을 위해 개발된 metric으로 평가된다. 이러한 평가는 최종 생성 샘플에서 계산된 metric에 의존하며, 학습된 denoiser의 model approximation error와 sampling dynamics에서 기인하는 sampler-induced error를 혼동한다. 우리는 learned denoiser를 ground-truth Markov chain에서 도출한 oracle Hidden Markov Model posterior로 대체하는 sampler-centric oracle evaluation framework를 도입하여, 통제되고 방법 일관적인 설정에서 sampler-induced error를 분리한다. 우리는 몇 단계의 discrete diffusion sampler가, 정확한 oracle denoiser가 주어져도, 분포적으로 올바르지 않다는 것을 보이며, 작은 step 수에서는 transition level의 상당한 distributional mismatch가 지속되고 diffusion step 수가 sequence length에 가까워질 때에만 사라짐을 보인다. 또한 dLLM 평가를 위한 현재 metric은 충분하지 않다는 점도 발견했다. negative log-likelihood, generative perplexity, 또는 MAUVE의 개선이 올바른 sampling을 의미하지는 않는다.

"},{"id":"64182","en":"Measurement-Consistent Langevin Corrector for Stabilizing Latent Diffusion Inverse Problem Solvers","ko":"잠재 Diffusion 역문제 해결기 안정화를 위한 측정 일관 Langevin 보정기","authors":"Hyoseok Lee, Sohwi Lim, Eunju Cha, Tae-Hyun Oh","pos":"#2809","link":"https://openreview.net/forum?id=QC7fOKv1jg","abs":"

While latent diffusion models (LDMs) have emerged as powerful priors for inverse problems, existing LDM-based solvers frequently suffer from instability. In this work, we first identify the instability as a discrepancy between the solver dynamics and stable reverse diffusion dynamics learned by the diffusion model, and show that reducing this gap stabilizes the solver. Building on this, we introduce Measurement-Consistent Langevin Corrector (MCLC), a theoretically grounded plug-and-play stabilization module that remedies the LDM-based inverse problem solvers through measurement-consistent Langevin updates. Compared to prior approaches that rely on linear manifold assumptions, which often fail to hold in latent space, MCLC provides a principled stabilization mechanism, leading to more stable and reliable behavior in latent space.

","absKo":"

latent diffusion model (LDM)은 inverse problem을 위한 강력한 prior로 부상했지만, 기존의 LDM 기반 solver는 종종 불안정성 문제를 겪는다. 본 연구에서는 먼저 이러한 불안정성을 solver dynamics와 diffusion model이 학습한 stable reverse diffusion dynamics 간의 불일치로 식별하고, 이 차이를 줄이면 solver가 안정화됨을 보인다. 이를 바탕으로 우리는 measurement-consistent Langevin update를 통해 LDM 기반 inverse problem solver의 문제를 완화하는, 이론적으로 타당한 plug-and-play stabilization module인 Measurement-Consistent Langevin Corrector (MCLC)를 제안한다. latent space에서 자주 성립하지 않는 linear manifold assumption에 의존하는 기존 접근법과 비교할 때, MCLC는 원리적인 stabilization mechanism을 제공하여 latent space에서 더 안정적이고 신뢰할 수 있는 거동을 이끈다.

"},{"id":"61116","en":"Particle-Guided Diffusion Models for Partial Differential Equations","ko":"편미분방정식을 위한 Particle-Guided Diffusion Model","authors":"Andrew Millard, Fredrik Lindsten, Zheng Zhao","pos":"#2813","link":"https://openreview.net/forum?id=tX1k4GQtLC","abs":"

We introduce a guided stochastic sampling method that augments sampling from diffusion models with physics-based guidance derived from partial differential equation (PDE) residuals and observational constraints, ensuring generated samples remain physically admissible. We embed this sampling procedure within a new Sequential Monte Carlo (SMC) framework, yielding a scalable generative PDE solver. Across multiple benchmark PDE systems as well as multiphysics and interacting PDE systems, our method produces solution fields with lower numerical error than existing state-of-the-art generative methods.

","absKo":"

우리는 diffusion model에서의 sampling에 PDE residual과 observational constraint로부터 유도된 physics-based guidance를 결합하는 guided stochastic sampling method를 소개한다. 이를 통해 생성된 sample이 물리적으로 허용 가능한 상태로 유지되도록 한다. 우리는 이 sampling procedure를 새로운 Sequential Monte Carlo (SMC) framework 안에 통합하여, 확장 가능한 generative PDE solver를 구현한다. 여러 benchmark PDE system은 물론 multiphysics 및 interacting PDE system 전반에서, 우리의 방법은 기존 state-of-the-art generative method보다 더 낮은 numerical error를 가진 solution field를 생성한다.

"},{"id":"66597","en":"Fine-Tuning Masked Diffusion for Provable Self-Correction","ko":"증명 가능한 자기 교정을 위한 Masked Diffusion의 Fine-Tuning","authors":"Jaeyeon Kim, Seunggeun Kim, Taekyun Lee, David Pan, Hyeji Kim, Sham Kakade, Sitan Chen","pos":"#2900","link":"https://openreview.net/forum?id=2LRaNghfoM","abs":"

A natural desideratum for generative models is \\emph{self-correction}--detecting and revising low-quality tokens at inference. While Masked Diffusion Models (MDMs) have emerged as a promising approach for generative modeling in discrete spaces, their capacity for self-correction remains poorly understood. Prior attempts to incorporate self-correction into MDMs either require overhauling MDM architectures/training or rely on imprecise proxies for token quality, limiting their applicability. Motivated by this, we introduce PRISM--Plug-in Remasking for Inference-time Self-correction of Masked Diffusions--a lightweight, model-agnostic approach that applies to any pretrained MDM. Theoretically, PRISM defines a self-correction loss that provably learns per-token quality scores, without RL or a verifier. These quality scores are computed in the same forward pass with MDM and used to detect low-quality tokens. Empirically, PRISM advances MDM inference across domains and scales: Sudoku; unconditional text (170M); and code with LLaDA (8B).

","absKo":"생성 모델에 대한 자연스러운 desideratum은 \\emph{self-correction}이다. 즉, inference 시 낮은 품질의 token을 탐지하고 수정하는 것이다. Masked Diffusion Model(MDM)은 discrete space에서의 생성 모델링에 대한 유망한 접근법으로 부상했지만, self-correction 능력은 아직 충분히 이해되지 않았다. MDM에 self-correction을 통합하려는 기존 시도는 MDM architecture와 training을 전면 개편해야 하거나, token 품질에 대한 부정확한 proxy에 의존해야 했기 때문에 적용 가능성이 제한되었다. 이러한 동기에 따라 우리는 PRISM--Plug-in Remasking for Inference-time Self-correction of Masked Diffusions--을 소개한다. 이는 사전 학습된 어떤 MDM에도 적용 가능한 경량의 model-agnostic 접근법이다. 이론적으로 PRISM은 RL이나 verifier 없이도 token별 품질 점수를 올바르게 학습하는 self-correction loss를 정의한다. 이 품질 점수는 MDM과 동일한 forward pass에서 계산되며 낮은 품질의 token을 탐지하는 데 사용된다. 실험적으로 PRISM은 Sudoku, unconditional text(170M), 그리고 LLaDA(8B)를 사용한 code 전반에서 MDM 추론을 발전시킨다.

"},{"id":"66143","en":"Efficient Diffusion Models under Nonconvex Equality and Inequality constraints via Landing","ko":"Landing을 통한 비볼록 등식 및 부등식 제약 하의 효율적인 Diffusion Model","authors":"Kijung Jeon, Michael Muehlebach, Molei Tao","pos":"#2902","link":"https://openreview.net/forum?id=6VEqdKmfKh","abs":"Generative modeling within constrained sets is essential for scientific and engineering applications involving physical, geometric, or safety requirements (e.g., molecular generation, robotics). We present a unified framework for constrained diffusion models on generic nonconvex feasible sets $\\Sigma$ that simultaneously enforces equality and inequality constraints throughout the diffusion process. Our framework incorporates both overdamped and underdamped dynamics for forward and backward sampling. A key algorithmic innovation is a computationally efficient landing mechanism that replaces costly and often ill-defined projections onto $\\Sigma$, ensuring feasibility without iterative Newton solves or projection failures. By leveraging underdamped dynamics, we accelerate mixing toward the prior distribution, effectively alleviating the high simulation costs typically associated with constrained diffusion. Empirically, this approach reduces function evaluations and memory usage during both training and inference while preserving sample quality. On benchmarks featuring equality and mixed constraints, our method achieves comparable sample quality to state-of-the-art baselines while significantly reducing computational cost, providing a practical and scalable solution for diffusion on nonconvex feasible sets.","absKo":"제약 집합 내 generative modeling은 물리적, 기하학적 또는 안전 요구사항(예: molecular generation, robotics)을 포함하는 과학 및 공학 응용에서 필수적이다. 우리는 일반적인 비볼록 feasible set $\\Sigma$ 위의 constrained diffusion model을 위한 통합 프레임워크를 제시하며, diffusion 과정 전반에서 equality constraint와 inequality constraint를 동시에 강제한다. 우리의 프레임워크는 forward 및 backward sampling 모두에 대해 overdamped와 underdamped dynamics를 포함한다. 핵심 알고리즘 혁신은 $\\Sigma$에 대한 비용이 크고 종종 잘 정의되지도 않은 projection을 대체하는 계산적으로 효율적인 landing mechanism으로, 반복적인 Newton solve나 projection failure 없이 feasibility를 보장한다. underdamped dynamics를 활용함으로써 우리는 prior distribution으로의 mixing을 가속하고, constrained diffusion에 일반적으로 수반되는 높은 simulation cost를 효과적으로 완화한다. 경험적으로 이 접근은 sample quality를 유지하면서 training과 inference 모두에서 function evaluation 수와 memory 사용량을 줄인다. equality 및 mixed constraint를 포함한 benchmark에서 우리의 방법은 최첨단 baseline과 동등한 sample quality를 달성하면서 계산 비용을 크게 줄여, 비볼록 feasible set 위의 diffusion을 위한 실용적이고 확장 가능한 해법을 제공한다.

"},{"id":"62822","en":"CountsDiff: A diffusion model on the natural numbers for generation and imputation of count-based data","ko":"CountsDiff: count 기반 데이터의 생성 및 보간을 위한 자연수 상의 diffusion model","authors":"Renzo Soatto, Anders Hoel, Greycen Ren, Shorna Alam, Stephen Bates, Nikolaos Daskalakis, Caroline Uhler, Maria Skoularidou","pos":"#2904","link":"https://openreview.net/forum?id=cvlPAWSeo3","abs":"

Diffusion models have excelled at generative tasks for both continuous and token-based domains, but their application to discrete ordinal data remains underdeveloped. We present \\emph{CountsDiff}, a diffusion framework designed to natively model distributions on the natural numbers. CountsDiff extends the Blackout diffusion framework by simplifying its formulation through a direct parameterization in terms of a survival probability schedule and an explicit loss weighting. This introduces flexibility through design parameters with direct analogues in existing diffusion modeling frameworks. Beyond this reparameterization, CountsDiff introduces features from modern diffusion models, previously absent in counts-based domains, including continuous-time training, classifier-free guidance, and churn/remasking reverse dynamics that allow non-monotone reverse trajectories. We propose an initial instantiation of CountsDiff and validate it on natural image datasets (CIFAR-10, CelebA), demonstrating the benefits of the proposed design space and that the framework scales to complex, high-dimensional data domains. We then highlight biological count assays as a natural use case, evaluating CountsDiff on single-cell RNA-seq imputation in a fetal cell and heart cell atlas. Remarkably, we find that even this simple instantiation matches or surpasses the performance of a state-of-the-art discrete generative model and leading RNA-seq imputation methods, while leaving substantial headroom for further gains through optimized design choices in future work.

","absKo":"

Diffusion model은 연속형 및 token 기반 도메인 모두에서 생성 task에 뛰어난 성능을 보여 왔지만, discrete ordinal data에 대한 적용은 아직 충분히 발전하지 않았다. 우리는 자연수 분포를 직접 모델링하도록 설계된 diffusion framework인 \\emph{CountsDiff}를 제시한다. CountsDiff는 survival probability schedule과 명시적 loss weighting을 통한 직접 parameterization으로 formulation을 단순화함으로써 Blackout diffusion framework를 확장한다. 이는 기존 diffusion modeling framework와 직접 대응되는 design parameter를 통해 유연성을 제공한다. 이러한 reparameterization을 넘어, CountsDiff는 counts-based domain에서는 이전에 없던 최신 diffusion model의 요소들을 도입한다. 여기에는 continuous-time training, classifier-free guidance, 그리고 비단조 reverse trajectory를 허용하는 churn/remasking reverse dynamics가 포함된다. 우리는 CountsDiff의 초기 구현을 제안하고 이를 natural image dataset (CIFAR-10, CelebA)에서 검증하여 제안한 design space의 이점을 보여주며, 이 framework가 복잡하고 고차원적인 data domain으로 확장됨을 입증한다. 이어서 biological count assay를 자연스러운 사용 사례로 제시하고, fetal cell 및 heart cell atlas에서 single-cell RNA-seq imputation에 대해 CountsDiff를 평가한다. 놀랍게도, 이 단순한 구현만으로도 state-of-the-art discrete generative model과 선도적인 RNA-seq imputation method의 성능에 필적하거나 이를 능가했으며, 향후 최적화된 design choice를 통해 더 큰 향상을 얻을 수 있는 상당한 여지를 남겼다.

"},{"id":"60502","en":"Condition-Aware Graph Flow Matching for Modeling the Distributions of Complex Physical Systems","ko":"복잡한 물리 시스템 분포 모델링을 위한 Condition-Aware Graph Flow Matching","authors":"Xiaochao Deng, Jie Chen, Xiaogang Deng","pos":"#2905","link":"https://openreview.net/forum?id=zclW70fItC","abs":"

Accurately modeling the full distributions of possible states is crucial for understanding statistical properties and enabling reliable predictions in complex, unsteady physical systems. Recently, diffusion models and flow matching have shown promise in these tasks. However, they remain limited in uncovering the general principles of systems from multiple short trajectories across condition space. In addition, they exhibit inferior adaptability to large irregular geometries, particularly in regions with sharp gradients. In this paper, we propose a condition-aware graph flow matching (CGFM) method that combines condition-aware flow matching with a hierarchical graph structure to learn the full distributions of physical systems from incomplete training data. Specifically, CGFM constructs a flow enabling smooth interpolation across physical conditions and parameterizes the graph-conditioned vector field through HieraGraphNet. HieraGraphNet performs message passing across multilevel graphs to capture multi-scale dynamics and facilitate long-range information interactions in physical systems. Moreover, we introduce a topology- and geometry-aware graph coarsening scheme that incorporates topological connectivity and local geometric density to construct reliable coarse graphs. We validate the effectiveness of CGFM on three canonical scenarios across both 2D and 3D dynamics, which demonstrate its superior performance compared with that of state-of-the-art baselines.

","absKo":"

가능한 상태들의 전체 분포를 정확하게 모델링하는 것은 복잡하고 비정상적인 물리 시스템에서 통계적 성질을 이해하고 신뢰할 수 있는 예측을 가능하게 하는 데 매우 중요하다. 최근 diffusion models와 flow matching이 이러한 과제에서 가능성을 보여주었다. 그러나 이들 방법은 condition space 전반에 걸친 여러 개의 짧은 trajectory로부터 시스템의 일반 원리를 밝혀내는 데에는 여전히 한계가 있다. 또한 대규모 불규칙 기하에 대한 적응성이 떨어지며, 특히 급격한 gradient가 있는 영역에서 그러하다. 본 논문에서는 불완전한 training data로부터 물리 시스템의 전체 분포를 학습하기 위해 condition-aware flow matching과 hierarchical graph structure를 결합한 condition-aware graph flow matching(CGFM) 방법을 제안한다. 구체적으로, CGFM은 물리적 condition 전반에 걸친 smooth interpolation을 가능하게 하는 flow를 구성하고, HieraGraphNet을 통해 graph-conditioned vector field를 parameterize한다. HieraGraphNet은 multilevel graph 전반에서 message passing을 수행하여 multi-scale dynamics를 포착하고 물리 시스템 내의 장거리 정보 상호작용을 촉진한다. 또한 topological connectivity와 local geometric density를 통합하여 신뢰할 수 있는 coarse graph를 구성하는 topology- and geometry-aware graph coarsening scheme을 제안한다. 우리는 2D와 3D dynamics 모두에 걸친 세 가지 canonical scenario에서 CGFM의 효과를 검증했으며, 이는 state-of-the-art baselines에 비해 우수한 성능을 보인다.

"},{"id":"62370","en":"Are First-Order Diffusion Samplers Really Slower? A Fast Forward-Value Approach","ko":"1차 Diffusion Sampler는 정말 느린가? Fast Forward-Value 접근","authors":"Yuchen Jiao, Na Li, Changxiao Cai, Gen Li","pos":"#2909","link":"https://openreview.net/forum?id=hEQDZg1ZeD","abs":"

Higher-order ODE solvers have become a standard tool for accelerating diffusion probabilistic model (DPM) sampling, motivating the widespread view that first-order methods are inherently slower and that increasing discretization order is the primary path to faster generation. This paper challenges this belief and revisits acceleration from a complementary angle: beyond solver order, the placement of DPM evaluations along the reverse-time dynamics can substantially affect sampling accuracy in the low-neural function evaluation (NFE) regime. We propose a novel training-free, first-order sampler named Forward DPMSolver (F-DPMSolver), whose leading discretization error has the opposite sign to that of DDIM. Algorithmically, the method approximates the forward-value evaluation via a cheap one-step lookahead predictor. We provide theoretical guarantees showing that the resulting sampler provably approximates the ideal forward-value trajectory while retaining first-order convergence. Empirically, across standard image generation benchmarks, the proposed sampler consistently improves sample quality under the same NFE budget and can be competitive with, and sometimes outperform, state-of-the-art higher-order samplers. Overall, the results suggest that the placement of DPM evaluations provides an additional and largely independent design angle for accelerating diffusion sampling. Our code is available at https://anonymous.4open.science/r/F-DPMSolver.

","absKo":"

고차 ODE solver는 diffusion probabilistic model (DPM) sampling을 가속하는 표준 도구가 되었고, 이에 따라 first-order method는 본질적으로 더 느리며 discretization order를 높이는 것이 더 빠른 generation을 위한 핵심 경로라는 견해가 널리 퍼졌다. 이 논문은 이러한 믿음에 도전하며, solver order를 넘어 reverse-time dynamics를 따라 DPM evaluation을 어디에 배치하느냐가 low-neural function evaluation (NFE) 영역에서 sampling accuracy에 상당한 영향을 미칠 수 있음을 재검토한다. 우리는 leading discretization error의 부호가 DDIM과 반대인, training-free first-order sampler인 Forward DPMSolver (F-DPMSolver)를 제안한다. 알고리즘적으로는, 이 방법은 저렴한 one-step lookahead predictor를 통해 forward-value evaluation을 근사한다. 우리는 이로부터 얻어지는 sampler가 이상적인 forward-value trajectory를 provably 근사하면서도 first-order convergence를 유지함을 보이는 이론적 보장을 제공한다. 실험적으로, 표준 image generation benchmark 전반에서 제안 sampler는 동일한 NFE budget 하에서 sample quality를 일관되게 향상시키며, state-of-the-art higher-order sampler와 견줄 만하고 때로는 이를 능가한다. 전반적으로, 이 결과는 DPM evaluation의 배치가 diffusion sampling을 가속하기 위한 추가적이고 상당 부분 독립적인 설계 축임을 시사한다. 코드는 https://anonymous.4open.science/r/F-DPMSolver 에서 제공된다.

"},{"id":"61508","en":"Robust Parallel Diffusion Sampling via Dynamic Jacobian Bandwidth","ko":"Dynamic Jacobian Bandwidth를 통한 강건한 Parallel Diffusion Sampling","authors":"Zile Huang, Ser-Nam Lim","pos":"#302","link":"https://openreview.net/forum?id=pRsFAKBQKz","abs":"Recent years have witnessed significant progress in developing effective diffusion models. Parallel sampling is a promising recent approach that reformulates the sequential denoising process as solving a system of nonlinear equations, and it can be combined with other acceleration techniques. However, current progress is limited by the trade-off between high fidelity and computational efficiency. This paper addresses the challenge of scaling to high-dimensional, multi-modal generation. Specifically, we present ROPA (Robust Parallel Diffusion Sampling), which takes into account the properties of the denoising process and solves the linear system using adaptive local sparsity to achieve stable parallel sampling. Extensive experiments demonstrate ROPA’s effectiveness: it significantly accelerates sampling across diverse image and video diffusion models, achieving up to $2.9\\times$ speedup with eight core, an improvement of 52\\% over baselines without sacrificing sample quality. ROPA enables parallel sampling methods to provide a solid foundation for real-time, high-fidelity diffusion generation.","absKo":"최근 몇 년간 효과적인 diffusion model 개발에서 상당한 진전이 있었다. Parallel sampling은 순차적 denoising process를 nonlinear equation system을 푸는 문제로 재구성하는 유망한 최근 접근법이며, 다른 acceleration technique과 결합할 수 있다. 그러나 현재의 진전은 high fidelity와 computational efficiency 사이의 trade-off에 의해 제한된다. 이 논문은 high-dimensional, multi-modal generation으로의 scaling 문제를 다룬다. 구체적으로 우리는 denoising process의 특성을 고려하고 adaptive local sparsity를 사용해 linear system을 풀어 안정적인 parallel sampling을 달성하는 ROPA(Robust Parallel Diffusion Sampling)를 제안한다. 광범위한 실험은 ROPA의 효과를 입증한다. 다양한 image 및 video diffusion model 전반에서 sampling을 크게 가속하며, sample quality를 희생하지 않으면서 eight core에서 최대 $2.9\\times$ speedup, baseline 대비 52\\% 향상을 달성한다. ROPA는 parallel sampling method가 real-time, high-fidelity diffusion generation을 위한 견고한 기반을 제공하도록 한다."},{"id":"62087","en":"Temporal Difference Learning for Diffusion Models","ko":"Diffusion Model을 위한 Temporal Difference Learning","authors":"Qizhen Ying, Yangchen Pan, Victor Prisacariu, Junfeng Wen","pos":"#307","link":"https://openreview.net/forum?id=jwLSuhDyDc","abs":"

Diffusion models are typically trained with objectives that focus on local denoising targets at individual time steps (or adjacent pairs), which do not enforce consistency between predictions along the denoising trajectory. This lack of cross-time consistency can degrade performance, especially for few-step samplers. We introduce a temporal difference (TD) objective that penalizes inconsistency of the model’s \\emph{multi-step} progress along the denoising path. By reformulating the diffusion process as a Markov reward process and casting denoising as a policy evaluation problem in reinforcement learning, we derive a unified TD approach that applies to both discrete- and continuous-time diffusion formulations. We further propose a principled sample-based reweighting method that stabilizes training. Empirically, we show that using our TD training can significantly improve sample quality measured by FID, with stronger advantages when the number of sampling steps is small, highlighting its practical utility under low-computation-budget scenarios. We provide ablation studies to justify our design choices, including pairwise loss reweighting, regularization weight, and one-step stride. Overall, our TD approach can be a general drop-in that enforces cross-time consistency and improves generation quality across different diffusion generative models.

","absKo":"

Diffusion model은 일반적으로 개별 time step(또는 인접한 pair)의 local denoising target에 초점을 맞춘 objective로 학습되며, 이 objective는 denoising trajectory를 따라가는 예측들 사이의 일관성을 강제하지 않는다. 이러한 cross-time consistency의 부재는 특히 few-step sampler에서 성능 저하를 초래할 수 있다. 우리는 denoising path를 따라 모델의 \\emph{multi-step} 진행이 일관되지 않을 때 이를 패널티하는 temporal difference (TD) objective를 제안한다. diffusion process를 Markov reward process로 재구성하고 denoising을 reinforcement learning의 policy evaluation 문제로 정식화함으로써, discrete-time과 continuous-time diffusion formulation 모두에 적용 가능한 통합 TD approach를 도출한다. 또한 학습을 안정화하는 원칙적인 sample-based reweighting 방법을 추가로 제안한다. 실험적으로, 우리의 TD training을 사용하면 FID로 측정한 sample quality가 크게 향상되며, sampling step 수가 적을수록 더 강한 이점을 보인다. 이는 낮은 computation budget 상황에서의 실용적 유용성을 강조한다. 우리는 pairwise loss reweighting, regularization weight, one-step stride를 포함한 설계 선택을 정당화하기 위해 ablation study도 제시한다. 전반적으로, 우리의 TD approach는 cross-time consistency를 강제하고 다양한 diffusion generative model 전반에서 generation quality를 향상시키는 일반적인 drop-in 방식이 될 수 있다.

"},{"id":"61885","en":"Categorical Reparameterization with Denoising Diffusion models","ko":"범주형 Reparameterization with Denoising Diffusion 모델","authors":"Samson Gourevitch, Alain Oliviero Durmus, Jimmy Olsson, Eric Moulines, Yazid Janati","pos":"#3700","link":"https://openreview.net/forum?id=lxKyBy8NDn","abs":"

Learning models with categorical variables requires optimizing expectations over discrete distributions, a setting in which stochastic gradient-based optimization is challenging due to the non-differentiability of categorical sampling. A common workaround is to replace the discrete distribution with a continuous relaxation, yielding a smooth surrogate that admits reparameterized gradient estimates via the reparameterization trick. Building on this idea, we introduce ReDGE, a novel and efficient diffusion-based soft reparameterization method for categorical distributions. Our approach defines a flexible class of gradient estimators that includes the Straight-Through estimator as a special case. Experiments spanning latent variable models and inference-time reward guidance in discrete diffusion models demonstrate ReDGE consistently matches or outperforms existing gradient-based methods.

","absKo":"

범주형 변수를 가진 모델을 학습하려면 이산 분포에 대한 기댓값을 최적화해야 하며, 범주형 샘플링의 비미분성 때문에 확률적 gradient 기반 최적화가 어려운 설정이 된다. 흔한 우회 방법은 이산 분포를 연속적 relaxation으로 대체하여, reparameterization trick을 통해 reparameterized gradient 추정을 허용하는 매끄러운 surrogate를 얻는 것이다. 이러한 아이디어를 바탕으로, 우리는 범주형 분포를 위한 새롭고 효율적인 diffusion-based soft reparameterization 방법인 ReDGE를 제안한다. 우리의 접근법은 Straight-Through estimator를 특수한 경우로 포함하는 유연한 gradient estimator 클래스를 정의한다. latent variable model과 discrete diffusion model에서의 inference-time reward guidance에 걸친 실험은 ReDGE가 기존 gradient 기반 방법과 일관되게 동등하거나 더 우수함을 보여준다.

"},{"id":"66727","en":"Markovian Projection of Star-Shaped Diffusion for Exponential Family Distributions","ko":"지수족 분포를 위한 Star-Shaped Diffusion의 Markov 사영","authors":"François Bertholom, Khalid Oublal","pos":"#3902","link":"https://openreview.net/forum?id=0yBwJcNrHa","abs":"

Diffusion models achieve state-of-the-art performance in generative modeling but are limited by their reliance on Gaussian noise and the high computational cost of iterative sampling. Star-shaped diffusion addresses the former by introducing a non-Markovian forward process, yet this comes at the expense of temporal coherence in the reverse process. We propose a novel framework that resolves this trade-off by learning a Markovian projection of a star-shaped forward process, and its reversal. This design enables learning over a broad class of exponential models and recovers DDPM as a special case. It is particularly well-suited for knowledge distillation, allowing few-step or even single-step generation. Experiments demonstrate the effectiveness and flexibility of our approach across multiple generative tasks. Code is available at \\url{https://anonymous.4open.science/w/MStar-Diffusion-B3EE/}.

","absKo":"

Diffusion model은 generative modeling에서 state-of-the-art 성능을 달성하지만, Gaussian noise에 대한 의존성과 반복적 sampling의 높은 계산 비용에 의해 제한됩니다. Star-shaped diffusion은 비-Markovian forward process를 도입하여 전자를 해결하지만, 이는 reverse process의 temporal coherence를 희생하는 대가를 치릅니다. 우리는 star-shaped forward process의 Markovian projection과 그 역과정을 학습함으로써 이러한 trade-off를 해결하는 새로운 framework를 제안합니다. 이 설계는 광범위한 exponential model class에 대한 학습을 가능하게 하고 DDPM을 특별한 경우로 복원합니다. 이는 knowledge distillation에 특히 잘 맞으며, few-step 또는 심지어 single-step generation을 가능하게 합니다. 실험은 여러 generative task 전반에 걸쳐 우리의 접근법의 효과성과 유연성을 보여줍니다. 코드는 \\url{https://anonymous.4open.science/w/MStar-Diffusion-B3EE/}에서 제공됩니다.

"},{"id":"66818","en":"CANDI: Hybrid Discrete-Continuous Diffusion Models","ko":"CANDI: Hybrid Discrete-Continuous Diffusion 모델","authors":"Patrick Pynadath, Jiaxin Shi, Ruqi Zhang","pos":"#3904","link":"https://openreview.net/forum?id=03KQGUvw1e","abs":"

While continuous diffusion has shown remarkable success in continuous domains such as image generation, its direct application to discrete data has underperformed pure discrete formulations. To understand this gap, we introduce token identifiability, an analytical framework characterizing how Gaussian noise corrupts discrete data through two mechanisms: discrete identity corruption and continuous rank degradation. We reveal that these mechanisms scale differently with vocabulary size, creating a temporal dissonance that forces a tradeoff between learning continuous geometry and discrete structure. To address this, we propose CANDI (Continuous ANd DIscrete diffusion), a hybrid framework that decouples discrete and continuous corruption, enabling simultaneous learning of both. This unlocks the benefits of continuous diffusion for discrete spaces: on controlled generation, CANDI enables classifier-based guidance with off-the-shelf classifiers through simple gradient addition; on text generation, CANDI outperforms masked diffusion at low NFE, demonstrating the value of learning continuous gradients for discrete spaces.

","absKo":"

continuous diffusion은 image generation과 같은 continuous domain에서 놀라운 성공을 거두었지만, discrete data에 직접 적용했을 때는 pure discrete formulation보다 성능이 떨어졌다. 이 격차를 이해하기 위해 우리는 token identifiability를 도입한다. 이는 Gaussian noise가 discrete data를 두 가지 메커니즘, 즉 discrete identity corruptioncontinuous rank degradation을 통해 어떻게 훼손하는지를 특징짓는 분석적 framework이다. 우리는 이 메커니즘들이 vocabulary size에 따라 서로 다르게 스케일링되어, continuous geometry 학습과 discrete structure 학습 사이의 절충을 강제하는 temporal dissonance를 만든다는 점을 밝힌다. 이를 해결하기 위해 우리는 CANDI(Continuous ANd DIscrete diffusion)를 제안한다. 이는 discrete corruption과 continuous corruption을 분리하는 hybrid framework로, 두 가지를 동시에 학습할 수 있게 한다. 이로써 discrete space에서 continuous diffusion의 이점을 활용할 수 있다: controlled generation에서는 CANDI가 off-the-shelf classifier를 사용한 classifier-based guidance를 단순한 gradient addition만으로 가능하게 하고; text generation에서는 낮은 NFE에서 masked diffusion보다 우수한 성능을 보여 discrete space에서 continuous gradient를 학습하는 가치가 있음을 입증한다.

"},{"id":"64124","en":"Unitary Convolutions for Message-passing and Positional Encodings on Directed Graphs","ko":"방향 그래프에서 메시지 전달 및 위치 인코딩을 위한 유니터리 합성곱","authors":"Lukas Fesser, Bobak Kiani, Melanie Weber","pos":"#4603","link":"https://openreview.net/forum?id=QhNO3qmXST","abs":"

In many real-world networks, relationships are inherently directional, yet most graph neural networks (GNNs) assume undirected edges, and naïve adaptations of undirected GNNs to directed graphs amplify oversmoothing and gradient pathologies that cap model depth. Unitary graph convolutions (UniConv) provably prevent representational collapse and oversmoothing, but cannot incorporate edge directionality or edge features. In this paper, we introduce a directed unitary GNN with edge features (Dune), which retains these guarantees while overcoming UniConv’s limitations by incorporating edge directionality and edge features. Dune keeps gradient norms bounded at any number of layers, allowing it to benefit from neural network depth, unlike existing directed GNNs. The same unitary operator can be embedded in hybrid architectures with graph transformers, where its wavelike propagation supplies positional information and reduces the importance of random-walk or Laplacian-based encodings. We prove that Dune avoids exponential oversmoothing that plagues existing directed GNNs and empirically show that it achieves state-of-the-art performance on 12 directed-graph benchmarks while remaining trainable beyond 100 layers, improving performance by up to 18 percentage points over strong baselines. Our results establish unitary convolutions as a scalable, geometry-aware foundation for deep learning on directed graphs.

","absKo":"

많은 실제 네트워크에서 관계는 본질적으로 방향성을 띠지만, 대부분의 graph neural network(GNN)는 undirected edge를 가정한다. 또한 undirected GNN을 directed graph에 단순히 적용하면 oversmoothing과 gradient pathology가 심화되어 model depth가 제한된다. Unitary graph convolution(UniConv)은 representational collapse와 oversmoothing을 이론적으로 방지하지만, edge 방향성과 edge feature를 통합할 수는 없다. 이 논문에서는 이러한 보장을 유지하면서 UniConv의 한계를 극복하기 위해 edge directionality와 edge feature를 포함하는 directed unitary GNN with edge features(Dune)를 제안한다. Dune은 layer 수와 무관하게 gradient norm을 bounded하게 유지하므로, 기존 directed GNN과 달리 neural network depth의 이점을 얻을 수 있다. 동일한 unitary operator는 graph transformer와의 hybrid architecture에도 내장될 수 있으며, 이때 wave-like propagation은 positional information을 제공하고 random-walk 또는 Laplacian-based encoding의 중요성을 낮춘다. 우리는 Dune이 기존 directed GNN을 괴롭히는 exponential oversmoothing을 피함을 증명하고, 12개의 directed-graph benchmark에서 state-of-the-art 성능을 달성하면서 100층을 넘어서도 학습 가능함을 실증적으로 보인다. 또한 강력한 baseline 대비 최대 18 percentage point의 성능 향상을 달성한다. 우리의 결과는 unitary convolution이 directed graph에서 deep learning을 위한 확장 가능하고 geometry-aware한 기반임을 확립한다.

"},{"id":"63366","en":"VCG-Bench: Towards A Unified Visual-Centric Benchmark for Structured Generation and Editing","ko":"VCG-Bench: 구조화된 생성과 편집을 위한 통합 시각 중심 벤치마크를 향하여","authors":"Xiaoyan Su, Peijie Dong, Zhenheng Tang, Song Tang, Yuyao Zhai, Kaitao Lin, Liang Chen, Gai Yuhang, Yuyu Luo, Qiang Wang, Xiaowen Chu","pos":"#4602","link":"https://openreview.net/forum?id=XjSd2CtV20","abs":"

Despite the rapid advancements in Vision-Language Models (VLMs), a critical gap remains in their ability to handle structured, controllable diagrammatic tasks essential for professional workflows, as existing methods predominantly rely on pixel-based synthesis which operates in probabilistic pixel spaces and is inherently limited in editability and fidelity; instead, we propose a new \"Diagram-as-Code\" paradigm with symbolic logic that leverages mxGraph Extensible Markup Language (XML) for precise diagram generation and editing, and we present VCG-Bench, a unified benchmark for visual-centric mxGraph tasks comprising (1) a taxonomized dataset of 1,449 diverse diagrams spanning 6 domains and 15 sub-domains, (2) a paradigm definition that integrates Generation (Vision-to-Code) and Editability (Code-to-Code), and (3) a Tailored Evaluation Protocol employing multi-dimensional metrics such as mxGraph Execution Success Rate and Style Consistency Score (SCS), where experimental results highlight the challenges faced by current State-of-the-Art (SOTA) VLMs in structured fidelity and instruction compliance, reflecting their vision and reasoning capabilities.

","absKo":"

Vision-Language Model(VLM)의 빠른 발전에도 불구하고, 전문적 워크플로에 필수적인 구조화되고 제어 가능한 diagrammatic task를 처리하는 능력에는 여전히 중요한 격차가 남아 있다. 기존 방법은 주로 pixel-based synthesis에 의존하는데, 이는 probabilistic pixel space에서 동작하며 editability와 fidelity가 본질적으로 제한된다. 이에 우리는 precise diagram generation과 editing을 위해 mxGraph Extensible Markup Language(XML)를 활용하는 symbolic logic 기반의 새로운 \"Diagram-as-Code\" paradigm을 제안하고, visual-centric mxGraph task를 위한 통합 benchmark인 VCG-Bench를 제시한다. 이 benchmark는 (1) 6개 domain과 15개 sub-domain에 걸친 1,449개 다양한 diagram으로 이루어진 taxonomized dataset, (2) Generation(Vision-to-Code)과 Editability(Code-to-Code)를 통합하는 paradigm definition, (3) mxGraph Execution Success Rate와 Style Consistency Score(SCS) 같은 multi-dimensional metric을 사용하는 Tailored Evaluation Protocol을 포함한다. 실험 결과는 현재 State-of-the-Art(SOTA) VLM들이 structured fidelity와 instruction compliance에서 겪는 어려움을 부각시키며, 이는 그들의 vision 및 reasoning 능력을 반영한다.

"},{"id":"63436","en":"xKV: Cross-Layer KV-Cache Compression via Aligned Singular Vector Extraction","ko":"xKV: 정렬된 특이 벡터 추출을 통한 교차 레이어 KV-Cache 압축","authors":"Chi-Chih Chang, Wei-Cheng Lin, Chien-Yu Lin, Hung-Yueh Chiang, Yash Akhauri, Xilai Dai, Huiqiang Jiang, Yucheng Li, Kai-Chiang Wu, Luis Ceze, Mohamed Abdelfattah","pos":"#1700","link":"https://openreview.net/forum?id=XDb69lykA8","abs":"

Long-context Large Language Models (LLMs) enable powerful applications but incur high memory costs due to the key–value states (KV-Cache). Recent studies attempt to share KV-Cache across layers, but these approaches either require expensive pretraining or rely on per-token cross-layer cosine similarity that is often limited in practice. We show, via Centered Kernel Alignment (CKA), that the dominant singular vectors of KV-Cache are well aligned across layers. Motivated by this observation, we propose xKV, a post-training compression method that jointly factorizes grouped-layer KV-Cache into a shared low-rank subspace, substantially reducing KV-Cache memory. Across widely used LLMs, xKV achieves up to 8× KV-Cache compression while preserving accuracy on long-context tasks and in multi-turn settings. To further improve efficiency, we introduce Selective Reconstruction (SR) at decode time. Combined with SR, xKV achieves up to 4.23× end-to-end speedup, surpassing notable baselines with 30% higher throughput under a similar accuracy level. Overall, xKV provides a plug-and-play approach to reduce both memory and latency for long-context LLM inference. Our code will be open-sourced.

","absKo":"

Long-context Large Language Models(LLMs)은 강력한 응용을 가능하게 하지만 key-value state(KV-Cache) 때문에 높은 메모리 비용을 발생시킨다. 최근 연구들은 layer 간 KV-Cache를 공유하려고 시도했지만, 이러한 접근은 비용이 큰 pretraining을 요구하거나 token별 cross-layer cosine similarity에 의존하며, 이는 실제로 종종 한계가 있다. 우리는 Centered Kernel Alignment(CKA)를 통해 KV-Cache의 지배적인 singular vector가 layer 간에 잘 정렬되어 있음을 보인다. 이러한 관찰에 동기를 얻어, 우리는 grouped-layer KV-Cache를 shared low-rank subspace로 공동 factorization하는 post-training compression method인 xKV를 제안하며, 이를 통해 KV-Cache 메모리를 크게 줄인다. 널리 사용되는 LLM 전반에서 xKV는 long-context task와 multi-turn setting에서 정확도를 유지하면서 KV-Cache를 최대 8배 압축한다. 효율성을 더 높이기 위해 decode time에서 Selective Reconstruction(SR)을 도입한다. SR과 결합하면 xKV는 end-to-end 속도를 최대 4.23배 향상시키며, 유사한 정확도 수준에서 30% 더 높은 throughput을 제공하는 주목할 만한 baseline들을 능가한다. 전반적으로 xKV는 long-context LLM inference에서 메모리와 latency를 모두 줄이는 plug-and-play 접근법을 제공한다. 코드는 오픈소스로 공개할 예정이다.

"},{"id":"64146","en":"ZipMoE: Efficient On-Device MoE Serving via Lossless Compression and Cache-Affinity Scheduling","ko":"ZipMoE: 무손실 압축 및 캐시 친화 스케줄링을 통한 효율적 온디바이스 MoE 서빙","authors":"Yuchen Yang, Yaru Zhao, Pu Yang, Shaowei Wang, Zhi-Hua Zhou","pos":"#1701","link":"https://openreview.net/forum?id=QUL7jp4QDQ","abs":"While Mixture-of-Experts (MoE) architectures substantially bolster the expressive power of large-language models, their prohibitive memory footprint severely impedes the practical deployment on resource-constrained edge devices, especially when model behavior must be preserved without relying on lossy quantization. In this paper, we present ZipMoE, an efficient and semantically lossless on-device MoE serving system. ZipMoE exploits the synergy between the hardware properties of edge devices and the statistical redundancy inherent to MoE parameters via a caching-scheduling co-design with provable performance guarantee. Fundamentally, our design shifts the paradigm of on-device MoE inference from an I/O-bound bottleneck to a compute-centric workflow that enables efficient parallelization. We implement a prototype of ZipMoE and conduct extensive experiments on representative edge computing platforms using popular open-source MoE models and real-world workloads. Our evaluation reveals that ZipMoE achieves up to 72.77% inference latency reduction and up to $6.76\\times$ higher throughput than the state-of-the-art systems. Our code is available at: https://anonymous.4open.science/r/s3fg2i1dn/.","absKo":"Mixture-of-Experts(MoE) 아키텍처는 대규모 언어 모델의 표현력을 크게 강화하지만, 막대한 메모리 사용량 때문에 자원 제약이 있는 엣지 디바이스에 실용적으로 배치하기가 매우 어렵고, 특히 손실이 있는 quantization에 의존하지 않고 모델 동작을 보존해야 할 때 그 제약이 더 심각해진다. 본 논문에서는 효율적이고 의미적으로 무손실인 on-device MoE serving system인 ZipMoE를 제안한다. ZipMoE는 엣지 디바이스의 하드웨어 특성과 MoE 파라미터에 내재된 통계적 중복성을, 성능 보장이 증명 가능한 caching-scheduling co-design을 통해 활용한다. 근본적으로, 우리의 설계는 on-device MoE inference의 패러다임을 I/O-bound 병목에서 compute-centric workflow로 전환하여 효율적인 병렬화를 가능하게 한다. 우리는 ZipMoE의 프로토타입을 구현하고, 대표적인 엣지 컴퓨팅 플랫폼에서 널리 사용되는 오픈소스 MoE 모델과 실제 워크로드를 사용해 광범위한 실험을 수행했다. 평가 결과, ZipMoE는 최신 시스템들보다 최대 72.77%의 inference latency 감소와 최대 $6.76\\times$의 더 높은 throughput을 달성한다. 코드: https://anonymous.4open.science/r/s3fg2i1dn/."},{"id":"61959","en":"WinQ: Accelerating Quantization-Aware Training of Large Language Models around Saddle Points","ko":"WinQ: saddle point 근방에서 Large Language Model의 Quantization-Aware Training 가속","authors":"Dongyue Li, Zechun Liu, Kai Yi, Changsheng Zhao, Zhenshuo Zhang, Raghuraman Krishnamoorthi, Harshit Khaitan, Hongyang Zhang, Steven Li","pos":"#1703","link":"https://openreview.net/forum?id=lIQnUtwWVb","abs":"Quantization-aware training is widely used for language model quantization in sub-4-bit precision, by training full-precision weights with gradients computed on the quantized model. The main bottleneck for this training approach is its slow convergence and plateauing of test performance, which gets worse in lower bit-widths. While observed in prior work, its precise cause has not been carefully studied. In this paper, we analyze the convergence by computing the Hessian spectrum of the model loss throughout quantization-aware training. We find the key reason is that the model weights converge to flat surfaces near saddle points, with a large fraction of Hessian eigenvalues concentrated around zero, and the magnitude of both positive and negative eigenvalues decreases over training. Additionally, the convergence speed is slower in lower bit-widths with significantly smaller Hessian eigenvalue magnitude. Motivated by these findings, we propose an approach to accelerate quantized training with minimal overhead named WinQ. This approach periodically performs linear weight interpolation between the full-precision and quantized weights and computes gradients on noise-injected weights. Both techniques effectively regularize the Hessian and accelerate training, resulting in an algorithm broadly applicable to quantization methods. Extensive experiments show that WinQ accelerates various quantized training methods by up to 4$\\times$. Under the same training budget, WinQ improves state-of-the-art sub-4-bit quantization performance by up to 8.8% relatively. Additionally, WinQ remains consistently effective across 16 settings of different language models, quantization methods, and bit-widths.","absKo":"Quantization-aware training은 정밀도 4-bit 미만에서 language model quantization에 널리 사용되며, quantized model에서 계산한 gradient로 full-precision weight를 학습한다. 이 학습 접근법의 주요 병목은 수렴이 느리고 test performance가 plateau에 머문다는 점이며, bit-width가 낮아질수록 이 문제는 더 심해진다. 이는 이전 연구에서도 관찰되었지만, 정확한 원인은 충분히 면밀히 연구되지 않았다. 본 논문에서는 quantization-aware training 전반에 걸쳐 model loss의 Hessian spectrum을 계산하여 수렴을 분석한다. 그 결과 핵심 원인은 model weight가 saddle point 근처의 flat surface로 수렴하면서 Hessian eigenvalue의 큰 부분이 0 주변에 집중되고, 학습이 진행될수록 양의 eigenvalue와 음의 eigenvalue의 크기 모두 감소하기 때문임을 발견했다. 또한 낮은 bit-width에서는 Hessian eigenvalue의 크기가 훨씬 작아 수렴 속도도 더 느리다. 이러한 발견에 착안해, 최소한의 오버헤드로 quantized training을 가속하는 방법인 WinQ를 제안한다. 이 접근법은 full-precision weight와 quantized weight 사이에서 선형 weight interpolation을 주기적으로 수행하고, noise가 주입된 weight에 대해 gradient를 계산한다. 이 두 기법은 모두 Hessian을 효과적으로 정규화하고 학습을 가속하여, quantization method 전반에 폭넓게 적용 가능한 알고리즘을 만든다. 광범위한 실험 결과 WinQ는 다양한 quantized training method를 최대 4$\\times$까지 가속한다. 동일한 training budget에서 WinQ는 state-of-the-art sub-4-bit quantization 성능을 최대 8.8% 상대적으로 향상시킨다. 또한 WinQ는 서로 다른 language model, quantization method, bit-width의 16개 설정 전반에서 일관되게 효과적이다."},{"id":"63038","en":"TwinQuant: Learnable Subspace Decomposition for 4-Bit LLM Quantization","ko":"TwinQuant: 4비트 LLM 양자화를 위한 학습 가능한 부분공간 분해","authors":"Haodong WANG, Junjie Liu, Zicong Hong, Qianli Liu, Jian Lin, Song Guo, Xu Chen","pos":"#1714","link":"https://openreview.net/forum?id=ahYRAjx7UG","abs":"4-bit quantization reduces the memory footprint and latency of large language model inference, but its aggressive precision reduction can severely degrade accuracy. Prior methods address this by decomposing each weight matrix into two components (e.g., via singular value decomposition) and quantizing them separately, assigning the bulk of values to a low-precision residual component while handling outliers with a high-precision low-rank component. However, such decompositions are designed to minimize the real-valued energy of the residual, rather than the post-quantization error of the residual and low-rank components. We propose TwinQuant, a 4-bit quantization framework that learns quantization-friendly decomposed subspaces and jointly reshapes both the low-rank and residual components. TwinQuant learns component-specific transformations via a joint optimization over the Stiefel and general linear manifolds, flattening their distributions and reducing dynamic-range imbalance. To enable efficient end-to-end execution, we further design a fused dual-component kernel that pipelines the two-stage low-rank computation on-chip and merges both components with a single epilogue, avoiding intermediate global-memory traffic. Across LLaMA3 and Qwen3 models, TwinQuant preserves near-FP16 accuracy and delivers up to $2.11\\times$ end-to-end speedup over an FP16 baseline.","absKo":"4-bit quantization은 large language model inference의 memory footprint와 latency를 줄여 주지만, 공격적인 precision reduction은 accuracy를 심각하게 저하시킬 수 있다. 기존 methods는 각 weight matrix를 두 component로 분해한 뒤(예: singular value decomposition을 통해) 이를 별도로 quantize하며, 값의 대부분은 low-precision residual component에 할당하고 outlier는 high-precision low-rank component으로 처리한다. 그러나 이러한 decomposition은 residual의 post-quantization error가 아니라 residual의 real-valued energy를 최소화하도록 설계되어 있다. 우리는 quantization-friendly decomposed subspaces를 학습하고 low-rank 및 residual components를 함께 재구성하는 4-bit quantization framework인 TwinQuant를 제안한다. TwinQuant는 Stiefel manifold와 general linear manifold에 대한 joint optimization을 통해 component-specific transformations을 학습하여, distribution을 평탄화하고 dynamic-range imbalance를 줄인다. 효율적인 end-to-end execution을 위해, 우리는 on-chip에서 two-stage low-rank computation을 파이프라인화하고 single epilogue로 두 component를 결합하는 fused dual-component kernel도 설계하여 intermediate global-memory traffic을 피한다. LLaMA3와 Qwen3 모델 전반에서 TwinQuant는 FP16에 가까운 accuracy를 유지하면서, FP16 baseline 대비 end-to-end speedup 최대 2.11x를 제공한다."},{"id":"63745","en":"The Geometry of Narrow Fine-Tuning Degradation: Trajectory Lock-in and Spectral Bifurcation","ko":"좁은 Fine-Tuning 성능 저하의 기하학: 궤적 고착과 스펙트럼 분기","authors":"Jia Liu, Jiaxin Luo, Xinhao Qiu, Yixue Hao, Min Chen","pos":"#1807","link":"https://openreview.net/forum?id=U7HBymjLQl","abs":"

Magnitude-based stability proxies such as parameter drift are widely used in narrow-task fine-tuning, yet they do not reliably indicate degradation of broad capabilities. We identify trajectory lock-in: under fixed training conditions for narrow adaptation, the joint evolution of task loss and broad generalization collapses onto a shared low-dimensional degradation curve, so many stabilizers primarily change the rate of progress along this curve rather than altering the curve itself. This yields a drift paradox, in which comparable Euclidean displacement can still correspond to divergent generalization outcomes. To diagnose the underlying structure, we introduce objective-agnostic geometric probes that track the effective update subspace, together with an online harm signal that reflects curvature-dominated channeling toward directions associated with broad degradation. Finally, we show that escaping lock-in requires a spectral bifurcation, namely a qualitative reorientation of the update subspace toward softer curvature modes, thereby improving broad generalization while maintaining matched training performance. We validate these findings across model scales and modalities in narrow-task settings, and report practical deployment procedures and overhead measurements.

","absKo":"

parameter drift와 같은 magnitude-based stability proxy는 좁은 task fine-tuning에서 널리 사용되지만, broad capability의 저하를 신뢰성 있게 나타내지는 못한다. 우리는 trajectory lock-in을 규명한다. 즉, 좁은 적응을 위한 고정된 training condition 아래에서 task loss와 broad generalization의 결합된 진화가 하나의 공유된 저차원 degradation curve로 수렴하므로, 많은 stabilizer는 곡선 자체를 바꾸기보다 그 곡선을 따라 진행되는 속도만 바꾼다. 이로 인해 유사한 Euclidean displacement가 서로 다른 generalization 결과에 대응할 수 있는 drift paradox가 발생한다. 이러한 내부 구조를 진단하기 위해, 우리는 effective update subspace를 추적하는 objective-agnostic geometric probe와 함께, broad degradation과 연관된 방향으로 curvature가 지배적으로 채널링되는 양상을 반영하는 online harm signal을 도입한다. 마지막으로, lock-in에서 벗어나려면 spectral bifurcation, 즉 update subspace를 더 완만한 curvature mode 쪽으로 질적으로 재지향해야 하며, 이를 통해 학습 성능은 유지하면서 broad generalization을 개선할 수 있음을 보인다. 우리는 좁은 task 설정에서 모델 규모와 modality 전반에 걸쳐 이 결과를 검증하고, 실제 배포 절차와 오버헤드 측정 결과를 보고한다.

"},{"id":"62561","en":"Scaling Law for Quantization-Aware Training","ko":"Quantization-Aware Training의 스케일링 법칙","authors":"Mengzhao Chen, Chaoyi Zhang, Jing Liu, Zeng, Zeyue Xue, Zhiheng Liu, Yunshui Li, Jin Ma, Jie Huang, zhou Xun, Ping Luo","pos":"#1907","link":"https://openreview.net/forum?id=fXr3uPr1G5","abs":"

Large language models (LLMs) demand substantial computational and memory resources, creating deployment challenges. Quantization-aware training (QAT) addresses these challenges by reducing model precision while maintaining performance. However, the scaling behavior of QAT, especially at 4-bit precision (W4A4), is not well understood. Existing QAT scaling laws often ignore key factors such as the number of training tokens and quantization granularity, which limits their applicability. This paper proposes a unified scaling law for QAT that models quantization error as a function of model size, training data volume, and quantization group size. Through 268 QAT experiments, we show that quantization error decreases as model size increases, but rises with more training tokens and coarser quantization granularity. To identify the sources of W4A4 quantization error, we decompose it into weight and activation components. Both components follow the overall trend of W4A4 quantization error, but with different sensitivities. Specifically, weight quantization error increases more rapidly with more training tokens. Further analysis shows that the activation quantization error in the FC2 layer, caused by outliers, is the primary bottleneck of W4A4 QAT quantization error. By applying mixed-precision quantization to address this bottleneck, we demonstrate that weight and activation quantization errors can converge to similar levels. Additionally, with more training data, weight quantization error eventually exceeds activation quantization error, suggesting that reducing weight quantization error is also important in such scenarios. These findings offer key insights for improving QAT research and development.

","absKo":"

Large language models (LLMs)는 막대한 연산 및 메모리 자원을 요구하여 배포에 어려움을 만든다. Quantization-aware training (QAT)은 모델 정밀도를 낮추면서 성능을 유지함으로써 이러한 문제를 해결한다. 그러나 특히 4-bit precision (W4A4)에서 QAT의 scaling behavior는 아직 잘 이해되지 않았다. 기존 QAT scaling law는 training tokens의 수와 quantization granularity 같은 핵심 요인을 종종 무시하여 적용 가능성이 제한된다. 본 논문은 모델 크기, training data volume, quantization group size의 함수로 quantization error를 모델링하는 QAT를 위한 통합 scaling law를 제안한다. 268회의 QAT 실험을 통해 우리는 모델 크기가 증가할수록 quantization error는 감소하지만, training tokens가 많아지고 quantization granularity가 거칠어질수록 error는 증가함을 보였다. W4A4 quantization error의 원인을 식별하기 위해 이를 weight component와 activation component로 분해했다. 두 component 모두 전체 W4A4 quantization error의 추세를 따르지만, 민감도는 서로 다르다. 구체적으로, weight quantization error는 training tokens가 많아질수록 더 빠르게 증가한다. 추가 분석은 outlier에 의해 발생하는 FC2 layer의 activation quantization error가 W4A4 QAT quantization error의 주요 병목임을 보여준다. 이 병목을 해결하기 위해 mixed-precision quantization을 적용한 결과, weight와 activation quantization error가 유사한 수준으로 수렴할 수 있음을 보였다. 또한 더 많은 training data를 사용할 경우, 결국 weight quantization error가 activation quantization error를 초과하게 되므로, 이러한 상황에서는 weight quantization error를 줄이는 것도 중요함을 시사한다. 이러한 결과는 QAT 연구 및 개발을 개선하기 위한 핵심 통찰을 제공한다.

"},{"id":"61958","en":"STAR-KV: Low-Rank KV Cache Compression via Soft Thresholding for Adaptive Rank Control","ko":"STAR-KV: adaptive rank control을 위한 soft-thresholding 기반 low-rank KV cache compression","authors":"Priyansh Bhatnagar, Ashkan Moradifirouzabadi, Se-Hyun Yang, SeungJae Lee, Jungwook Choi, Mingu Kang","pos":"#1910","link":"https://openreview.net/forum?id=lJjH1q6RwY","abs":"Low-rank projection has emerged as a promising approach for compressing the KV cache by exploiting hidden-dimension redundancy. However, prior methods rely on fixed or heuristic rank selection and struggle to achieve aggressive compression with minimal accuracy degradation. We propose STAR-KV, an adaptive low-rank KV cache compression framework with fine-grained rank control. STAR-KV encompasses 1) a differentiable thresholding mechanism that enables optimal rank selection at both attention-head and block levels, 2) a hybrid decomposition strategy that applies different low-rank factorizations according to the sensitivity of key and value projections, and 3) a low-rank--aware mixed precision quantization that leverages data statistics for near lossless low-bit quantization. Evaluated across multiple LLMs and benchmarks, STAR-KV achieves up to 75\\% KV cache compression and up to 20$\\times$ overall KV cache reduction when combined with quantization. Enabled by custom Triton-based GPU kernels, STAR-KV delivers up to 6.9$\\times$ speedup for the attention module and 3.1$\\times$ end-to-end generation throughput. The source code will be publicly available in the future.","absKo":"Low-rank projection은 hidden-dimension redundancy를 활용하여 KV cache를 압축하는 유망한 접근법으로 떠올랐다. 그러나 기존 방법들은 고정된 rank 선택 또는 heuristic rank 선택에 의존하며, 정확도 저하를 최소화하면서 공격적인 압축을 달성하는 데 어려움을 겪는다. 우리는 세밀한 rank control을 갖춘 adaptive low-rank KV cache compression framework인 STAR-KV를 제안한다. STAR-KV는 1) attention-head 및 block 수준 모두에서 최적의 rank 선택을 가능하게 하는 differentiable thresholding mechanism, 2) key와 value projection의 sensitivity에 따라 서로 다른 low-rank factorization을 적용하는 hybrid decomposition strategy, 3) data statistics를 활용하여 거의 손실 없는 low-bit quantization을 수행하는 low-rank-aware mixed precision quantization을 포함한다. 여러 LLM과 benchmark에서 평가한 결과, STAR-KV는 최대 75\\%의 KV cache compression과 quantization과 결합 시 최대 20$\\times$의 전체 KV cache reduction을 달성한다. 또한 custom Triton-based GPU kernel의 지원을 받아 attention module에서 최대 6.9$\\times$, end-to-end generation throughput에서 3.1$\\times$의 speedup을 제공한다. source code는 향후 공개될 예정이다."},{"id":"65000","en":"QTALE: Quantization-Robust Token-Adaptive Layer Execution for LLMs","ko":"QTALE: LLM을 위한 Quantization-Robust Token-Adaptive Layer Execution","authors":"Kanghyun Noh, Jinheon Choi, Yulhwa Kim","pos":"#2008","link":"https://openreview.net/forum?id=HvBHq4Ycvn","abs":"

Large language models (LLMs) demand substantial computational and memory resources, posing challenges for efficient deployment. Two complementary approaches have emerged to address these issues: token-adaptive layer execution, which reduces floating-point operations (FLOPs) by selectively bypassing layers, and quantization, which lowers memory footprint by reducing weight precision. However, naively integrating these techniques leads to additional accuracy degradation due to reduced redundancy in token-adaptive models. We propose QTALE (Quantization-Robust Token-Adaptive Layer Execution for LLMs), a novel framework that enables seamless integration of token-adaptive execution with quantization while preserving accuracy. Conventional token-adaptive methods reduce redundancy in two ways: (1) by limiting the diversity of training paths explored during fine-tuning, and (2) by lowering the number of parameters actively involved in inference. To overcome these limitations, QTALE introduces two key components: (1) a training strategy that ensures diverse execution paths are actively explored during fine-tuning, and (2) a post-training mechanism that allows flexible adjustment of the execution ratio at inference to reintroduce redundancy when needed. Experimental results show that QTALE enables seamless integration of token-adaptive layer execution with quantization, showing no noticeable accuracy difference, with the gap to quantization-only models kept below 0.5\\% on CommonsenseQA benchmarks. By combining token-adaptive execution for FLOPs reduction and quantization for memory savings, QTALE provides an effective solution for efficient LLM deployment.

","absKo":"

Large language models(LLMs)는 상당한 연산 및 메모리 자원을 요구하여 효율적인 배포에 어려움을 초래한다. 이러한 문제를 해결하기 위해 두 가지 상보적 접근이 등장했다. 하나는 layer를 선택적으로 우회하여 floating-point operations(FLOPs)을 줄이는 token-adaptive layer execution이고, 다른 하나는 weight precision을 낮춰 메모리 사용량을 줄이는 quantization이다. 그러나 이 기법들을 단순히 통합하면 token-adaptive model의 redundancy 감소로 인해 추가적인 정확도 저하가 발생한다. 우리는 정확도를 보존하면서 token-adaptive execution과 quantization을 매끄럽게 통합할 수 있는 새로운 프레임워크인 QTALE(Quantization-Robust Token-Adaptive Layer Execution for LLMs)를 제안한다. 기존 token-adaptive method는 redundancy를 두 가지 방식으로 줄인다. (1) fine-tuning 동안 탐색되는 training path의 다양성을 제한하고, (2) inference에 실제로 관여하는 parameter 수를 낮춘다. 이러한 한계를 극복하기 위해 QTALE는 두 가지 핵심 구성요소를 도입한다: (1) fine-tuning 동안 다양한 execution path가 적극적으로 탐색되도록 보장하는 training strategy, (2) 필요할 때 redundancy를 다시 도입할 수 있도록 inference 시 execution ratio를 유연하게 조정하는 post-training mechanism. 실험 결과, QTALE는 token-adaptive layer execution과 quantization의 매끄러운 통합을 가능하게 하며, CommonsenseQA benchmark에서 quantization-only model과의 성능 차이를 0.5\\% 이하로 유지하면서 눈에 띄는 정확도 차이를 보이지 않았다. FLOPs 절감을 위한 token-adaptive execution과 메모리 절감을 위한 quantization을 결합함으로써, QTALE는 효율적인 LLM 배포를 위한 효과적인 해법을 제공한다.

"},{"id":"66109","en":"PGS: Effective LLM Code Refinement via Property-Oriented and Structurally Minimal Feedback","ko":"PGS: 속성 지향적이고 구조적으로 최소인 피드백을 통한 효과적 LLM 코드 정제","authors":"Lehan He, Zeren Chen, Zhe Zhang, Xiang Gao, Lu Sheng","pos":"#2101","link":"https://openreview.net/forum?id=6xeSRifGeL","abs":"

LLMs excel at code generation, yet ensuring the functional correctness of their outputs remains a persistent challenge. While recent studies have applied Test-Driven Development (TDD) to refine code, these methods are often undermined by poor feedback quality, stemming from the scarcity of high-quality test cases and noisy signals from auto-generated ones. In this work, we shift the focus from test quantity to feedback quality. We introduce the Property-Generated Solver (PGS), a novel paradigm designed to generate highly effective feedback via two principles: it must be property-oriented, to provide semantic guidance beyond simple I/O mismatches, and structurally minimal, to reduce cognitive load and isolate root causes. PGS operates by checking high-level program properties (e.g., a sorting function must produce a non-decreasing sequence) then providing the simplest failing counterexample to the LLM. This property-driven, minimal feedback steers LLMs toward correct and generalizable solutions. Across diverse benchmarks, PGS demonstrates superior performance, achieving a bug fix rate 1.4x-1.6x higher than the strongest debugging-based approaches and establishing a new state-of-the-art in automated code refinement. Source code and data are available in the supplementary.

","absKo":"

LLM은 code generation에 뛰어나지만, 출력의 functional correctness를 보장하는 일은 여전히 지속적인 과제이다. 최근 연구들은 code를 정제하기 위해 Test-Driven Development(TDD)를 적용했지만, 이러한 방법은 고품질 test case의 부족과 auto-generated test에서 비롯되는 noisy signal 때문에 종종 feedback quality가 저하된다. 본 연구에서는 test의 양에서 feedback의 질로 초점을 옮긴다. 우리는 Property-Generated Solver(PGS)를 제안하는데, 이는 두 가지 원칙을 통해 매우 효과적인 feedback을 생성하도록 설계된 새로운 paradigm이다. 첫째, 단순한 I/O mismatch를 넘어 semantic guidance를 제공하기 위해 property-oriented여야 하고, 둘째, cognitive load를 줄이고 근본 원인을 고립하기 위해 structurally minimal해야 한다. PGS는 고수준 program property를 검사한 뒤(예: sorting function은 non-decreasing sequence를 만들어야 함), LLM에 가장 단순한 failing counterexample을 제공하는 방식으로 동작한다. 이러한 property-driven, minimal feedback은 LLM이 정확하고 일반화 가능한 solution으로 향하도록 유도한다. 다양한 benchmark 전반에서 PGS는 우수한 성능을 보이며, 가장 강력한 debugging-based approach보다 1.4배~1.6배 높은 bug fix rate를 달성하고 automated code refinement에서 새로운 state-of-the-art를 확립한다. source code와 data는 supplementary에 제공된다.

"},{"id":"61771","en":"OTora: A Unified Red Teaming Framework for Reasoning-Level Denial-of-Service in LLM Agents","ko":"OTora: LLM Agent의 추론 수준 Denial-of-Service를 위한 통합 Red Teaming Framework","authors":"Xinyu Li, ronghui mu, Lin Li, Tianjin Huang, Gaojie Jin","pos":"#2109","link":"https://openreview.net/forum?id=n6smFKXQWg","abs":"

Large Language Models (LLMs) are increasingly deployed as autonomous agents that execute tool-augmented, multi-step tasks, where latency is a critical factor for real-world applications. Yet an overlooked threat is Reasoning-Level Denial-of-Service (R-DoS), in which an attacker preserves task correctness but degrades availability by inflating an agent’s reasoning depth or tool-use budget. We introduce OTora, the first unified, two-stage red-teaming framework for instantiating R-DoS attacks. Stage I optimizes an adversarial trigger that induces targeted tool invocations using insertion-aware scoring and dynamic target co-evolution, supporting both black-box and white-box settings. Stage II generates agent-aware reasoning payloads via an ICL-guided genetic search that amplifies overthinking while maintaining correct task outcomes. Across WebShop, Email, and OS agents built on multiple backbone models such as LLaMA-70B and GPT-OSS-120B, OTora achieves up to 10× increases in reasoning tokens and order-of-magnitude latency slowdowns, all while preserving near-baseline task accuracy. Finally, we discuss mitigation strategies for detecting and constraining abnormal reasoning and latency spikes.

","absKo":"

Large Language Models (LLMs)는 tool-augmented, multi-step tasks를 수행하는 autonomous agents로 점점 더 많이 배치되고 있으며, 이때 latency는 실제 응용에서 매우 중요한 요소이다. 그러나 간과된 위협은 Reasoning-Level Denial-of-Service (R-DoS)로, 공격자가 task correctness는 유지하면서 agent의 reasoning depth 또는 tool-use budget을 부풀려 availability를 저하시키는 방식이다. 우리는 R-DoS attacks를 구현하기 위한 최초의 통합형 2-stage red-teaming framework인 OTora를 제안한다. Stage I은 insertion-aware scoring과 dynamic target co-evolution을 사용하여 표적 tool invocation을 유도하는 adversarial trigger를 최적화하며, black-box와 white-box 설정을 모두 지원한다. Stage II는 ICL-guided genetic search를 통해 agent-aware reasoning payload를 생성하여, 올바른 task 결과를 유지하면서 overthinking을 증폭시킨다. LLaMA-70B와 GPT-OSS-120B 같은 여러 backbone models 위에 구축된 WebShop, Email, OS agents 전반에서 OTora는 reasoning tokens를 최대 10배까지 증가시키고 latency를 자릿수 수준으로 악화시키면서도, task accuracy는 거의 baseline 수준으로 유지한다. 마지막으로, 비정상적인 reasoning과 latency spike를 탐지하고 제한하기 위한 mitigation strategies를 논의한다.

"},{"id":"62571","en":"Not All Invariants Are Equal: Curating Training Data to Accelerate Program Verification with SLMs","ko":"모든 불변량이 동일하지 않다: SLM을 이용해 프로그램 검증 속도를 높이는 학습 데이터 큐레이션","authors":"Ido Pinto, Yizhak Elboher, Haoze Wu, Nina Narodytska, Guy Katz","pos":"#2111","link":"https://openreview.net/forum?id=fS28SOioQd","abs":"

The synthesis of inductive loop invariants is a critical bottleneck in automated program verification. While Large Language Models (LLMs) show promise in mitigating this issue, they often fail on hard instances, generating invariants that are invalid or computationally ineffective. While fine-tuning is a natural route to mitigate this limitation, obtaining high-quality training data for invariant generation remains an open challenge. We present a rigorous data curation pipeline designed to extract high-quality training signals from raw verifier-generated invariants. First, we formalize the properties required for a high-quality training invariant. Second, we propose \\textsc{Wonda}, a pipeline that refines noisy data via AST-based normalization, followed by LLM-driven semantic rewriting and augmentation with provable quality guarantees. We demonstrate that fine-tuning Small Language Models (SLMs) on this curated dataset result in consistent and significant performance gain. In particular, a fine-tuned 4B parameter model matches the utility of a GPT-OSS-120B baseline and approaches the state-of-the-art GPT-5.2, without incurring reasoning-time overhead. On challenging instances from the recent InvBench evaluation suite, our approach doubles the invariant correctness rate of base models; and improves their Virtual Best Performance (VBP) rates on the verification task by up to 14.2\\%.

","absKo":"

inductive loop invariant의 합성은 automated program verification의 중대한 병목이다. Large Language Model (LLM)이 이 문제를 완화할 잠재력을 보이지만, 난이도 높은 사례에서는 무효하거나 계산적으로 비효율적인 invariant를 생성하며 자주 실패한다. fine-tuning은 이러한 한계를 완화하는 자연스러운 경로이지만, invariant generation을 위한 고품질 training data를 확보하는 문제는 여전히 열려 있다. 우리는 raw verifier-generated invariant로부터 고품질 training signal을 추출하도록 설계된 엄격한 data curation pipeline을 제시한다. 먼저, 고품질 training invariant에 필요한 속성을 formalize한다. 다음으로, AST-based normalization에 이어 LLM-driven semantic rewriting과 provable quality guarantee를 갖는 augmentation을 통해 noisy data를 정제하는 \\textsc{Wonda} pipeline을 제안한다. 우리는 이처럼 정제된 dataset으로 Small Language Model (SLM)을 fine-tuning하면 일관되고 유의미한 성능 향상이 나타남을 보인다. 특히, fine-tuned 4B parameter model은 GPT-OSS-120B baseline의 효용과 맞먹고 reasoning-time overhead 없이 state-of-the-art GPT-5.2에 근접한다. 최근 InvBench evaluation suite의 까다로운 사례들에서, 우리의 접근은 base model의 invariant correctness rate를 두 배로 높였으며 verification task에서의 Virtual Best Performance (VBP) rate를 최대 14.2\\%까지 향상시켰다.

"},{"id":"64362","en":"LLM-MatLogic: Executable Exchange Contracts for Knowledge-Graph Query Answering with Scoped Negation","ko":"LLM-MatLogic: 범위 지정 부정을 이용한 지식 그래프 쿼리 응답을 위한 실행 가능한 교환 계약","authors":"Dezhuang Miao, Xiaoming Zhang, Bo Zhang, Yibin Du, Xiang Li, Ruilin Zeng, Yirui QI","pos":"#2206","link":"https://openreview.net/forum?id=OL4hndlNDS","abs":"LLM-to-KG systems frequently fail on exclusion-rich questions because natural-language negation is both scope-sensitive and evidence-dependent: it may constrain only one subgoal/branch and only certain supporting paths, yet such attachment is rarely explicit in text. We propose the Executable Exchange Contract (EEC) to bridge this gap, specifying scope-bound exclusions as executable control metadata exchanged between a specifier and an executor. Our executor, MatLogic, compiles exclusions into scope-local masks applied during multi-hop propagation and executes requests under a unified P$\\rightarrow$N$\\rightarrow$C/D schedule, ensuring exclusions are enforced before witness loss and branch entanglement. The system can also return compact witness pointers to keep support types distinguishable when needed. We evaluate on both structured complex queries and end-to-end natural-language KGQA, and introduce contract-aligned diagnostics that isolate errors from specification versus execution and verify the necessity of scoped enforcement.","absKo":"LLM-to-KG system은 배제(exclusion)가 풍부한 질문에서 자주 실패합니다. 자연어 negation은 scope-sensitive하고 evidence-dependent하기 때문인데, 이는 하나의 subgoal/branch만을 제약하고 특정 supporting path에만 적용될 수 있지만, 이러한 부착 관계는 텍스트에서 명시적으로 드러나는 경우가 드뭅니다. 우리는 specifier와 executor 사이에서 교환되는 executable control metadata로서 scope-bound exclusion을 명세하는 Executable Exchange Contract(EEC)를 제안하여 이 간극을 메웁니다. 우리의 executor인 MatLogic은 exclusion을 multi-hop propagation 동안 적용되는 scope-local mask로 compile하고, 통합된 P$\\rightarrow$N$\\rightarrow$C/D schedule 아래에서 요청을 실행하여 witness loss와 branch entanglement가 발생하기 전에 exclusion이 강제되도록 합니다. 또한 이 system은 필요할 때 support type을 구분 가능하게 유지하기 위해 compact witness pointer를 반환할 수 있습니다. 우리는 구조화된 complex query와 end-to-end natural-language KGQA 모두에서 평가를 수행하고, specification과 execution에서 비롯된 오류를 분리하며 scoped enforcement의 필요성을 검증하는 contract-aligned diagnostic을 도입합니다."},{"id":"66697","en":"INT vs. FP: A Comprehensive Study of Fine-Grained Low-bit Quantization Formats","ko":"INT vs. FP: Fine-Grained Low-bit Quantization 포맷에 대한 포괄적 연구","authors":"Mengzhao Chen, Meng Wu, Hui Jin, Zhihang Yuan, Jing Liu, Chaoyi Zhang, Yunshui Li, Jie Huang, Jin Ma, Zeyue Xue, Zhiheng Liu, Xingyan Bin, Ping Luo","pos":"#2301","link":"https://openreview.net/forum?id=1GIYHWO9S5","abs":"

Modern AI hardware, such as Nvidia's Blackwell architecture, is increasingly embracing low-precision floating-point (FP) formats to handle the pervasive activation outliers in Large Language Models (LLMs). Despite this industry trend, a unified comparison of FP and integer (INT) quantization across varying granularities has been missing, leaving algorithm and hardware co-design without clear guidance. This paper fills that gap by systematically investigating the trade-offs between FP and INT formats. We reveal a critical performance crossover: while FP excels in coarse-grained quantization, INT consistently surpasses it as the quantization block size shrinks. Our comprehensive comparison demonstrates that for popular fine-grained formats like MX (block size 32), MXINT8 and MXINT4 are superior to their FP counterparts in both algorithmic accuracy and hardware efficiency. We also introduce a symmetric clipping method that resolves gradient bias in fine-grained low-bit INT training, enabling nearly lossless performance for MXINT8 training. These findings challenge the current hardware trajectory and advocate for prioritizing fine-grained INT formats in future AI accelerators to achieve a better balance of accuracy, power, and efficiency.

","absKo":"

Nvidia의 Blackwell architecture와 같은 현대 AI 하드웨어는 Large Language Models(LLMs)에 만연한 activation outlier를 처리하기 위해 저정밀 floating-point(FP) format을 점점 더 채택하고 있다. 그러나 이러한 산업적 흐름에도 불구하고, 서로 다른 granularity 전반에서 FP와 integer(INT) quantization을 통합적으로 비교한 연구는 부족하여 algorithm과 hardware co-design에 명확한 지침이 없었다. 본 논문은 FP와 INT format 사이의 trade-off를 체계적으로 조사하여 이 공백을 메운다. 우리는 중요한 performance crossover를 밝혀낸다. coarse-grained quantization에서는 FP가 우수하지만, quantization block size가 작아질수록 INT가 일관되게 이를 능가한다. 우리의 포괄적 비교는 MX(block size 32)와 같은 널리 사용되는 fine-grained format에서 MXINT8과 MXINT4가 알고리즘 정확도와 hardware 효율성 모두에서 대응하는 FP format보다 우수함을 보여준다. 또한 fine-grained low-bit INT training에서 gradient bias를 해결하는 symmetric clipping method를 도입하여 MXINT8 training에서 거의 손실 없는 성능을 가능하게 한다. 이러한 결과는 현재의 hardware 방향에 도전하며, 정확도, 전력, 효율성의 균형을 더 잘 맞추기 위해 향후 AI accelerator에서 fine-grained INT format을 우선시해야 한다는 점을 강하게 시사한다.

"},{"id":"64581","en":"From LLM-Generated Conjectures to Lean Formalizations: Automated Polynomial Inequality Proving via Sum-of-Squares Certificates","ko":"LLM 생성 추측에서 Lean 형식화로: Sum-of-Squares 인증서를 통한 다항식 부등식 자동 증명","authors":"Ruobing Zuo, Hanrui Zhao, Gaolei He, Zhengfeng Yang, Jianlin Wang","pos":"#2313","link":"https://openreview.net/forum?id=Lz8rHDmmKt","abs":"

Automated proving of polynomial inequalities is a fundamental challenge in automated mathematical reasoning, where rich algebraic structure and a rapidly growing certificate search space hinder scalability. Purely symbolic approaches provide strong guarantees but often scale poorly as the number of variables or the degree increases, due to expensive algebraic manipulations and rapidly growing intermediate expressions. In parallel, LLM-guided methods have made notable progress, particularly on competition-style inequalities with a small number of variables. To address the remaining scalability challenges, we propose NSPI, a neuro-symbolic framework that combines the complementary strengths of LLMs and symbolic computation for polynomial-inequality proving. Concretely, an LLM proposes a conjecture in the form of an approximate polynomial Sum-Of-Squares (SOS) decomposition; we refine it via symbolic computation to obtain an exact polynomial SOS representation, which directly proves the target inequality, and we further certify the proof in Lean, yielding an end-to-end pipeline from heuristic discovery to machine-checked proof. Experiments on challenging benchmarks involving polynomials with up to 10 variables demonstrate the effectiveness and scalability of the proposed method.

","absKo":"

다항 부등식의 자동 증명은 자동 수학적 추론의 근본적인 과제로, 풍부한 algebraic structure와 급격히 커지는 certificate search space가 확장성을 저해한다. 순수 symbolic approach는 강력한 보장을 제공하지만, 변수 수나 degree가 증가할수록 비용이 큰 algebraic manipulation과 빠르게 커지는 intermediate expression 때문에 확장성이 떨어지는 경우가 많다. 한편, LLM-guided 방법은 특히 변수 수가 적은 competition-style inequality에서 주목할 만한 진전을 보여 왔다. 남은 확장성 문제를 해결하기 위해, 우리는 다항 부등식 증명을 위해 LLM과 symbolic computation의 상보적 강점을 결합한 neuro-symbolic framework인 NSPI를 제안한다. 구체적으로, LLM은 근사적인 polynomial Sum-Of-Squares (SOS) decomposition 형태의 conjecture를 제안한다. 우리는 이를 symbolic computation으로 정교화하여 정확한 polynomial SOS representation을 얻고, 이는 목표 부등식을 직접 증명한다. 또한 Lean에서 proof를 추가로 검증하여, heuristic discovery에서 machine-checked proof에 이르는 end-to-end pipeline을 구성한다. 최대 10개의 변수를 포함하는 다항식을 다루는 어려운 benchmark에 대한 실험은 제안한 방법의 효과성과 확장성을 입증한다.

"},{"id":"64041","en":"Escaping the Subspace Trap: The Role of Optimizer Geometry in Model Width Expansion","ko":"부분공간 함정 탈출: 모델 폭 확장에서 옵티마이저 기하학의 역할","authors":"Jiabei Chen, Haoyu Wang, Yang Yu, Yao Xu, Liangdong Wang, Guang Liu, Shizhu He, Jun Zhao, Kang Liu","pos":"#2408","link":"https://openreview.net/forum?id=RUgiEJJ9Fo","abs":"

Pre-training large language models from scratch is prohibitively expensive as model scales increase. A practical alternative is Model Width Expansion (MWE), which grows a larger model from a well-pretrained ''seed'' model to inherit existing capabilities at initialization. However, we identify a phenomenon termed the Subspace Trap: during continual pre-training, parameter updates largely stagnate within a low-dimensional subspace aligned with the initialization, limiting the effective capacity of the expanded model. Our theoretical analysis investigates this issue by attributing it to the function-preserving properties of width expansion. In particular, element-wise adaptive optimizers remain confined to the trap, whereas optimizers that yield an isotropic geometry of parameter updates can escape. To demonstrate the impact of the subspace trap on model performance, we conduct empirical experiments across different model sizes and model families, which show that escaping the trap is principally effective in improving training efficiency and overall model performance. Detailed mechanistic analyses further confirm that escaping the trap indeed activates the new dimensions to encode general knowledge. Our code is available at https://anonymous.4open.science/r/MWE-1B46.

","absKo":"

대형 language model을 scratch부터 pre-training하는 것은 model scale이 커질수록 매우 큰 비용이 든다. 실용적인 대안은 Model Width Expansion (MWE)으로, 잘 pre-train된 ''seed'' model에서 더 큰 model을 확장하여 초기화 시점에 기존 능력을 계승한다. 그러나 우리는 Subspace Trap이라는 현상을 확인했다. continual pre-training 동안 parameter update가 initialization과 정렬된 저차원 subspace 내에 주로 정체되어, 확장된 model의 유효 capacity가 제한된다. 우리의 이론 분석은 이 문제를 width expansion의 function-preserving property에 기인하는 것으로 설명한다. 특히, element-wise adaptive optimizer는 trap에 갇힌 상태를 유지하는 반면, parameter update에 isotropic geometry를 부여하는 optimizer는 이를 탈출할 수 있다. subspace trap이 model performance에 미치는 영향을 보여주기 위해, 서로 다른 model size와 model family에 걸쳐 실험을 수행했으며, trap에서 탈출하는 것이 주로 training efficiency와 전체 model performance 향상에 효과적임을 보였다. 자세한 mechanistic analysis는 trap에서 벗어나는 과정이 실제로 새로운 dimension을 활성화하여 일반 지식을 인코딩함을 추가로 확인한다. 코드는 https://anonymous.4open.science/r/MWE-1B46 에서 उपलब्ध하다.

"},{"id":"66310","en":"ExCyTIn-Bench: Evaluating LLM agents on Cyber Threat Investigation","ko":"ExCyTIn-Bench: 사이버 위협 조사에서 LLM 에이전트 평가","authors":"Yiran Wu, Mauricio Velazco, Andrew Zhao, Manuel Luján, Srisuma Movva, Yogesh Roy, Quang Nguyen, Roberto Rodriguez, Qingyun Wu, Michael Albada, Julia Kiseleva, Anand Mudgerikar","pos":"#2410","link":"https://openreview.net/forum?id=4pQVVSKdkH","abs":"

We present \\textbf{ExCyTIn-Bench}, the first benchmark to \\textbf{E}valuate an LLM agent \\textbf{X} on the task of \\textbf{Cy}ber \\textbf{T}hreat \\textbf{In}vestigation through security questions derived from investigation graphs. Real‑world security analysts must sift through a large number of heterogeneous security logs, follow multi‑hop chains of evidence to investigate threats. With the developments of LLMs, building LLM-based agents for automatic threat investigation is a promising direction. We construct a benchmark from a controlled Azure tenant including a SQL environment covering 57 log tables from Microsoft Sentinel and related services, and 7542 generated questions. We leverage security logs extracted with expert-crafted detection logic to build threat investigation graphs, and then generate questions with LLMs using paired nodes on the graph, taking the start node as background context and the end node as answer. Anchoring each question to these explicit nodes and edges not only provides automatic, explainable ground truth answers but also makes the pipeline reusable and readily extensible to new logs. Our comprehensive experiments on the test set with different models confirm the difficulty of the task: the best model so far can achieve a reward of 0.606, leaving much headroom for future research.

","absKo":"

We present \\textbf{ExCyTIn-Bench}, the first benchmark to \\textbf{E}valuate an LLM agent \\textbf{X} on the task of \\textbf{Cy}ber \\textbf{T}hreat \\textbf{In}vestigation through security questions derived from investigation graphs. 실제 보안 분석가들은 많은 이질적인 보안 로그를 훑고, 위협을 조사하기 위해 evidence의 multi-hop chain을 따라가야 합니다. LLM의 발전과 함께, 자동 threat investigation을 위한 LLM-based agent를 구축하는 것은 유망한 방향입니다. 우리는 Microsoft Sentinel 및 관련 서비스의 57개 log table을 포함하는 SQL environment를 갖춘 통제된 Azure tenant에서 벤치마크를 구성했으며, 7542개의 생성된 질문을 포함합니다. 우리는 expert-crafted detection logic으로 추출한 security log를 활용해 threat investigation graph를 만들고, 그다음 그래프의 짝지어진 node를 사용하여 LLM으로 질문을 생성합니다. 이때 start node는 background context로, end node는 answer로 사용합니다. 각 질문을 이러한 명시적 node와 edge에 고정함으로써 자동화되고 설명 가능한 ground truth answer를 제공할 뿐 아니라, 파이프라인을 재사용 가능하고 새로운 log에 쉽게 확장 가능하게 만듭니다. 테스트 세트에 대한 다양한 모델의 포괄적 실험은 이 task의 난이도를 확인해 주며, 현재까지의 최고 모델도 0.606의 reward만 달성하여 향후 연구를 위한 충분한 여지를 남깁니다.

"},{"id":"66801","en":"Attacks on Machine-Text Detectors Retain Stylistic Fingerprints","ko":"Machine-Text 탐지기에 대한 공격은 문체적 지문을 보존한다","authors":"Rafael Rivera Soto, Barry Chen, Nicholas Andrews","pos":"#2602","link":"https://openreview.net/forum?id=0BfnRKdTXH","abs":"

Despite considerable progress in the development of machine-text detectors, the ease with which machine-text can be manipulated to evade detection has led to suggestions that the problem is inherently intractable. In this work, we investigate the limits of such evasion strategies. We demonstrate that while current attacks, ranging from prompt engineering to detector-guided optimization can effectively degrade performance of standard detectors, they fail to erase the underlying stylistic \"fingerprints\" of machine text. We show that few-shot detectors that utilize the stylistic feature space are robust to these evasion attempts, reliably detecting samples even from models explicitly tuned to prevent detection. This raises the question: does style represent a universal defense against machine-detection attacks? We demonstrate that the answer is \"no\" by introducing a novel paraphrasing approach that simultaneously optimizes for undetectability and adherence to specific human styles. We show that unlike prior methods, this attack effectively evades all considered detectors, including those that utilize writing style. However, we find that this evasion is not absolute: as the number of documents available for analysis grows, the human and machine distributions become distinguishable again. Overall, our findings suggest that reliable machine-text detection requires moving beyond single-document analysis to multi-document analysis.

","absKo":"

machine-text detector의 발전에도 불구하고, machine-text가 detection을 회피하도록 손쉽게 조작될 수 있다는 점 때문에 이 문제는 본질적으로 다루기 어렵다는 견해가 제기되어 왔다. 본 연구에서는 이러한 evasion strategy의 한계를 조사한다. 우리는 prompt engineering에서 detector-guided optimization에 이르는 현재의 공격들이 standard detector의 성능은 효과적으로 저하시킬 수 있지만, machine text의 근본적인 문체적 \"fingerprint\"를 지우지는 못한다는 것을 보인다. 또한 stylistic feature space를 활용하는 few-shot detector는 이러한 회피 시도에 견고하며, detection을 피하도록 명시적으로 조정된 model로부터 나온 sample조차도 안정적으로 탐지함을 보인다. 이는 다음 질문을 제기한다: style이 machine-detection attack에 대한 보편적 방어인가? 우리는 새로운 paraphrasing approach를 도입하여 답이 \"아니오\"임을 보인다. 이 방법은 탐지 불가능성과 특정 human style 준수를 동시에 최적화한다. 우리는 이전 방법들과 달리 이 공격이 writing style을 활용하는 detector를 포함하여 고려한 모든 detector를 효과적으로 회피함을 보인다. 그러나 이러한 회피가 절대적인 것은 아님을 발견한다. 분석에 사용할 수 있는 document 수가 늘어날수록 human distribution과 machine distribution은 다시 구별 가능해진다. 전반적으로, 우리의 결과는 신뢰할 수 있는 machine-text detection이 single-document analysis를 넘어 multi-document analysis로 나아가야 함을 시사한다.

"},{"id":"61567","en":"Attribution-Guided and Coverage-Maximized Pruning for Structural MoE Compression","ko":"Structural MoE Compression을 위한 Attribution-Guided 및 Coverage-Maximized Pruning","authors":"Yifu Ding, jiacheng wang, Ge Yang, Yongcheng Jing, Jinyang Guo, Xianglong Liu, Dacheng Tao","pos":"#2603","link":"https://openreview.net/forum?id=oreET6Wz52","abs":"Mixture-of-Experts (MoE) models scale compute efficiently, yet they remain expensive to deploy due to substantial memory footprint and inference overhead. Prior methods mainly operate at the expert level, either removing whole experts or ranking experts by importance. However, such expert-wise decisions are too coarse to identify redundancy, and often misallocate pruning budgets and limits compression. This issue worsens in large MoEs with dynamic routing and heterogeneous experts. To alleviate this dilemma, we for the first time observe that information in MoE experts is highly concentrated in a few channels, leaving substantial redundancy even in \"high importance\" experts. Accordingly, we propose a structural pruning framework tailored for MoEs, reforming the prune-ratio objective to maximizing channel-score coverage via an efficient attribution-based approximation. Experiments on DeepSeek and Qwen MoEs retain accuracy under 50\\% or 25\\% pruning joinly with 4-bit quantization, reducing the memory footprint of Qwen3-30B-A3B by 5.27$\\times$, and outperforming state-of-the-art baselines under diverse benchmarks.","absKo":"Mixture-of-Experts(MoE) 모델은 compute를 효율적으로 확장하지만, 상당한 memory footprint와 inference overhead 때문에 배포 비용이 여전히 크다. 기존 방법은 주로 expert 수준에서 동작하며, 전체 expert를 제거하거나 중요도에 따라 expert를 순위화한다. 그러나 이러한 expert 단위 결정은 중복성을 식별하기에는 너무 거칠고, 종종 pruning budget을 잘못 배분하며 compression을 제한한다. 이 문제는 dynamic routing과 heterogeneous expert를 가진 대형 MoE에서 더욱 심각해진다. 이를 완화하기 위해, 우리는 처음으로 MoE expert 내부의 정보가 소수의 channel에 매우 집중되어 있으며, \"high importance\" expert에서도 상당한 redundancy가 남아 있음을 관찰했다. 이에 따라 우리는 MoE에 특화된 structural pruning framework를 제안하며, 효율적인 attribution-based approximation을 통해 prune-ratio objective를 channel-score coverage를 최대화하는 방향으로 재구성한다. DeepSeek과 Qwen MoE에 대한 실험에서 4-bit quantization과 결합된 50% 또는 25% pruning 하에서도 정확도를 유지하였고, Qwen3-30B-A3B의 memory footprint를 5.27$\\times$ 줄였으며, 다양한 benchmark에서 state-of-the-art baseline을 능가했다."},{"id":"65383","en":"AutoQRA: Joint Optimization of Mixed-Precision Quantization and Low-rank Adapters for Efficient LLM Fine-Tuning","ko":"AutoQRA: 효율적 LLM 파인튜닝을 위한 혼합 정밀도 양자화와 Low-rank Adapter의 결합 최적화","authors":"Changhai Zhou, Shiyang Zhang, Yuhua Zhou, Qian Qiao, Jun Gao, Cheng Jin, KAIZHOU QIN, Weizhong Zhang","pos":"#2604","link":"https://openreview.net/forum?id=EB8MHApOav","abs":"

Quantization followed by parameter-efficient fine-tuning has emerged as a promising paradigm for downstream adaptation under tight GPU memory constraints. However, this sequential pipeline fails to leverage the intricate interaction between quantization bit-width and LoRA rank. Specifically, a carefully optimized quantization allocation with low quantization error does not always translate to strong fine-tuning performance, and different bit-width and rank configurations can lead to significantly varying outcomes under the same memory budget. To address this limitation, we propose AutoQRA, a joint optimization framework that simultaneously optimizes the bit-width and LoRA rank configuration for each layer during the mixed quantized fine-tuning process. To tackle the challenges posed by the large discrete search space and the high evaluation cost associated with frequent fine-tuning iterations, AutoQRA decomposes the optimization process into two stages. First, it first conducts a global multi-fidelity evolutionary search, where the initial population is warm-started by injecting layer-wise importance priors. This stage employs specific operators and a performance model to efficiently screen candidate configurations. Second, trust-region Bayesian optimization is applied to locally refine promising regions of the search space and identify optimal configurations under the given memory budget. This approach enables active compensation for quantization noise in specific layers during training. Experiments show that AutoQRA achieves performance close to full-precision fine-tuning with a memory footprint comparable to uniform 4-bit methods.

","absKo":"

quantization 뒤에 parameter-efficient fine-tuning을 적용하는 방식은 제한된 GPU memory 제약 하에서 downstream adaptation을 위한 유망한 paradigm으로 부상했다. 그러나 이 순차적 pipeline은 quantization bit-width와 LoRA rank 사이의 정교한 상호작용을 활용하지 못한다. 구체적으로, quantization error가 낮도록 세심하게 최적화된 allocation이 항상 강한 fine-tuning 성능으로 이어지는 것은 아니며, 같은 memory budget 아래에서도 서로 다른 bit-width와 rank configuration이 크게 다른 결과를 낳을 수 있다. 이러한 한계를 해결하기 위해, 우리는 mixed quantized fine-tuning 과정에서 각 layer의 bit-width와 LoRA rank configuration을 동시에 최적화하는 joint optimization framework인 AutoQRA를 제안한다. 큰 discrete search space와 빈번한 fine-tuning iteration에 수반되는 높은 평가 비용이 초래하는 문제를 다루기 위해, AutoQRA는 optimization 과정을 두 단계로 분해한다. 먼저, layer-wise importance prior를 주입해 warm-start된 초기 population을 사용하는 global multi-fidelity evolutionary search를 수행한다. 이 단계에서는 특정 operator와 performance model을 사용해 candidate configuration을 효율적으로 선별한다. 둘째, trust-region Bayesian optimization을 적용해 유망한 search space 영역을 국소적으로 정제하고 주어진 memory budget 하에서 최적 configuration을 찾는다. 이 접근은 training 중 특정 layer의 quantization noise를 능동적으로 보상할 수 있게 한다. 실험 결과, AutoQRA는 uniform 4-bit 방법과 비슷한 memory footprint로 full-precision fine-tuning에 가까운 성능을 달성한다.

"},{"id":"66562","en":"AutoSizer: Automatic Sizing of Analog and Mixed-Signal Circuits via Large Language Model (LLM) Agents","ko":"AutoSizer: 대규모 언어 모델(LLM) 에이전트를 통한 아날로그 및 혼합 신호 회로의 자동 크기 조정","authors":"Xi Yu, Dmitrii Torbunov, Soumyajit Mandal, Yihui Ren","pos":"#2605","link":"https://openreview.net/forum?id=2b4nmWhDBt","abs":"

The design of Analog and Mixed-Signal (AMS) integrated circuits remains heavily reliant on expert knowledge, with transistor sizing a major bottleneck due to nonlinear behavior, high-dimensional design spaces, and strict performance constraints. Existing Electronic Design Automation (EDA) methods typically frame sizing as static black-box optimization, resulting in inefficient and less robust solutions. Although Large Language Models (LLMs) exhibit strong reasoning abilities, they are not suited for precise numerical optimization in AMS sizing. To address this gap, we propose \\textsc{AutoSizer}, a reflective LLM-driven meta-optimization framework that unifies circuit understanding, adaptive search-space construction, and optimization orchestration in a closed loop. It employs a two-loop optimization framework, with an inner loop for circuit sizing and an outer loop that analyzes optimization dynamics and constraints to iteratively refine the search space from simulation feedback. We further introduce \\textsc{AMS-SizingBench}, an open benchmark comprising 24 diverse AMS circuits in SKY130 CMOS technology, designed to evaluate adaptive optimization policies under realistic simulator-based constraints. \\textsc{AutoSizer} experimentally achieves higher solution quality, faster convergence, and higher success rate across varying circuit difficulties, outperforming both traditional optimization methods and existing LLM-based agents.

","absKo":"

Analog and Mixed-Signal (AMS) integrated circuit 설계는 여전히 전문가 지식에 크게 의존하며, transistor sizing은 nonlinear behavior, 고차원 design space, 엄격한 성능 제약 때문에 주요 병목으로 남아 있다. 기존 Electronic Design Automation (EDA) 방법은 대개 sizing을 정적인 black-box optimization으로 다루어 비효율적이고 덜 robust한 해를 낳는다. Large Language Models (LLMs)는 강한 reasoning 능력을 보이지만, AMS sizing의 정밀한 수치 최적화에는 적합하지 않다. 이 간극을 해결하기 위해, 우리는 circuit understanding, adaptive search-space construction, optimization orchestration을 closed loop에서 통합하는 reflective LLM-driven meta-optimization framework인 \\textsc{AutoSizer}를 제안한다. 이 framework는 두 개의 loop를 사용하는 optimization 구조를 채택하며, inner loop는 circuit sizing을 담당하고 outer loop는 optimization dynamics와 제약을 분석하여 simulation feedback으로부터 search space를 반복적으로 정제한다. 또한 우리는 현실적인 simulator-based constraint 하에서 adaptive optimization policy를 평가하도록 설계된, SKY130 CMOS technology의 24개 다양한 AMS circuit으로 구성된 공개 benchmark인 \\textsc{AMS-SizingBench}를 도입한다. \\textsc{AutoSizer}는 실험적으로 서로 다른 circuit difficulty 전반에서 더 높은 solution quality, 더 빠른 convergence, 더 높은 success rate를 달성하며, 전통적 optimization 방법과 기존 LLM-based agent를 모두 능가한다.

"},{"id":"63340","en":"Automated Formal Proofs of Combinatorial Identities via Wilf–Zeilberger Guidance and LLMs","ko":"Wilf-Zeilberger 유도와 LLM을 통한 조합 항등식의 자동 형식 증명","authors":"Beibei Xiong, Hangyu Lv, Junqi Liu, Yisen Wang, Shaoshi Chen, Jianlin Wang, Zhengfeng Yang, Lihong Zhi","pos":"#2606","link":"https://openreview.net/forum?id=Xxq7fcQUNR","abs":"

Automating formal proofs of combinatorial identities is challenging for LLM-based provers, as long-horizon proof planning is required and unconstrained search quickly explodes. Symbolic methods such as the Wilf--Zeilberger (WZ) method can achieve a mechanized proof of combinatorial identities by constructing special auxiliary functions and demonstrating that they satisfy specific recurrence relations. We propose WZ-LLM, a neuro-symbolic framework that turns WZ proof plans into executable proof sketches in Lean~4 and uses an LLM-based prover to discharge the resulting machine-checkable subgoals. We also train a dedicated WZ-Prover via a Lean-kernel-verified bootstrapping loop with expert-verified iteration, followed by DAPO-based refinement. Experiments show that WZ-LLM achieves a 34\\% proof success rate on LCI-Test (100 classical combinatorial identities), outperforming strong baselines such as DeepSeek-V3 and Goedel-Prover-V2; moreover, on LCI-Test it proves 5 identities on which the symbolic-only baseline fails. WZ-LLM also improves performance on CombiBench and PutnamBench-Comb, suggesting the effectiveness of coupling symbolic proof sketches with learned formal reasoning. Experiments show that WZ-LLM achieves a 34\\% proof success rate on LCI-Test (100 classic combinatorial identities), outperforming strong baselines such as DeepSeek-V3 and Goedel-Prover-V2, and delivering consistent gains on CombiBench and PutnamBench-Comb. These results indicate that our framework provides two complementary strengths: improved direct proving for identities beyond the scope of WZ, and substantially higher end-to-end success when WZ sketches guide a specialized prover.

","absKo":"

조합 항등식의 formal proof를 자동화하는 일은 LLM 기반 prover에게 어렵다. 긴 horizon의 proof planning이 필요하고, 제약 없는 search는 빠르게 폭발하기 때문이다. Wilf--Zeilberger(WZ) method와 같은 symbolic method는 특별한 auxiliary function을 구성하고 그것이 특정 recurrence relation을 만족함을 보임으로써 조합 항등식의 mechanized proof를 달성할 수 있다. 우리는 WZ proof plan을 Lean~4에서 실행 가능한 proof sketch로 바꾸고, LLM 기반 prover를 사용해 그 결과로 생기는 machine-checkable subgoal을 해결하는 neuro-symbolic framework인 WZ-LLM을 제안한다. 또한 expert verification iteration이 포함된 Lean-kernel-verified bootstrapping loop에 이어 DAPO-based refinement를 통해 전용 WZ-Prover를 학습한다. 실험 결과 WZ-LLM은 LCI-Test(100개의 고전적 조합 항등식)에서 34\\%의 proof success rate를 달성하여 DeepSeek-V3와 Goedel-Prover-V2 같은 강력한 baseline을 능가했으며, 더 나아가 LCI-Test에서 symbolic-only baseline이 실패한 5개의 항등식을 증명했다. WZ-LLM은 CombiBench와 PutnamBench-Comb에서도 성능을 향상시켜, symbolic proof sketch와 학습된 formal reasoning을 결합하는 효과를 시사한다. 실험은 WZ-LLM이 LCI-Test(100개의 고전 조합 항등식)에서 34\\%의 proof success rate를 달성하여 DeepSeek-V3와 Goedel-Prover-V2 같은 강력한 baseline을 능가하고, CombiBench와 PutnamBench-Comb에서도 일관된 향상을 보임을 보여준다. 이러한 결과는 우리의 framework가 두 가지 상호보완적 강점을 제공함을 시사한다. 하나는 WZ의 범위를 넘어서는 항등식에 대한 직접 증명 성능 향상이고, 다른 하나는 WZ sketch가 특화된 prover를 안내할 때 end-to-end 성공률이 크게 높아진다는 점이다.

"},{"id":"62359","en":"BPDQ: Bit-Plane Decomposition Quantization on a Variable Grid for Large Language Models","ko":"BPDQ: Large Language Models를 위한 가변 격자 기반 Bit-Plane Decomposition Quantization","authors":"Junyu Chen, Jungang Li, Jing Xiong, Wenjie Wang, Qingyao Yang, He Xiao, Zhen Li, Taiqiang Wu, Mengzhao Chen, Zhen Peng, Chaofan Tao, Long Shi, Hongxia Yang, Ngai Wong","pos":"#2608","link":"https://openreview.net/forum?id=hIFa2QxEiC","abs":"

Large language model (LLM) inference is often bounded by memory footprint and memory bandwidth in resource-constrained deployments, making quantization a fundamental technique for efficient serving. While post-training quantization (PTQ) maintains high fidelity at 4-bit, it deteriorates at 2–3 bits. Fundamentally, existing methods enforce a shape-invariant quantization grid (e.g., the fixed uniform intervals of UINT2) for each group, severely restricting the feasible set for error minimization. To address this, we propose Bit-Plane Decomposition Quantization (BPDQ), which constructs a variable quantization grid via bit-planes and scalar coefficients, and iteratively refines them using approximate second-order information while progressively compensating quantization errors to minimize output discrepancy. In the 2-bit regime, BPDQ enables serving Qwen2.5-72B on a single RTX 3090 with 83.85\\% GSM8K accuracy (vs. 90.83\\% at 16-bit). Moreover, we provide theoretical analysis showing that the variable grid expands the feasible set, and that the quantization process consistently aligns with the optimal objective in Hessian-induced geometry. Code is available in the supplementary materials and will be open-sourced.

","absKo":"

자원 제약이 있는 배포 환경에서 large language model(LLM) 추론은 종종 memory footprint와 memory bandwidth에 의해 제한되므로, quantization은 효율적 serving을 위한 근본적인 기법이다. post-training quantization(PTQ)은 4-bit에서 높은 충실도를 유지하지만, 2–3 bit에서는 성능이 저하된다. 근본적으로 기존 방법은 각 group에 대해 shape-invariant quantization grid(예: UINT2의 고정된 uniform interval)를 강제하여, error minimization을 위한 feasible set을 심각하게 제한한다. 이를 해결하기 위해, 우리는 bit-plane과 scalar coefficient를 통해 variable quantization grid를 구성하고, approximate second-order information을 사용해 이를 반복적으로 정교화하면서 quantization error를 점진적으로 보상하여 output discrepancy를 최소화하는 Bit-Plane Decomposition Quantization(BPDQ)을 제안한다. 2-bit 환경에서 BPDQ는 Qwen2.5-72B를 단일 RTX 3090에서 serving할 수 있게 하며, GSM8K 정확도는 83.85\\%를 달성한다(16-bit에서 90.83\\%). 더 나아가, 우리는 variable grid가 feasible set을 확장하고, quantization 과정이 Hessian-induced geometry에서 최적 objective와 일관되게 정렬됨을 보이는 이론적 분석을 제시한다. 코드는 supplementary materials에 제공되며 공개될 예정이다.

"},{"id":"65065","en":"Predicting Large Model Test Losses with a Noisy Quadratic System","ko":"Noisy Quadratic System으로 대규모 모델의 Test Loss 예측","authors":"Chuning Li, Chris Maddison","pos":"#2915","link":"https://openreview.net/forum?id=H5wTMVlukn","abs":"We introduce a predictive model that estimates the pre-training loss of large models from model size ($N$), batch size ($B$) and number of weight updates ($K$). This is the first loss prediction model that can handle changing batch size. The model outperforms Chinchilla's loss model, a model of the test loss using the batch size and number of tokens, in terms of projecting the loss at extrapolated compute budgets (up to 1000 folds). A natural use of the model is to find optimal $N,B,K$ configurations under explicit and compound resource constraints like time, memory and compute. In our experiments, the model-selected configurations are close to ground-truth optimal. Our work advocates for loss prediction as a better alternative to heuristic-based laws, which are growing in complexity.","absKo":"모델 크기($N$), batch size($B$), weight update 수($K$)로부터 대규모 모델의 pre-training loss를 추정하는 예측 모델을 소개한다. 이는 batch size가 변하는 경우를 처리할 수 있는 최초의 loss prediction model이다. 이 모델은 batch size와 token 수를 사용해 test loss를 예측하는 Chinchilla의 loss model보다, 외삽된 compute budget(최대 1000배)에서 loss를 추정하는 성능이 더 우수하다. 이 모델의 자연스러운 활용처는 시간, 메모리, compute 같은 명시적이고 복합적인 자원 제약 하에서 최적의 $N,B,K$ 구성을 찾는 것이다. 우리의 실험에서 모델이 선택한 구성은 실제 최적값에 가깝다. 우리의 연구는 점점 더 복잡해지는 heuristic 기반 법칙보다 loss prediction이 더 나은 대안임을 주장한다."},{"id":"62293","en":"Rapid Poison: Practical Poisoning Attacks Against the Rapid Response Framework","ko":"Rapid Poison: Rapid Response Framework에 대한 실용적 Poisoning 공격","authors":"David Huang, Jaewon Chang, Avidan Shah, Prateek Mittal, Chawin Sitawarin","pos":"#3000","link":"https://openreview.net/forum?id=hvI3Syn2U7","abs":"

The Rapid Response (RR) framework (Peng et al., 2024), deployed in production systems including Anthropic’s ASL-3 safeguards (Anthropic, 2025), dynamically adapts jailbreak detection classifiers by generating synthetic training data from emerging attacks. We reveal that prompt injection can infiltrate this pipeline to deliver poisoned samples into the classifier’s training set, enabling two attack objectives: (I) targeted poisoning attacks that create false positives on harmless samples by categorizing them as a jailbreak, with a specific desired feature (e.g., certain formatting, subject, or keyword), (II) concept-based backdoor attacks that induce false negatives on jailbreak inputs, generalizing even to jailbreaks from attack strategies the defender explicitly trained against, when the backdoor trigger is present. Importantly, our threat model restricts adversaries to modify- ing only jailbreak samples (not benign data or labels), a constraint unexplored by prior work that makes the second objective particularly challeng- ing. We address this with Omission Attack, which exploits a new phenomenon: when training on concept-absent unsafe samples, the classifier mis- associates that concept’s presence with the safe label. Both attacks flip nearly all target labels with only 1% poisoning rate. Code: anonymous.tbd.

","absKo":"

Rapid Response (RR) 프레임워크(Peng et al., 2024)는 Anthropic의 ASL-3 safeguards(Anthropic, 2025)를 포함한 production systems에 배포되어 있으며, 새롭게 등장하는 공격으로부터 synthetic training data를 생성함으로써 jailbreak detection classifiers를 동적으로 적응시킨다. 우리는 prompt injection이 이 파이프라인에 침투하여 poisoned samples를 classifier의 training set에 주입할 수 있음을 밝힌다. 이를 통해 두 가지 공격 목표가 가능해진다: (I) harmless samples를 jailbreak로 분류하여 false positive를 만들되, 특정한 원하는 feature(예: 특정 formatting, subject, 또는 keyword)를 갖게 하는 targeted poisoning attacks, (II) backdoor trigger가 존재할 때 false negative를 유도하고, defender가 명시적으로 학습한 attack strategies로부터 비롯된 jailbreak에 대해서도 일반화되는 concept-based backdoor attacks. 특히, 우리의 threat model은 adversary가 jailbreak samples만 수정할 수 있고(benign data나 labels는 수정 불가) 라는 제약을 두며, 이는 prior work에서 탐구되지 않았던 조건으로 두 번째 목표를 특히 어렵게 만든다. 우리는 이를 Omission Attack으로 해결한다. 이는 새로운 현상을 이용하는데, concept가 없는 unsafe samples로 학습할 때 classifier가 그 concept의 존재를 safe label과 잘못 연관시키는 현상이다. 두 공격 모두 1% poisoning rate만으로 거의 모든 target labels를 뒤집는다. Code: anonymous.tbd.

"},{"id":"62509","en":"SPARD: Defending Harmful Fine-Tuning Attack via Safety Projection with Relevance–Diversity Data Selection","ko":"SPARD: Relevance–Diversity 데이터 선택 기반 Safety Projection으로 유해한 Fine-Tuning 공격 방어","authors":"Shuhao Chen, Weisen Jiang, Yeqi Gong, Shengda Luo, Chengxiang Zhuo, Zang Li, James Kwok, Yu Zhang","pos":"#3001","link":"https://openreview.net/forum?id=fyM0jbe5DK","abs":"

Fine-tuning large language models often undermines their safety alignment, a problem further amplified by harmful fine-tuning attacks in which adversarial data removes safeguards and induces unsafe behaviors. We propose SPARD, a defense framework that integrates Safety-Projected Alternating optimization with Relevance-Diversity aware data selection. SPARD employs SPAG, which optimizes alternatively between utility updates and explicit safety projections with a set of safe data to enforce safety constraints. To curate safe data, we introduce a Relevance–Diversity Determinantal Point Process to select compact safe data, balancing task relevance and safety coverage. Experiments on GSM8K and OpenBookQA under four harmful fine-tuning attacks demonstrate that SPARD consistently achieves the lowest average attack success rates, substantially outperforming state-of-the-art defense methods, while maintaining high task accuracy.

","absKo":"

대규모 language model의 fine-tuning은 종종 safety alignment를 훼손하며, adversarial data가 안전장치를 제거하고 unsafe behavior를 유도하는 harmful fine-tuning attack에서는 이 문제가 더욱 심화된다. 우리는 Safety-Projected Alternating optimization과 Relevance-Diversity aware data selection을 통합한 defense framework인 SPARD를 제안한다. SPARD는 SPAG를 사용하며, SPAG는 utility update와 safe data 집합에 대한 explicit safety projection을 번갈아 최적화하여 safety constraint를 강제한다. 안전한 data를 정제하기 위해 우리는 Relevance–Diversity Determinantal Point Process를 도입하여 task relevance와 safety coverage 사이의 균형을 맞추면서 compact safe data를 선택한다. GSM8K와 OpenBookQA에서 네 가지 harmful fine-tuning attack 하에 수행한 실험은 SPARD가 일관되게 가장 낮은 평균 attack success rate를 달성하며, state-of-the-art defense method를 크게 능가하면서도 높은 task accuracy를 유지함을 보여준다.

"},{"id":"64556","en":"SafeHarbor: Defining Precise Decision Boundaries via Hierarchical Memory-Augmented Guardrail for LLM Agent Safety","ko":"SafeHarbor: 계층적 Memory-Augmented Guardrail을 통한 LLM Agent 안전성을 위한 정밀 결정 경계 정의","authors":"Zhe Liu, Zhe Liu, Wenxin Zhang, Quanchen Zou, Deyue Zhang, Dongdong Yang, Xiangzheng Zhang, Hao Peng","pos":"#3002","link":"https://openreview.net/forum?id=MEKj3d9Bf3","abs":"

With the rapid evolution of foundation models, Large Language Model (LLM) agents have demonstrated increasingly powerful tool-use capabilities. However, this proficiency introduces significant security risks, as malicious actors can manipulate agents into executing tools to generate harmful content. While existing defensive mechanisms are effective, they frequently suffer from the over-refusal problem, where increased safety strictness compromises the agent's utility on benign tasks. To mitigate this trade-off, we propose \\textsc{SafeHarbor}, a novel framework designed to establish precise decision boundaries for LLM agents. Unlike static guidelines, \\textsc{SafeHarbor} extracts context-aware defense rules through enhanced adversarial generation. We design a local hierarchical memory system for dynamic rule injection, offering a training-free, efficient, and plug-and-play solution. Furthermore, we introduce an information entropy-based self-evolution mechanism that continuously optimizes the memory structure through dynamic node splitting and merging. Extensive experiments demonstrate that \\textsc{SafeHarbor} achieves state-of-the-art performance on both ambiguous benign tasks and explicit malicious attacks, notably attaining a peak benign utility of 63.6\\% on GPT-4o while maintaining a robust refusal rate exceeding 93\\% against harmful requests.

","absKo":"

foundation model의 빠른 진화와 함께 Large Language Model(LLM) agent는 점점 더 강력한 tool-use capability를 보여 주고 있다. 그러나 이러한 숙련도는 심각한 보안 위험도 수반하는데, 악의적 행위자가 agent를 조작해 tool을 실행시키고 유해한 콘텐츠를 생성하도록 만들 수 있기 때문이다. 기존 방어 메커니즘은 효과적이지만, 안전 기준을 높일수록 benign task에서 agent의 유용성이 떨어지는 over-refusal problem을 자주 겪는다. 이러한 trade-off를 완화하기 위해 우리는 LLM agent를 위한 정밀한 decision boundary를 구축하도록 설계된 새로운 프레임워크 \\textsc{SafeHarbor}를 제안한다. 정적인 guideline과 달리, \\textsc{SafeHarbor}는 강화된 adversarial generation을 통해 context-aware defense rule을 추출한다. 또한 동적 rule injection을 위한 local hierarchical memory system을 설계하여, training-free이면서 효율적이고 plug-and-play 방식의 해결책을 제공한다. 더 나아가, information entropy 기반 self-evolution mechanism을 도입해 dynamic node splitting과 merging을 통해 memory structure를 지속적으로 최적화한다. 광범위한 실험은 \\textsc{SafeHarbor}가 모호한 benign task와 명시적인 악성 공격 모두에서 state-of-the-art 성능을 달성함을 보여 주며, 특히 GPT-4o에서 benign utility 최고 63.6\\%를 기록하면서도 유해한 요청에 대해 93\\%를 넘는 강건한 refusal rate를 유지한다.

"},{"id":"65816","en":"CAT-Q: Cost-efficient and Accurate Ternary Quantization for LLMs","ko":"CAT-Q: LLM을 위한 비용 효율적이고 정확한 삼진 양자화","authors":"Shigeng Wang, Chao Li, Yangyuxuan Kang, Jiawei Fan, Anbang Yao","pos":"#3006","link":"https://openreview.net/forum?id=9uZJLXt7fq","abs":"

In this paper, we present CAT-Q, Cost-efficient and Accurate Ternary Quantization, to compress LLMs. Unlike current state-of-the-art ternary quantization methods that rely on data-intensive and costly quantization-aware training to mitigate severe performance degradation, CAT-Q employs a simple yet effective post-training quantization scheme, thereby is easily applicable to LLMs with diverse architectures and model sizes. It has two key components, learnable modulation (LM) and softened ternarization (ST), which are coupled from an optimization perspective. LM leverages a composition of learnable factors to modulate the distribution of high-precision weights and the ternary threshold, making them less sensitive to ternarization. ST further introduces a novel transition function to guide the ternarization process toward stable convergence. We show that, for pre-trained LLMs with 1.7B to 8B parameters, CAT-Q can quantize them into ternary models using merely 512 calibration samples, while achieving competitive performance to the seminal BitNet 1.58-bit v1 and v2 families (with 1.3B to 7B parameters) trained with 100B tokens, yielding about a 100,000x reduction in training tokens. Moreover, we show for the first time that CAT-Q can quantize even larger pre-trained LLMs having 14B to 235B parameters into leading ternary models within 8 to 60 hours on 8 A100-80GB GPUs. Code will be made publicly available.

","absKo":"

본 논문에서는 LLM 압축을 위한 Cost-efficient and Accurate Ternary Quantization, 즉 CAT-Q를 제안한다. 기존의 state-of-the-art ternary quantization 방법들이 심각한 성능 저하를 완화하기 위해 data-intensive하고 비용이 많이 드는 quantization-aware training에 의존하는 것과 달리, CAT-Q는 간단하면서도 효과적인 post-training quantization scheme을 사용하므로, 다양한 architecture와 model size를 가진 LLM에 쉽게 적용할 수 있다. 이 방법은 optimization 관점에서 서로 결합된 두 가지 핵심 구성요소, learnable modulation (LM)과 softened ternarization (ST)로 이루어진다. LM은 learnable factor의 조합을 활용하여 high-precision weight와 ternary threshold의 분포를 조절함으로써, 이들이 ternarization에 덜 민감하도록 만든다. ST는 이어서 안정적인 수렴을 향해 ternarization 과정을 유도하는 새로운 transition function을 도입한다. 우리는 1.7B에서 8B parameter를 가진 pre-trained LLM에 대해, CAT-Q가 단지 512개의 calibration sample만으로 이를 ternary model로 양자화할 수 있으며, 100B token으로 학습된(1.3B에서 7B parameter를 가진) seminal BitNet 1.58-bit v1 및 v2 family와 경쟁력 있는 성능을 달성함을 보인다. 이는 training token 수를 약 100,000배 줄인 결과이다. 더 나아가, 우리는 처음으로 CAT-Q가 14B에서 235B parameter를 가진 더 큰 pre-trained LLM조차 8개의 A100-80GB GPU에서 8~60시간 내에 선도적인 ternary model로 양자화할 수 있음을 보인다. 코드는 공개될 예정이다.

"},{"id":"62315","en":"SpecMD: A Comprehensive Study On Speculative Expert Prefetching","ko":"SpecMD: Speculative Expert Prefetching에 대한 종합적 연구","authors":"Duc Hoang, Mohammad Samragh, Ajay Jaiswal, Minsik Cho","pos":"#312","link":"https://openreview.net/forum?id=hjmbstLYc9","abs":"Mixture-of-Experts (MoE) models enable sparse expert activation, meaning that only a subset of the model’s parameters is used during each inference. However, to translate this sparsity into practical performance, an expert caching mechanism is required. Previous works have proposed hardware-centric caching policies, but how these various caching policies interact with each other and different hardware specification remains poorly understood. To address this gap, we develop \\textbf{SpecMD}, a standardized framework for benchmarking ad-hoc cache policies on various hardware configurations. Using SpecMD, we perform an exhaustive benchmarking of several MoE caching strategies, reproducing and extending prior approaches in controlled settings with realistic constraints. Our experiments reveal that MoE expert access is not consistent with temporal locality assumptions (e.g LRU, LFU). Motivated by this observation, we propose \\textbf{Least-Stale}, a novel eviction policy that exploits MoE's predictable expert access patterns to reduce collision misses by up to $85\\times$ over LRU. With such gains, we achieve over $88\\%$ hit rates with up to $34.7\\%$ Time-to-first-token (TTFT) reduction on OLMoE at only $5\\%$ or $0.6GB$ of VRAM cache capacity.","absKo":"Mixture-of-Experts (MoE) model은 sparse expert activation을 가능하게 하며, 이는 각 inference 동안 model parameter의 일부만 사용된다는 뜻이다. 그러나 이러한 sparsity를 실제 성능으로 전환하려면 expert caching mechanism이 필요하다. 기존 연구들은 hardware-centric caching policy를 제안했지만, 이러한 다양한 caching policy가 서로 및 서로 다른 hardware specification과 어떻게 상호작용하는지는 여전히 충분히 이해되지 않았다. 이 격차를 해소하기 위해, 우리는 다양한 hardware configuration에서 ad-hoc cache policy를 benchmarking하기 위한 표준화된 framework인 \\textbf{SpecMD}를 개발한다. SpecMD를 사용하여 우리는 여러 MoE caching strategy를 철저히 benchmark하고, 현실적인 제약이 있는 통제된 설정에서 기존 접근법을 재현하고 확장한다. 우리의 실험은 MoE expert access가 temporal locality 가정(e.g LRU, LFU)과 일치하지 않음을 보여준다. 이 관찰에 착안하여, 우리는 MoE의 예측 가능한 expert access pattern을 활용해 LRU 대비 collision miss를 최대 $85\\\\times$까지 줄이는 새로운 eviction policy인 \\textbf{Least-Stale}을 제안한다. 이러한 향상을 바탕으로, 우리는 OLMoE에서 VRAM cache capacity를 단 $5\\%$ 또는 $0.6GB$만 사용하면서도 $88\\%$가 넘는 hit rate와 최대 $34.7\\%$의 Time-to-first-token (TTFT) 감소를 달성한다."},{"id":"65432","en":"Tracing the Dynamics of Refusal: Exploiting Latent Refusal Trajectories for Robust Jailbreak Detection","ko":"거부의 동역학 추적: 강건한 Jailbreak 탐지를 위한 잠재 거부 궤적 활용","authors":"Xulin Hu, CHE WANG, Wei Yang Bryan Lim, Jianbo Gao, Zhong Chen","pos":"#3205","link":"https://openreview.net/forum?id=DjX8eh4PIm","abs":"Representation Engineering typically relies on static refusal vectors derived from terminal representations. We move beyond this paradigm, demonstrating that refusal is a dynamic and sparse process rather than a localized outcome. Using Causal Tracing, we uncover the Refusal Trajectory—a persistent upstream signature that remains intact even when adversarial attacks (e.g., GCG) suppress terminal signals. Leveraging this, we propose SALO (Sparse Activation Localization Operator), an inference-time detector designed to capture these latent patterns. SALO effectively recovers defense capabilities against forced-decoding attacks, improving detection rates from $\\sim$0% to $>$90% where methods relying on terminal states perform poorly.","absKo":"Representation Engineering은 일반적으로 terminal representation에서 유도된 정적 refusal vector에 의존한다. 우리는 이 패러다임을 넘어, refusal이 국소적인 결과가 아니라 동적이고 sparse한 과정임을 보여준다. Causal Tracing을 사용해, adversarial attack(예: GCG)이 terminal signal을 억제하더라도 그대로 유지되는 upstream signature인 Refusal Trajectory를 밝혀낸다. 이를 활용해 우리는 inference-time detector인 SALO (Sparse Activation Localization Operator)를 제안하며, 이는 이러한 latent pattern을 포착하도록 설계되었다. SALO는 forced-decoding attack에 대한 방어 능력을 효과적으로 복원하며, terminal state에 의존하는 방법들이 성능이 저조한 상황에서 detection rate를 약 0%에서 90% 이상으로 향상시킨다."},{"id":"65822","en":"OSAQ: Outlier Self-Absorption for Accurate Low-bit LLM Quantization","ko":"OSAQ: 정확한 저비트 LLM 양자화를 위한 이상치 자기흡수","authors":"Zhikai Li, Zhen Dong, Xuewen Liu, Jing Zhang, Qingyi Gu","pos":"#3914","link":"https://openreview.net/forum?id=9mo5O8m7vv","abs":"

Large Language Models (LLMs) have demonstrated remarkable capabilities in understanding and generation tasks. However, their massive parameter scale leads to significant resource consumption and latency during inference. Post-training weight-only quantization offers a promising solution by reducing model size and accelerating token generation through alleviating the memory-bound issue. Nevertheless, there are inherent systematic outliers in weights, and although some efforts have attempted to address them, such as scaling and rotation, the performance of low-bit quantization remains far from satisfactory. In this paper, we propose Outlier Self-Absorption Quantization (OSAQ), which performs second-order low-rank derived additive weight suppression for low-bit weight-only LLM quantization. Specifically, we observe that Hessian exhibits low-rank consistency across different inputs, with certain directions persistently lacking strength. Leveraging this property, we construct an additive weight transformation based on the Hessian’s null space, thereby suppressing weight outliers without affecting the task loss. This additive transformation can be absorbed into the weights offline, requiring no inter-layer transformations and introducing no inference overhead. Moreover, the construction is efficiently achieved by a closed-form solution, without resource-intensive training or iterative procedures. Extensive experiments across models of varying scales and tasks are conducted, and the results show that OSAQ effectively suppresses outliers and improves low-bit quantization performance.

","absKo":"

Large Language Models (LLMs)는 이해와 생성 과제에서 놀라운 성능을 보여 왔다. 그러나 막대한 parameter scale 때문에 inference 시 자원 소모와 latency가 크게 증가한다. Post-training weight-only quantization은 memory-bound 문제를 완화함으로써 model size를 줄이고 token generation을 가속할 수 있어 유망한 해법이다. 그럼에도 불구하고 weight에는 본질적인 systematic outlier가 존재하며, scaling과 rotation과 같은 일부 시도가 이를 해결하려 했지만 low-bit quantization의 성능은 여전히 충분하지 않다. 본 논문에서는 low-bit weight-only LLM quantization을 위해 2차 low-rank 유도 additive weight suppression을 수행하는 Outlier Self-Absorption Quantization (OSAQ)을 제안한다. 구체적으로, Hessian이 서로 다른 입력 전반에서 low-rank consistency를 보이며 특정 방향에서는 지속적으로 strength가 부족하다는 것을 관찰한다. 이 성질을 활용해 Hessian의 null space를 기반으로 additive weight transformation을 구성함으로써 task loss에 영향을 주지 않으면서 weight outlier를 억제한다. 이 additive transformation은 오프라인에서 weight에 흡수될 수 있으므로 inter-layer transformation이 필요 없고 inference overhead도 추가되지 않는다. 더 나아가 이 구성은 자원 집약적인 training이나 반복 절차 없이 closed-form solution으로 효율적으로 얻을 수 있다. 다양한 규모의 model과 task에 걸친 광범위한 실험 결과, OSAQ가 outlier를 효과적으로 억제하고 low-bit quantization 성능을 향상시킴을 보여준다.

"},{"id":"65556","en":"Deep sequence models tend to memorize geometrically; it is unclear why.","ko":"심층 시퀀스 모델은 기하학적으로 암기하는 경향이 있다; 그 이유는 불분명하다.","authors":"Shahriar Noroozizadeh, Vaishnavh Nagarajan, Elan Rosenfeld, Sanjiv Kumar","pos":"#1611","link":"https://openreview.net/forum?id=CT2tSmahVQ","abs":"Deep sequence models are said to store atomic facts predominantly in the form of associative memory: a brute-force lookup of co-occurring entities. We identify a dramatically different form of storage of atomic facts that we term as geometric memory. Here, the model has synthesized embeddings encoding novel global relationships between all entities, including ones that do not co-occur in training. Such storage is powerful: for instance, we show how it transforms a hard reasoning task involving an $\\ell$-fold composition into an easy-to-learn $1$-step navigation task. From this phenomenon, we extract fundamental aspects of neural embedding geometries that are hard to explain. We argue that the rise of such a geometry, as against a lookup of local associations, cannot be straightforwardly attributed to typical supervisory, architectural, or optimizational pressures. Counterintuitively, a geometry is learned even when it is more complex than the brute-force lookup. Then, by analyzing a connection to Node2Vec, we demonstrate how the geometry stems from a spectral bias that---in contrast to prevailing theories---indeed arises naturally despite the lack of various pressures. This analysis also points out to practitioners a visible headroom to make Transformer memory more strongly geometric. We hope the geometric view of parametric memory encourages revisiting the default intuitions that guide researchers in areas like knowledge acquisition, capacity, discovery, and unlearning.","absKo":"Deep sequence model은 원자적 사실을 주로 associative memory, 즉 공출현하는 entity에 대한 brute-force lookup 형태로 저장한다고 알려져 있다. 우리는 원자적 사실의 저장이 이와는 극적으로 다른 형태로 나타나는 것을 발견했으며, 이를 geometric memory라고 부른다. 여기서 model은 training에서 함께 등장하지 않은 entity까지 포함하여, 모든 entity 사이의 새로운 global relationship을 인코딩한 embedding을 합성한다. 이러한 저장 방식은 강력하다. 예를 들어, 우리는 $\\ell$-fold composition을 포함하는 어려운 reasoning task가 쉽게 학습되는 $1$-step navigation task로 변환되는 방식을 보인다. 이 현상으로부터 우리는 설명하기 어려운 neural embedding geometry의 근본적 측면을 추출한다. 우리는 local association의 lookup과 달리 이러한 geometry의 등장은 전형적인 supervisory, architectural, optimization pressure로는 단순하게 설명될 수 없다고 주장한다. 직관과 달리, brute-force lookup보다 더 복잡하더라도 geometry는 학습된다. 이어서 Node2Vec와의 연결을 분석함으로써, 우리는 geometry가 spectral bias에서 비롯되며, 이는 지배적인 이론들과 달리 다양한 pressure가 부족함에도 불구하고 실제로 자연스럽게 발생함을 보인다. 이 분석은 또한 practitioner에게 Transformer memory를 더 강하게 geometric하게 만들 수 있는 눈에 보이는 여지를 제시한다. 우리는 parametric memory에 대한 geometric view가 knowledge acquisition, capacity, discovery, unlearning과 같은 영역에서 연구자들을 이끄는 기본 직관을 다시 검토하도록 장려하기를 바란다."},{"id":"62684","en":"Unbiased Alignment for Large Language Models with Noisy Preferences","ko":"노이즈가 있는 선호도에서 LLM의 Unbiased Alignment","authors":"Jialiang Wang, Xianming Liu, Xiong Zhou, Hui Liu, Haoliang Li","pos":"#1712","link":"https://openreview.net/forum?id=eMQsdioK8z","abs":"

The alignment of large language models with human preferences is typically achieved via Reinforcement Learning from Human Feedback or Direct Preference Optimization. However, these methods are susceptible to the significant noise prevalent in real-world preference datasets. To address this critical issue, we present a theoretical framework for unbiased alignment, introducing the Unbiased Reward Model (URM) loss and the Unbiased Direct Preference Optimization (UDPO) loss. These novel objectives allow for the training of unbiased models directly from noisy preferences by mathematically correcting for label noise without requiring clean ground-truth supervision. We provide rigorous theoretical analyses demonstrating that our methods are noise-tolerant, parameter downward compatible, and classification-calibrated. Comprehensive experiments across diverse datasets demonstrate that our approaches outperform state-of-the-art baselines.

","absKo":"

large language model을 human preference에 맞추는 작업은 일반적으로 Reinforcement Learning from Human Feedback 또는 Direct Preference Optimization을 통해 달성된다. 그러나 이러한 방법은 실제 preference dataset에 널리 존재하는 상당한 noise에 취약하다. 이 중대한 문제를 해결하기 위해, 우리는 unbiased alignment를 위한 이론적 framework를 제시하며 Unbiased Reward Model (URM) loss와 Unbiased Direct Preference Optimization (UDPO) loss를 도입한다. 이 새로운 objective는 clean ground-truth supervision 없이도 label noise를 수학적으로 보정하여 noisy preference로부터 직접 unbiased model을 학습할 수 있게 한다. 우리는 우리의 방법이 noise-tolerant하고, parameter downward compatible하며, classification-calibrated임을 보이는 엄밀한 이론 분석을 제공한다. 다양한 dataset 전반의 포괄적인 실험은 우리의 접근법이 state-of-the-art baseline을 능가함을 보여준다.

"},{"id":"63833","en":"Statistical Early Stopping for Reasoning Models","ko":"추론 모델을 위한 통계적 조기 종료","authors":"Yangxinyu Xie, Tao Wang, Soham Mallick, Yan Sun, Georgy Noarov, Mengxin Yu, Tanwi Mallick, Edgar Dobriban","pos":"#1800","link":"https://openreview.net/forum?id=TJshRZDdyW","abs":"

While LLMs have seen substantial improvement in reasoning capabilities, they also sometimes overthink, generating unnecessary reasoning steps, particularly under uncertainty, given ill-posed or ambiguous queries. We introduce statistically principled early stopping methods that monitor uncertainty signals during generation to mitigate this issue. Our first approach is nonparametric and provides finite-sample guarantees on the probability of halting too early on well-posed queries. Our second approach is parametric: it models inter-arrival times of uncertainty keywords as a renewal process and applies sequential testing for stopping. We conduct empirical evaluations on reasoning tasks across several domains and models. Our results indicate that uncertainty-aware early stopping can improve both efficiency and reliability in LLM reasoning. The performance varies across domains, and we observe especially significant gains for math reasoning.

","absKo":"

LLM은 reasoning 능력에서 크게 향상되었지만, 불확실한 상황에서 특히 ill-posed이거나 ambiguous한 query에 대해 불필요한 reasoning step을 생성하며 과도하게 생각하는 경우도 있다. 우리는 생성 중 불확실성 신호를 모니터링하여 이 문제를 완화하는 통계적으로 정당화된 early stopping 방법을 제안한다. 첫 번째 접근법은 nonparametric이며, well-posed query에서 너무 일찍 중단할 확률에 대해 finite-sample guarantee를 제공한다. 두 번째 접근법은 parametric으로, 불확실성 키워드의 inter-arrival time을 renewal process로 모델링하고 stopping을 위해 sequential testing을 적용한다. 우리는 여러 도메인과 model에서 reasoning task에 대한 실험 평가를 수행한다. 결과는 uncertainty-aware early stopping이 LLM reasoning의 효율성과 신뢰성을 모두 개선할 수 있음을 보여준다. 성능은 도메인에 따라 달라지며, 특히 수학 reasoning에서 두드러진 향상을 관찰한다.

"},{"id":"64494","en":"Syntax vs. Semantics: How Transformers Learn Deep Dependencies","ko":"구문 대 의미론: Transformer가 깊은 의존성을 학습하는 방식","authors":"Jiangrui Zhao, Xiaoting Du","pos":"#1803","link":"https://openreview.net/forum?id=N0NjjQ9V3w","abs":"

Large Language Models demonstrate remarkable syntactic fluency, yet the optimization dynamics governing their acquisition of deep semantic dependencies remain poorly understood. We propose a mechanistic framework that models this learning process as a competition between Surface Statistics and Deep Semantics. Our theoretical analysis identifies a ``Gradient Starvation\" phenomenon where the error signals for sparse semantic dependencies are actively suppressed during early optimization. This suppression impedes the learning of structural reasoning and causes its emergence to manifest as a sudden phase transition. Furthermore, this framework offers a mechanistic basis for the effectiveness of Chain-of-Thought (CoT) strategies. By externalizing intermediate reasoning steps into concrete tokens, CoT effectively bypasses the suppression regime inherent to implicit reasoning. We validate these findings across scales ranging from toy transformers to production models (Llama-3.1-8B, Qwen2.5-Coder-7B). Finally, guided by this theory, we propose a topology-aligned contrastive objective that explicitly rectifies the gradient geometry. Experiments on variable binding tasks demonstrate that our method achieves an improvement that is over 2× larger than that obtained via standard cross-entropy fine-tuning.

","absKo":"Large Language Models는 놀라운 syntactic fluency를 보여주지만, 깊은 semantic dependency를 습득하는 과정을 지배하는 optimization dynamics는 여전히 잘 이해되지 않고 있습니다. 우리는 이 학습 과정을 Surface Statistics와 Deep Semantics 사이의 경쟁으로 모델링하는 mechanistic framework를 제안합니다. 이론적 분석을 통해, 희소한 semantic dependency에 대한 error signal이 초기 optimization 동안 적극적으로 억제되는 ``Gradient Starvation'' 현상을 식별합니다. 이러한 억제는 structural reasoning의 학습을 방해하며, 그 출현이 갑작스러운 phase transition처럼 나타나게 만듭니다. 또한 이 framework는 Chain-of-Thought(CoT) 전략의 효과에 대한 mechanistic basis를 제공합니다. 중간 추론 단계를 concrete token으로 외재화함으로써, CoT는 implicit reasoning에 내재된 suppression regime을 효과적으로 우회합니다. 우리는 toy transformer부터 production model(Llama-3.1-8B, Qwen2.5-Coder-7B)에 이르는 다양한 규모에서 이러한 결과를 검증합니다. 마지막으로, 이 이론에 기반해 gradient geometry를 명시적으로 수정하는 topology-aligned contrastive objective를 제안합니다. variable binding task 실험은 우리의 방법이 표준 cross-entropy fine-tuning으로 얻은 향상보다 2배 이상 큰 개선을 달성함을 보여줍니다.

"},{"id":"66378","en":"Toward Cybersecurity-Expert Small Language Models","ko":"사이버보안 전문가 소형 언어 모델을 향하여","authors":"Matan Levi, Daniel Ohayon, Ariel Blobstein, Ravid Sa, Ian Molloy, Yair Allouche","pos":"#1812","link":"https://openreview.net/forum?id=4Ec15Uj8Nd","abs":"

Large language models (LLMs) are transforming everyday applications, yet they lag behind in specialized fields, such as cybersecurity, due to a lack of high-quality, domain-specific models and training datasets. To address this gap, we present CyberPal 2.0, a family of cybersecurity-expert small language models (SLMs) ranging from 4B–20B parameters. To train CyberPal 2.0, we generate an enriched chain-of-thought cybersecurity instruction dataset built with our data enrichment and formatting pipeline, SecKnowledge 2.0, which integrates expert-in-the-loop steering of reasoning formats alongside LLM-driven multi-step grounding, yielding higher-fidelity, task-grounded reasoning traces for security tasks. Across diverse cybersecurity benchmarks, CyberPal 2.0 consistently outperforms its baselines and matches or surpasses various open and closed-source frontier models, while remaining a fraction of their size. On core threat-investigation tasks, such as correlating vulnerabilities and bug tickets with weaknesses, our best 20B-parameter model outperforms GPT-4o, o1, o3-mini, and Sec-Gemini v1, ranking first, while our smallest 4B-parameter model ranks second. On core cyber threat intelligence knowledge tasks, our models outperform almost all tested frontier models, ranking second only to Sec-Gemini v1. To foster reproducibility and practical adoption, we will release our models as open source.

","absKo":"

대규모 언어 모델(LLMs)은 일상적인 애플리케이션을 변화시키고 있지만, 고품질의 도메인 특화 모델과 학습 데이터셋이 부족해 cybersecurity 같은 특수 분야에서는 여전히 뒤처져 있다. 이러한 격차를 해소하기 위해, 우리는 4B–20B 파라미터 범위의 cybersecurity 전문가 소형 언어 모델(SLMs) 패밀리인 CyberPal 2.0을 제시한다. CyberPal 2.0을 학습하기 위해 우리는 데이터 증강 및 포맷팅 파이프라인 SecKnowledge 2.0으로 구축한 향상된 chain-of-thought cybersecurity instruction 데이터셋을 생성했으며, 이는 expert-in-the-loop에 의한 reasoning 형식 제어와 LLM 기반 다단계 grounding을 결합하여, security task에 대해 더 높은 fidelity의 task-grounded reasoning trace를 제공한다. 다양한 cybersecurity benchmark 전반에서 CyberPal 2.0은 일관되게 baseline을 능가하고 여러 open-source 및 closed-source frontier model과 동등하거나 그 이상을 달성하며, 그 크기는 훨씬 작다. 핵심 threat-investigation task, 예를 들어 vulnerability와 bug ticket을 weakness와 연결하는 작업에서, 우리의 최고 20B-parameter 모델은 GPT-4o, o1, o3-mini, Sec-Gemini v1을 능가하며 1위를 차지했고, 가장 작은 4B-parameter 모델은 2위를 기록했다. 핵심 cyber threat intelligence knowledge task에서는 우리 모델들이 테스트된 거의 모든 frontier model보다 우수했으며, Sec-Gemini v1에만 이어 2위를 차지했다. 재현성과 실용적 도입을 촉진하기 위해, 우리는 이 모델들을 open source로 공개할 예정이다.

"},{"id":"66393","en":"Towards Understanding Continual Factual Knowledge Acquisition of Language Models: From Theory to Algorithm","ko":"언어 모델의 지속적 사실 지식 습득 이해를 향하여: 이론에서 알고리즘까지","authors":"Haoyu Wang, yifan shang, Zhongxiang Sun, Weijie Yu, Xiao Zhang, Jun Xu","pos":"#1813","link":"https://openreview.net/forum?id=47xV4icN1B","abs":"

Continual Pre-Training (CPT) is essential for enabling Language Models (LMs) to integrate new factual knowledge without erasing old. While classical CPT techniques like data replay have become the standard paradigm, the mechanisms underlying how LMs acquire and retain facts over time, termed as continual Factual Knowledge Acquisition (cFKA), remain unclear. In this work, we present a theoretical framework that characterizes the training dynamics of cFKA using a single-layer Transformer with linear attention, offering a unified explanation for the behavior of popular CPT methods. Our analysis reveals that regularization-based methods merely adjust the convergence rate of parameters without altering the inherent forgetting tendency, whereas data replay methods shift convergence dynamics and stabilize pretrained knowledge. Building on these insights, we propose a novel generative data replay approach, called Selecting Tokens via attentiOn Contribution (STOC), which identifies influential factual snippets to guide replay generation. Extensive experiments on both synthetic and real-world datasets validate our theoretical findings and demonstrate that STOC effectively enhances cFKA by mitigating catastrophic forgetting.

","absKo":"

Continual Pre-Training (CPT)은 Language Models (LMs)가 오래된 정보를 지우지 않으면서 새로운 factual knowledge를 통합할 수 있게 하는 데 필수적이다. data replay와 같은 고전적인 CPT 기법이 표준 paradigm이 되었지만, LMs가 시간에 따라 fact를 획득하고 보존하는 메커니즘, 즉 continual Factual Knowledge Acquisition (cFKA)은 여전히 명확하지 않다. 이 연구에서는 linear attention을 사용하는 single-layer Transformer를 통해 cFKA의 training dynamics를 특징짓는 이론적 framework를 제시하여, 널리 사용되는 CPT method들의 동작에 대한 통합된 설명을 제공한다. 우리의 분석은 regularization-based method가 parameter의 고유한 forgetting tendency를 바꾸지 않은 채 수렴 속도만 조정하는 반면, data replay method는 수렴 dynamics를 변화시키고 pretrained knowledge를 안정화한다는 점을 보여준다. 이러한 통찰을 바탕으로, 우리는 Selecting Tokens via attentiOn Contribution (STOC)이라 불리는 새로운 generative data replay approach를 제안한다. STOC는 replay generation을 안내할 영향력 있는 factual snippet을 식별한다. synthetic dataset과 real-world dataset 모두에서 수행한 광범위한 실험은 우리의 이론적 발견을 검증하며, STOC가 catastrophic forgetting을 완화함으로써 cFKA를 효과적으로 향상시킨다는 점을 보여준다.

"},{"id":"64413","en":"Scaling Generative Verifiers For Natural Language Mathematical Proof Verification And Selection","ko":"자연어 수학 증명 검증 및 선택을 위한 생성적 검증기 확장","authors":"Sadegh Mahdavi, Branislav Kisacanin, Shubham Toshniwal, Wei Du, Ivan Moshkov, George Armstrong, Renjie Liao, Christos Thrampoulidis, Igor Gitman","pos":"#1908","link":"https://openreview.net/forum?id=NsC3NFsavi","abs":"

Large language models have achieved remarkable success on final-answer mathematical problems, largely due to the ease of applying reinforcement learning with verifiable rewards. However, the reasoning underlying these solutions is often flawed. Advancing to rigorous proof-based mathematics requires reliable proof verification capabilities. We begin by analyzing multiple evaluation setups and show that focusing on a single benchmark can lead to brittle or misleading conclusions. To address this, we evaluate both proof-based and final-answer reasoning to obtain a more reliable measure of model performance. We then scale two major generative verification methods (GenSelect and LLM-as-a-Judge) to millions of tokens and identify their combination as the most effective framework for solution verification and selection. We further show that the choice of prompt for LLM-as-a-Judge significantly affects the model's performance, but reinforcement learning can reduce this sensitivity. However, despite improving proof-level metrics, reinforcement learning does not enhance final-answer precision, indicating that current models often reward stylistic or procedural correctness rather than mathematical validity. Our results establish practical guidelines for designing and evaluating scalable proof-verification and selection systems.

","absKo":"

대규모 언어 모델은 최종 답만을 요구하는 수학 문제에서 놀라운 성과를 거두었는데, 이는 주로 검증 가능한 보상을 활용한 reinforcement learning을 쉽게 적용할 수 있기 때문이다. 그러나 이러한 해법의 배후에 있는 추론은 종종 결함이 있다. 엄밀한 증명 기반 수학으로 나아가려면 신뢰할 수 있는 proof verification 능력이 필요하다. 우리는 먼저 여러 평가 설정을 분석하고, 단일 benchmark에만 집중하면 취약하거나 오해의 소지가 있는 결론에 이를 수 있음을 보인다. 이를 해결하기 위해 우리는 proof 기반 추론과 final-answer 추론을 모두 평가하여 모델 성능을 더 신뢰할 수 있게 측정한다. 그런 다음 두 가지 주요 generative verification 방법(GenSelect와 LLM-as-a-Judge)을 수백만 token 규모로 확장하고, 이 둘의 조합이 해법 검증과 선택에 가장 효과적인 프레임워크임을 확인한다. 또한 LLM-as-a-Judge에 사용되는 prompt의 선택이 모델 성능에 큰 영향을 미치지만, reinforcement learning이 이러한 민감도를 줄일 수 있음을 보인다. 그러나 proof 수준 지표는 개선되더라도 reinforcement learning은 final-answer 정확도를 향상시키지 못하며, 이는 현재 모델들이 종종 수학적 타당성보다 문체적이거나 절차적인 정답성을 보상한다는 점을 시사한다. 우리의 결과는 확장 가능한 proof verification 및 selection 시스템을 설계하고 평가하기 위한 실용적 지침을 제시한다.

"},{"id":"65565","en":"Sampling from Your Language Model One Byte at a Time","ko":"한 번에 한 바이트씩 언어 모델에서 샘플링하기","authors":"Jonathan Hayase, Alisa Liu, Noah Smith, Sewoong Oh","pos":"#1909","link":"https://openreview.net/forum?id=COQV7D1dAW","abs":"

Tokenization is used almost universally by modern language models, enabling efficient text representation using multi-byte or multi-character tokens. However, prior work has shown that tokenization can introduce distortion into the model’s generations, an issue known as the Prompt Boundary Problem (PBP). For example, users are often advised not to end their prompts with a space because it prevents the model from including the space as part of the next token. While this heuristic is effective in English, the underlying PBP continues to affect languages such as Chinese as well as code generation, where tokens often do not line up with word and syntactic boundaries. In this work, we present an inference-time method to convert any autoregressive LM with a BPE tokenizer into a character-level or byte-level LM. Our method efficiently solves the PBP and is also able to unify the vocabularies of language models with different tokenizers, allowing one to ensemble LMs with different tokenizers at inference time or transfer the post-training from one model to another using proxy-tuning. We demonstrate in experiments that the ensemble and proxy-tuned models outperform their constituents on downstream evals

","absKo":"

Tokenization은 현대 language model에서 거의 보편적으로 사용되며, multi-byte 또는 multi-character token을 통해 효율적인 text representation을 가능하게 한다. 그러나 선행 연구는 tokenization이 모델 생성 결과에 distortion을 유발할 수 있음을 보였고, 이는 Prompt Boundary Problem (PBP)이라 불린다. 예를 들어, 사용자는 prompt를 space로 끝내지 말라고 자주 권고받는데, 이는 모델이 그 space를 다음 token의 일부로 포함하지 못하게 하기 때문이다. 이러한 heuristic은 English에서는 효과적이지만, 근본적인 PBP는 Chinese뿐 아니라 token이 단어 및 구문 경계와 잘 맞지 않는 code generation에도 계속 영향을 미친다. 본 연구에서는 어떤 autoregressive LM이든 BPE tokenizer를 사용한다면 inference-time에서 character-level 또는 byte-level LM으로 변환하는 방법을 제시한다. 우리의 방법은 PBP를 효율적으로 해결하며, 서로 다른 tokenizer를 가진 language model들의 vocabulary를 통합할 수도 있어, inference time에 서로 다른 tokenizer를 사용하는 LM들을 ensemble하거나 proxy-tuning을 사용해 한 모델의 post-training을 다른 모델로 transfer할 수 있게 한다. 실험을 통해 ensemble 및 proxy-tuned model이 downstream evals에서 구성 요소들을 능가함을 보인다

"},{"id":"66007","en":"SIMoE: A Probabilistic Framework for Cardinality-Constrained Routing in Mixture-of-Experts","ko":"SIMoE: Mixture-of-Experts에서 기수 제약 라우팅을 위한 확률적 프레임워크","authors":"Heng Zhao, Zilei Shao, Guy Van den Broeck, Zhe Zeng","pos":"#1912","link":"https://openreview.net/forum?id=7zOtnPt85B","abs":"Mixture-of-Experts (MoE) models scale by activating only a small subset of experts per token, but standard deterministic top-$k$ routing is non-differentiable and trained using surrogate gradients that ignore the discrete expert selection used at inference. We introduce SIMoE routing by modeling expert selection as a stochastic latent variable, casting it as probabilistic inference over discrete expert subsets under explicit cardinality constraints. First, we purpose SIMoE Exact-$k$ routing, which samples discrete $k$-expert subsets and propagates gradients through tractable inclusion probabilities for stochastic expert selection. We then extend this to SIMoE Dynamic-$k$ routing, where both training and inference constrain the routing cardinality to the same predefined range, allowing adaptive expert allocation per token. Across different benchmarks on OLMoE and Qwen MoE backbones, SIMoE Exact-$k$ improves expert utilization and routing diversity over competitive baselines, and Dynamic-$k$ achieves comparable performance with fewer activated experts.","absKo":"Mixture-of-Experts (MoE) model은 token마다 소수의 expert만 활성화함으로써 확장되지만, 표준 deterministic top-$k$ routing은 미분 가능하지 않으며, 추론 시 사용되는 discrete expert selection을 무시하는 surrogate gradient로 학습된다. 우리는 expert selection을 stochastic latent variable로 모델링하여, 명시적인 cardinality constraint 하에서 discrete expert subset에 대한 probabilistic inference로 해석하는 SIMoE routing을 도입한다. 먼저 우리는 SIMoE Exact-$k$ routing을 제안하는데, 이는 discrete한 $k$-expert subset을 샘플링하고, stochastic expert selection을 위해 tractable한 inclusion probability를 통해 gradient를 전달한다. 이어서 우리는 SIMoE Dynamic-$k$ routing으로 확장하는데, 여기서는 학습과 추론 모두 routing cardinality를 동일한 사전 정의 범위로 제한하여 token별로 적응적인 expert allocation을 가능하게 한다. OLMoE와 Qwen MoE backbone에 대한 다양한 benchmark에서, SIMoE Exact-$k$는 경쟁력 있는 baseline보다 expert 활용도와 routing 다양성을 향상시키며, Dynamic-$k$는 더 적은 활성 expert로도 비슷한 성능을 달성한다."},{"id":"61292","en":"Bits That Count: Quantifying and Predicting Capabilities of Language Models","ko":"중요한 비트: Language Model의 역량 정량화와 예측","authors":"Elizabeth Donoway, Hailey Joren, Michael R DeWeese, Ethan Perez, John Schulman, Fabien Roger, Jan Leike","pos":"#200","link":"https://openreview.net/forum?id=re9NtTb1Dy","abs":"

What and how do language models learn during training? When does learning elicit \\textit{existing} knowledge, and when does it primarily teach \\textit{new} capabilities? We find that the amount of generalizable information language models learn during training predicts the origins of their emergent capabilities. Minuscule amounts of information---in many cases, a few bits in a single example---can unlock large fractions of models' maximum performance when capabilities are \\textit{elicited} rather than \\textit{taught}. We quantify these learning regimes using excess description length (EDL), an information-theoretic measure of generalizable information learned during training. We find that elicitation and teaching exhibit distinct EDL signatures that characterize the predominant learning mechanism as information scales: elicitation requires orders of magnitude less information than teaching to comparable performance. We demonstrate that EDL provides a practical tool for quantitatively estimating the maximum amount of predictive information models can compress from data into trainable parameters during learning. These capacity limits describe optimal tradeoffs between data and parameter count that robustly predict when parameter-efficient fine-tuning methods (\\textit{e.g.}, LoRA) will underperform full fine-tuning.

","absKo":"

언어 모델은 학습 동안 무엇을, 어떻게 배우는가? 학습은 언제 \\textit{existing} knowledge를 끌어내고, 언제 주로 \\textit{new} capabilities를 가르치는가? 우리는 학습 중 언어 모델이 습득하는 generalizable information의 양이 emergent capability의 기원을 예측한다는 사실을 발견했다. 극히 적은 양의 정보, 많은 경우 단일 example에서 몇 bit에 불과한 정보만으로도, 능력이 \\textit{taught}되기보다 \\textit{elicited}될 때 모델의 최대 성능 중 상당 부분을 끌어낼 수 있다. 우리는 이러한 learning regime를 training 동안 학습된 generalizable information의 information-theoretic measure인 excess description length (EDL)로 정량화한다. 우리는 elicitation과 teaching이 서로 다른 EDL signature를 보이며, 정보가 증가함에 따라 지배적인 학습 메커니즘을 특징짓는다는 것을 발견했다. elicitation은 유사한 성능에 도달하는 데 teaching보다 몇 자릿수(order of magnitude) 적은 정보를 필요로 한다. 우리는 EDL이 학습 중 데이터로부터 trainable parameter에 압축될 수 있는 예측 정보의 최대량을 정량적으로 추정하는 실용적 도구임을 보인다. 이러한 capacity limit는 데이터와 parameter 수 사이의 최적 tradeoff를 설명하며, parameter-efficient fine-tuning method(\\textit{e.g.}, LoRA)가 full fine-tuning보다 성능이 떨어질 시점을 견고하게 예측한다.

"},{"id":"65410","en":"Probing How Scalable Table Data Enhances General Long-Context Reasoning","ko":"확장 가능한 표 데이터가 일반 장문 맥락 추론을 향상시키는 방식 탐구","authors":"Huaibing Xie, Guoliang Zhao, Yang Liu, Shihan Dou, Siming Huang, Yanling Xiao, Shaolei Wang, Yiting Liu, Cheng Zhang, Shaofan Liu, PlutoZhou","pos":"#2003","link":"https://openreview.net/forum?id=Dx3Z5QGAZB","abs":"

As real-world tasks grow increasingly complex, long-context reasoning has become a core capability for Large Language Models (LLMs). However, few studies explore which data types are effective for long-context reasoning and why. We find that structured table data with periodic structures shows strong potential for long-context reasoning. Motivated by this observation, we mathematically analyze tabular dependency structures using mutual information, revealing periodic non-vanishing dependencies in table data. Furthermore, we systematically analyze the capabilities of structured table data, conduct relevant scaling experiments, and validate its underlying mechanisms for enhancing long-context reasoning, yielding several meaningful insights. Leveraging these insights, we propose a simple yet scalable pipeline(TableLong) for synthesizing high-quality, diverse, and verifiable structured table data to boost long-context reasoning via RL. Extensive experimental results demonstrate that table data significantly enhances the long-context reasoning capability of LLMs across multiple long-context benchmarks (+8.24\\% on average), and even improves performance on out-of-domain benchmarks (+8.06\\% on average). We hope that our insights provide practical guidance for effective post-training data to enhance long-context reasoning in LLMs.

","absKo":"실제 세계의 task가 점점 더 복잡해짐에 따라, long-context reasoning은 Large Language Model(LLM)의 핵심 능력이 되었다. 그러나 어떤 data type이 long-context reasoning에 효과적인지, 그리고 왜 그런지에 대한 연구는 많지 않다. 우리는 periodic structure를 가진 structured table data가 long-context reasoning에 강한 잠재력을 보인다는 사실을 발견했다. 이러한 관찰에 동기부여를 받아, 우리는 mutual information을 사용해 tabular dependency structure를 수학적으로 분석하여 table data에서 주기적이고 소멸하지 않는 dependency를 밝혀냈다. 더 나아가, structured table data의 능력을 체계적으로 분석하고 관련 scaling experiment를 수행했으며, long-context reasoning을 향상시키는 underlying mechanism을 검증하여 여러 의미 있는 통찰을 도출했다. 이러한 통찰을 활용해, 우리는 RL을 통해 long-context reasoning을 강화하기 위해 고품질이고 다양한, 검증 가능한 structured table data를 합성하는 간단하면서도 확장 가능한 pipeline(TableLong)을 제안한다. 광범위한 실험 결과는 table data가 여러 long-context benchmark에서 LLM의 long-context reasoning 능력을 크게 향상시키고(평균 +8.24\\%), out-of-domain benchmark에서도 성능을 개선함을 보여준다(평균 +8.06\\%). 우리는 이러한 통찰이 LLM의 long-context reasoning을 향상시키는 효과적인 post-training data에 대한 실질적인 지침을 제공하기를 바란다.

"},{"id":"63923","en":"Reasoning Structure of Large Language Models","ko":"대규모 language model의 추론 구조","authors":"Frédéric Berdoz, Luca Lanzendörfer, Fabian Farestam, Roger Wattenhofer","pos":"#2014","link":"https://openreview.net/forum?id=SX7LSFG4IH","abs":"

Large reasoning models (LRMs) are often evaluated using metrics such as final-answer accuracy or token count. However, identical scores on these metrics can hide fundamentally different reasoning structures. To address this limitation, we introduce a scalable LRM benchmark of logic puzzles and a pipeline that converts unstructured traces into verifiable reasoning graphs of claims and dependencies. This turns reasoning into a structured, measurable object whose topology can be quantitatively analyzed. Building on this, we define a reasoning efficiency metric that quantifies how concentrated the model's logical flow is. Our analysis on open-source reasoning models shows that structural measurements separate behaviors that token count and accuracy conflate, providing a practical tool for diagnosing failure modes and comparing how reasoning scales with puzzle difficulty.

","absKo":"

Large reasoning models (LRMs)는 종종 final-answer accuracy나 token count 같은 metric으로 평가됩니다. 그러나 이러한 metric에서 동일한 점수는 근본적으로 다른 reasoning structure를 가릴 수 있습니다. 이 한계를 해결하기 위해 우리는 논리 퍼즐로 구성된 확장 가능한 LRM benchmark와, 비구조화된 trace를 claim과 dependency로 이루어진 검증 가능한 reasoning graph로 변환하는 pipeline을 도입합니다. 이를 통해 reasoning을 위상(topology)을 정량적으로 분석할 수 있는 구조화되고 측정 가능한 객체로 바꿉니다. 이를 바탕으로 우리는 model의 logical flow가 얼마나 집중되어 있는지를 정량화하는 reasoning efficiency metric을 정의합니다. 공개 reasoning model에 대한 분석 결과는 token count와 accuracy가 혼동하는 행동을 구조적 측정이 구분해 준다는 점을 보여주며, failure mode를 진단하고 puzzle 난이도에 따라 reasoning이 어떻게 확장되는지 비교하는 실용적 도구를 제공합니다.

"},{"id":"61802","en":"On the Generalization Gap in Self-Evolving Language Model Reasoning","ko":"자가 진화 Language Model 추론의 일반화 격차에 관하여","authors":"Zhenting Qi, Susanna Maria Baby, Stefanie Baby, Kan Yuan, Da-Cheng Juan, Tu Vu, Andrew Tomkins, Cyrus Rashtchian","pos":"#2106","link":"https://openreview.net/forum?id=mnUidYi5qO","abs":"

Recent work suggests that LLMs can improve their abilities through \\textit{self-evolution}, using only internally generated supervision. A central open question, however, is not whether self-evolution can help, but: \\textit{how far is it from oracle-supervised training under minimal assumptions?} To address this question, we present a controlled empirical analysis of LLM self-evolution under a strict formulation: self-evolution is allowed access only to (i) an unlabeled prompt set and (ii) a base language model, with all supervision signals generated from this model. Under this formulation, we can evaluate many self-evolution approaches in a unified preference optimization framework. Specifically, we analyze four representative self-evolution methods, ranging from single-round verification to multi-turn feedback, iterative training, and curriculum learning. For our primary analysis, we use a clean setting based on the Knights and Knaves logical reasoning dataset, which provides deterministic solutions, systematic verification, and a hierarchy of difficulty levels that enables an evaluation of easy-to-hard generalization. Across this controlled setting, we find that increasingly complex self-evolution strategies yield consistent but limited gains. In general, a substantial performance gap persists relative to oracle supervision. One strategy stands out as effective: we nearly match the oracle performance by using a larger model (Gemma 12B) with iterative revision based on natural language feedback. We also study self-evolution on the OpenThoughts reasoning corpus and evaluate on standard problem-solving benchmarks. In this regime, self-evolution only leads to modest improvements, including when using more resource-intensive strategies or online RL. Overall, our results shed new light on the empirical limits of various types of self-evolution.

","absKo":"

최근 연구는 LLM이 self-evolution, 즉 내부적으로 생성된 supervision만을 사용해 능력을 향상시킬 수 있음을 시사한다. 그러나 핵심적인 미해결 질문은 self-evolution이 도움이 되는지 여부가 아니라, 최소한의 가정 하에서 oracle-supervised training과 얼마나 차이가 나는가이다. 이 질문에 답하기 위해, 우리는 엄격한 정식화 아래에서 LLM self-evolution에 대한 통제된 실증 분석을 제시한다. 이 정식화에서는 self-evolution이 (i) unlabeled prompt set과 (ii) base language model에만 접근할 수 있으며, 모든 supervision signal은 이 모델에서 생성된다. 이 정식화 아래에서 우리는 하나의 통일된 preference optimization framework 안에서 다양한 self-evolution 접근법을 평가할 수 있다. 구체적으로, 우리는 single-round verification부터 multi-turn feedback, iterative training, curriculum learning에 이르는 네 가지 대표적 self-evolution 방법을 분석한다. 주요 분석에서는 deterministic solution, systematic verification, easy-to-hard generalization 평가를 가능하게 하는 난이도 계층을 제공하는 Knights and Knaves logical reasoning dataset을 기반으로 한 clean setting을 사용한다. 이 통제된 설정 전반에서, 점점 더 복잡한 self-evolution 전략이 일관되지만 제한적인 향상을 낳는다는 것을 발견했다. 일반적으로 oracle supervision에 비해 상당한 성능 격차가 여전히 남아 있다. 한 가지 전략은 특히 효과적이었는데, 자연어 feedback에 기반한 iterative revision을 사용하는 더 큰 모델(Gemma 12B)로 거의 oracle performance에 도달했다. 또한 우리는 OpenThoughts reasoning corpus에서 self-evolution을 연구하고 standard problem-solving benchmark에서 평가했다. 이 regime에서는 더 많은 자원을 요구하는 전략이나 online RL을 사용하더라도 self-evolution은 대체로 소폭의 개선만을 가져왔다. 전반적으로 우리의 결과는 다양한 유형의 self-evolution의 실증적 한계에 대한 새로운 통찰을 제공한다.

"},{"id":"63897","en":"On the Entropy Dynamics in Reinforcement Fine-Tuning of Large Language Models","ko":"대규모 language model의 reinforcement fine-tuning에서의 엔트로피 동역학에 대하여","authors":"Shumin Wang, Yuexiang Xie, Wenhao Zhang, Yuchang Sun, Yanxi Chen, Yaliang Li, Yanyong Zhang","pos":"#2107","link":"https://openreview.net/forum?id=Sj8NBI6aV0","abs":"

Entropy serves as a critical metric for measuring the diversity of outputs generated by large language models (LLMs), providing valuable insights into their exploration capabilities. While recent studies increasingly focus on monitoring and adjusting entropy to better balance exploration and exploitation in reinforcement fine-tuning (RFT), a principled understanding of entropy dynamics during this process is yet to be thoroughly investigated. In this paper, we establish a theoretical framework for analyzing the entropy dynamics during the RFT process, which begins with a discriminant expression that quantifies entropy change under a single logit update. This foundation enables the derivation of a first-order expression for entropy change, which can be further extended to the update formula of Group Relative Policy Optimization (GRPO). The corollaries and insights drawn from the theoretical analysis inspire the design of entropy control methods, and also offer a unified lens for interpreting various entropy-based methods in existing studies. We provide empirical evidence to support the main conclusions of our analysis and demonstrate the effectiveness of the derived entropy-discriminator clipping methods. This study yields novel insights into RFT training dynamics, providing theoretical support and practical strategies for optimizing the exploration-exploitation balance during LLM fine-tuning.

","absKo":"

Entropy는 대규모 언어 모델(LLMs)이 생성하는 출력의 다양성을 측정하는 데 중요한 지표로, 이들의 exploration capability에 대한 유용한 통찰을 제공한다. 최근 연구는 reinforcement fine-tuning (RFT)에서 exploration과 exploitation의 균형을 더 잘 맞추기 위해 entropy를 모니터링하고 조정하는 데 점점 더 주목하고 있지만, 이 과정에서의 entropy dynamics에 대한 원리적 이해는 아직 충분히 탐구되지 않았다. 본 논문에서는 단일 logit update 하에서 entropy 변화를 정량화하는 discriminant expression에서 시작하는 RFT 과정 동안의 entropy dynamics를 분석하기 위한 이론적 프레임워크를 확립한다. 이 기반 위에서 entropy change에 대한 first-order expression을 도출하며, 이는 더 나아가 Group Relative Policy Optimization (GRPO)의 update formula로 확장될 수 있다. 이론적 분석에서 도출된 corollary와 통찰은 entropy control method의 설계를 이끌며, 기존 연구의 다양한 entropy-based method를 해석하기 위한 통합적 관점을 제공한다. 우리는 분석의 주요 결론을 지지하는 실증적 증거를 제시하고, 도출된 entropy-discriminator clipping method의 효과를 입증한다. 본 연구는 RFT 학습 dynamics에 대한 새로운 통찰을 제공하며, LLM fine-tuning 동안 exploration-exploitation 균형을 최적화하기 위한 이론적 기반과 실용적 전략을 제시한다.

"},{"id":"66093","en":"Inference-Time Conformal Reasoning with Valid Factuality Control for Large Language Models","ko":"대형 언어 모델을 위한 유효 사실성 제어를 갖춘 추론 시점 Conformal 추론","authors":"Ting Wang, Yuanjie Shi, Yan Yan, Huan Zhang","pos":"#2200","link":"https://openreview.net/forum?id=76iNFmvmBa","abs":"

Large language models (LLMs) increasingly perform multi-step reasoning, where intermediate claims form implicit directed acyclic graphs whose node correctness is structurally conditioned on their ancestors. This makes factuality uncertainty structural, rather than a trivial accumulation of node-wise errors, and necessitates inference-time uncertainty quantification over the reasoning structure. While conformal prediction (CP) offers flexible user-specified factuality control, existing work remains post-hoc and cannot intervene during generation. To fill the gap between CP’s flexibility and its post-hoc limitation, we propose an Inference-Time Conformal Reasoning (ITCR) framework that integrates CP directly into reasoning graph generation. ITCR learns a structure-level factuality uncertainty function that aggregates claim-level factuality signals over reasoning graphs without complex modeling assumptions. We then design the non-conformity score based on graph-level factuality uncertainty and calibrate the conformal threshold to decide when to stop generation. We theoretically show such generation is nested, yielding valid coverage guarantees for factuality control. Experiments over multiple datasets and coverage objectives demonstrate empirically valid coverage. In downstream reasoning tasks, inference-time calibrated graphs yield more accurate generation than post-hoc pruned graphs.

","absKo":"

Large language models (LLMs)는 점점 더 multi-step reasoning을 수행하고 있으며, 이때 중간 claim은 implicit directed acyclic graph를 형성하고 그 노드의 정답성은 조상 노드들에 의해 구조적으로 조건화된다. 이는 factuality uncertainty가 단순한 node-wise error의 누적이 아니라 구조적인 문제임을 의미하며, reasoning structure 전반에 대한 inference-time uncertainty quantification을 필요로 한다. Conformal prediction (CP)은 유연한 사용자 지정 factuality control을 제공하지만, 기존 연구는 post-hoc에 머물러 generation 도중 개입할 수 없다. CP의 유연성과 post-hoc 한계 사이의 간극을 메우기 위해, 우리는 CP를 reasoning graph generation에 직접 통합하는 Inference-Time Conformal Reasoning (ITCR) framework를 제안한다. ITCR은 복잡한 modeling assumption 없이 reasoning graph 전반의 claim-level factuality signal을 집계하는 structure-level factuality uncertainty function을 학습한다. 이어서 graph-level factuality uncertainty를 기반으로 non-conformity score를 설계하고, 생성 중지를 결정하기 위해 conformal threshold를 보정한다. 우리는 이러한 생성이 nested하다는 점을 이론적으로 보이며, factuality control에 대한 valid coverage guarantee를 도출한다. 여러 dataset과 coverage objective에 대한 실험은 경험적으로 유효한 coverage를 보여준다. downstream reasoning task에서는 inference-time으로 보정된 graph가 post-hoc으로 pruned된 graph보다 더 정확한 생성을 제공한다.

"},{"id":"60884","en":"Information Flow Reveals When to Trust Language Models","ko":"정보 흐름은 Language Model을 언제 신뢰할지 드러낸다","authors":"Rui Xu, Yi Chen, Jiujiu Chen, Sihong Xie","pos":"#2202","link":"https://openreview.net/forum?id=vd8HzoFZ7v","abs":"

In retrieval-augmented generation, language models can generate incorrect responses if they fail to utilize query-relevant content from the retrieved evidence. This shifts the focus of uncertainty quantification (UQ) toward assessing contextual grounding, i.e., whether predictions are supported by query-relevant tokens. Recent UQ methods unpack language models to characterize how inputs are processed. Nevertheless, these methods focus on a few layers and overlook the whole progressive propagation within the model, thereby failing to fully capture the grounding dynamics essential for reliable uncertainty estimation. We use information flow to build a layer-wise trace that reveals each context token’s contribution to the output, providing an interpretable basis for assessing reliability. From this analysis, we introduce two measures to calibrate prediction confidence. The first, \\textit{simulatability}, posits that a prediction is more likely to be correct when context token contributions align closely with their true relevance. The second, \\textit{concentration}, asserts that a response is more likely to be correct when it is derived from a narrow, focused subset of tokens. Experiments show that our method achieves an average AUROC of 0.70, exceeding the runner-up performance of 0.65, while maintaining moderate computational cost.

","absKo":"

retrieval-augmented generation에서 language model은 검색된 evidence로부터 query-relevant content를 활용하지 못하면 잘못된 응답을 생성할 수 있다. 이는 uncertainty quantification (UQ)의 초점을 contextual grounding, 즉 예측이 query-relevant token에 의해 뒷받침되는지 평가하는 것으로 이동시킨다. 최근의 UQ 방법들은 language model을 분해하여 입력이 어떻게 처리되는지 특성화한다. 그러나 이러한 방법들은 몇 개의 layer에만 초점을 맞추고 model 내부의 전체 progressive propagation을 간과하므로, 신뢰할 수 있는 uncertainty estimation에 필수적인 grounding dynamics를 충분히 포착하지 못한다. 우리는 information flow를 사용해 각 context token이 output에 기여하는 방식을 드러내는 layer-wise trace를 구성하고, 신뢰성을 평가하기 위한 해석 가능한 기반을 제공한다. 이 분석으로부터 우리는 prediction confidence를 보정하기 위한 두 가지 measure를 제안한다. 첫째, \\textit{simulatability}는 context token의 기여가 실제 관련성과 밀접하게 정렬될 때 예측이 더 정확할 가능성이 높다고 본다. 둘째, \\textit{concentration}은 응답이 token의 좁고 집중된 subset으로부터 도출될 때 더 정확할 가능성이 높다고 주장한다. 실험 결과, 우리의 방법은 평균 AUROC 0.70을 달성하여 runner-up의 0.65를 상회하면서도 중간 수준의 계산 비용을 유지한다.

"},{"id":"61539","en":"Learning syntax without semantics: Disentangled tiny language models","ko":"의미 없이 구문 학습하기: 분리된 Tiny Language Models","authors":"Ezra Winston, Zico Kolter","pos":"#2207","link":"https://openreview.net/forum?id=p7HVrmZwWB","abs":"

Language models acquire syntax and world knowledge together, entangling the two in ways that limit efficiency and controllability. We show that syntax can be learned while suppressing semantic plausibility and world‑knowledge cues, yielding more efficient and controllable models. We train tiny LMs on grammatical nonsense — syntactically well-formed text with semantic content ablated via constrained relexicalization (SAMBAL). Models trained on this data perform comparably to standard pretraining on syntactic benchmarks (BLiMP, SyntaxGym) while scoring at chance on world knowledge probes (EWoK). On targeted grammar-plausibility conflict probes, content-neutral models prefer grammaticality where standard models prefer plausibility, and their representations show more syntactic vs lexical alignment. On efficiency, disentanglement yields substantial sample and parameter gains: in low‑resource regimes, a 5M‑parameter model matches a 30M‑parameter baseline at the same data budget. On controllability, content-neutral models adapt rapidly to a new domain with minimal exposure, suggesting the feasibility of modular post‑hoc knowledge specialization.

","absKo":"

Language model은 syntax와 world knowledge를 함께 습득하며, 이 둘을 서로 얽히게 만들어 효율성과 controllability를 제한한다. 우리는 syntax가 semantic plausibility와 world-knowledge cue를 억제한 상태에서도 학습될 수 있음을 보이며, 이를 통해 더 효율적이고 더 통제 가능한 model을 얻을 수 있음을 보여준다. 우리는 tiny LM을 grammatical nonsense, 즉 constrained relexicalization(SAMBAL)으로 semantic content를 제거한 채 syntactically well-formed한 text에서 학습한다. 이 데이터로 학습한 model은 syntactic benchmark(BLiMP, SyntaxGym)에서 standard pretraining과 비슷한 성능을 내는 동시에 world knowledge probe(EWoK)에서는 chance 수준의 점수를 보인다. targeted grammar-plausibility conflict probe에서는 content-neutral model이 standard model이 plausibility를 선호하는 곳에서 grammaticality를 선호하며, 그 representation은 lexical alignment보다 syntactic alignment가 더 강하게 나타난다. 효율성 측면에서 disentanglement는 상당한 sample 및 parameter 이득을 가져온다. low-resource regime에서 5M parameter model은 같은 data budget에서 30M parameter baseline과 맞먹는 성능을 보인다. controllability 측면에서 content-neutral model은 최소한의 exposure만으로 새로운 domain에 빠르게 적응하며, 이는 모듈형 post-hoc knowledge specialization의 가능성을 시사한다.

"},{"id":"63666","en":"MAS-ProVe: Understanding the Process Verification of Multi-Agent Systems","ko":"MAS-ProVe: 다중 에이전트 시스템의 process verification 이해","authors":"Vishal Venkataramani, Haizhou Shi, Zixuan Ke, Austin Xu, Xiaoxiao He, Yingbo Zhou, Semih Yavuz, Hao Wang, Shafiq Joty","pos":"#2210","link":"https://openreview.net/forum?id=UoM3G7nKr0","abs":"

Multi-Agent Systems (MAS) built on Large Language Models (LLMs) often exhibit high variance in their reasoning trajectories. Process verification, which evaluates intermediate steps in trajectories, has shown promise in general reasoning settings, and has been suggested as a potential tool for guiding coordination of MAS; however, its actual effectiveness in MAS remains unclear. To fill this gap, we present MAS-ProVe, a systematic empirical study of process verification for multi-agent systems (MAS). Our study spans three verification paradigms (LLM-as-a-Judge, reward models, and process reward models), evaluated across two levels of verification granularity (agent-level and iteration-level). We further examine five representative verifiers and four context management strategies, and conduct experiments over six diverse MAS frameworks on multiple reasoning benchmarks. We find that process-level verification does not consistently improve performance and frequently exhibits high variance, highlighting the difficulty of reliably evaluating partial multi-agent trajectories. Among the methods studied, LLM-as-a-Judge generally outperforms reward-based approaches, with trained judges surpassing general-purpose LLMs. We further observe a small performance gap between LLMs acting as judges and as single agents, and identify a context-length-performance trade-off in verification. Overall, our results suggest that effective and robust process verification for MAS remains an open challenge, requiring further advances beyond current paradigms.

","absKo":"

Large Language Models(LLMs) 위에 구축된 Multi-Agent Systems(MAS)는 종종 reasoning trajectory에서 높은 분산을 보인다. trajectory의 중간 단계를 평가하는 process verification은 일반적인 reasoning 설정에서 유망함을 보여 왔고, MAS의 coordination을 유도하는 잠재적 도구로 제안되어 왔지만, MAS에서의 실제 효과는 여전히 불분명하다. 이 공백을 메우기 위해, 우리는 multi-agent system(MAS)에 대한 process verification의 체계적 실증 연구인 MAS-ProVe를 प्रस्तुत한다. 우리의 연구는 세 가지 verification paradigm(LLM-as-a-Judge, reward model, process reward model)을 포함하며, 두 수준의 verification granularity(agent-level 및 iteration-level)에서 평가한다. 또한 다섯 개의 대표적인 verifier네 가지 context management strategy를 살펴보고, 여러 reasoning benchmark에서 여섯 개의 다양한 MAS framework를 대상으로 실험을 수행한다. 우리는 process-level verification이 일관되게 성능을 향상시키지 못하며, 높은 분산을 자주 보인다는 점을 발견했는데, 이는 부분적인 multi-agent trajectory를 신뢰성 있게 평가하는 어려움을 드러낸다. 연구한 방법들 중에서 LLM-as-a-Judge가 일반적으로 reward-based 접근법보다 우수하며, 학습된 judge는 범용 LLM을 능가한다. 또한 judge로 작동하는 LLM과 single agent로 작동하는 LLM 사이의 성능 격차가 작다는 점을 관찰했으며, verification에서 context length와 성능 간 trade-off를 확인했다. 전반적으로, 우리의 결과는 MAS를 위한 효과적이고 robust한 process verification이 여전히 열린 과제이며, 현재 paradigm을 넘어서는 추가적인 발전이 필요함을 시사한다.

"},{"id":"65264","en":"Distinguishable Deletion: Unifying Knowledge Erasure and Refusal for Large Language Model Unlearning","ko":"구별 가능한 삭제: Large Language Model 언러닝을 위한 지식 소거와 거부의 통합","authors":"Puning Yang, Junchi Yu, Qizhou Wang, Phil Torr, Bo Han, Xiuying Chen","pos":"#2401","link":"https://openreview.net/forum?id=FFEKmtsT4S","abs":"Mitigating sensitive and harmful outputs is fundamental to ensuring safe deployment of LLMs. Existing approaches typically follow two paradigms: Knowledge Deletion (KD), which erases undesirable information during training, and Distinguishable Refusal (DR), which steers models away from using sensitive knowledge during inference. Despite rapid progress, KD-based unlearning struggles with biased deletion due to suppressing specific token sequences as a substitute for complete knowledge removal, whereas DR-based unlearning risks the re-emergence of harmful knowledge because the underlying knowledge remains intact. To address these issues, we propose Distinguishable Deletion ($\\mathrm{D^2}$), a paradigm that restricts the response distribution in the latent space rather than specific tokens to erase undesirable knowledge, while distinguishing it from retained knowledge, enabling a refusal mechanism to handle unlearned inputs safely and coherently. To implement $\\mathrm{D^2}$, we introduce an energy index that quantifies the presence of knowledge and the separation between unlearned and retained content. Mathematical and empirical analyses show that energy is both accurate and efficient, enabling Energy-based Unlearning Alignment (EUA) to enforce energy-boundary unlearning during training and apply an energy-based refusal mechanism at inference. Extensive experiments demonstrate that EUA significantly outperforms previous methods, indicating the superiority of $\\mathrm{D^2}$.","absKo":"민감하고 유해한 출력을 완화하는 것은 LLM을 안전하게 배포하기 위한 기본 과제다. 기존 접근은 보통 두 가지 패러다임을 따른다. 하나는 Knowledge Deletion (KD)으로, 학습 중 바람직하지 않은 정보를 지우는 방식이고, 다른 하나는 Distinguishable Refusal (DR)로, inference 시 민감한 지식을 사용하지 않도록 모델을 유도하는 방식이다. 빠른 진전에도 불구하고, KD 기반 unlearning은 완전한 knowledge removal의 대체물로 특정 token sequence를 억제하는 데 그치기 때문에 biased deletion 문제에 취약하고, DR 기반 unlearning은 underlying knowledge가 그대로 남아 있어 유해한 지식이 재등장할 위험이 있다. 이러한 문제를 해결하기 위해, 우리는 Distinguishable Deletion ($\\mathrm{D^2}$)을 제안한다. 이는 특정 token이 아니라 latent space에서 response distribution을 제한하여 바람직하지 않은 지식을 지우는 동시에, 이를 retained knowledge와 구별함으로써, unlearned input을 안전하고 일관되게 처리할 수 있는 refusal mechanism을 가능하게 하는 패러다임이다. $\\mathrm{D^2}$를 구현하기 위해, 우리는 지식의 존재와 unlearned content와 retained content 사이의 분리를 정량화하는 energy index를 도입한다. 수학적 및 실증적 분석은 energy가 정확하고 효율적임을 보여주며, 이를 통해 Energy-based Unlearning Alignment (EUA)가 학습 중 energy-boundary unlearning을 강제하고 inference 시 energy-based refusal mechanism을 적용할 수 있게 한다. 대규모 실험은 EUA가 기존 방법들을 크게 능가함을 보여주며, $\\mathrm{D^2}$의 우수성을 입증한다."},{"id":"66144","en":"Do Vision and Text Cues Exhibit Evidential Coupling? UFO: A Benchmark for Compositional Multimodal Reasoning in Unified Models","ko":"시각과 텍스트 단서는 증거적 결합을 보이는가? UFO: 통합 모델에서의 구성적 멀티모달 추론 벤치마크","authors":"Zhongyu Yang, Dannong Xu, Yonghan Zhang, Kefan Chen, Xinyi Wang, Yang Xu, Wei Pang, Yingfang Yuan","pos":"#2402","link":"https://openreview.net/forum?id=6UKaYYRM3h","abs":"

Unified Foundation Models (UFMs), which support interleaved multimodal generation and understanding, have been proposed as a promising paradigm for reasoning about dynamic world states, yet it remains unclear whether the visual content they generate functions as grounded evidence for subsequent reasoning or merely as auxiliary output. Existing benchmarks largely evaluate generation and understanding as separate capabilities and do not test their functional dependence during reasoning. We introduce \\textbf{UFO}, a benchmark designed to evaluate whether UFMs generate and use image and text cues as evidence for compositional multimodal reasoning. UFO spans three cue types, state determination, state reconstruction, and state augmentation, which correspond to progressively smaller transformations of the underlying world state. Our analysis reveals a significant modality gap, as models often achieve high prediction accuracy even when the generated visual cues exert limited influence on their decisions, indicating weakened evidential coupling and a reliance on textual shortcuts rather than robust cross modal grounding.

","absKo":"

Interleaved multimodal generation과 understanding을 지원하는 Unified Foundation Models(UFMs)은 dynamic world states에 대해 추론하는 유망한 패러다임으로 제안되어 왔지만, 이들이 생성하는 visual content가 이후 추론을 위한 grounded evidence로 기능하는지, 아니면 단지 auxiliary output에 불과한지는 여전히 불분명하다. 기존 benchmark들은 대체로 generation과 understanding을 분리된 능력으로 평가하며, reasoning 동안 두 능력의 functional dependence를 시험하지는 않는다. 우리는 compositional multimodal reasoning을 위해 UFMs가 image와 text cues를 evidence로 생성하고 사용하는지를 평가하도록 설계된 benchmark인 \\textbf{UFO}를 제안한다. UFO는 three cue types, state determination, state reconstruction, state augmentation을 포괄하며, 이는 underlying world state에 대한 점진적으로 더 작은 transformations에 대응한다. 우리의 분석은 significant modality gap을 드러내는데, 모델은 생성된 visual cues가 그들의 결정에 미치는 영향이 제한적일 때조차 높은 prediction accuracy를 달성하는 경우가 많으며, 이는 robust cross modal grounding보다는 weakened evidential coupling과 textual shortcuts에 대한 의존을 시사한다.

"},{"id":"64855","en":"Entropy-Aware On-Policy Distillation of Language Models","ko":"Language Model의 Entropy-Aware On-Policy Distillation","authors":"Woogyeol Jin, Taywon Min, Yongjin Yang, Swanand Kadhe, Yi Zhou, Dennis Wei, Nathalie Baracaldo, Kimin Lee","pos":"#2407","link":"https://openreview.net/forum?id=J5i09faOOf","abs":"On-policy distillation is a promising approach for transferring knowledge between language models, where a student learns from dense token-level signals along its own trajectories. This framework typically uses reverse KL divergence, encouraging the student to match the teacher's high-confidence predictions. However, we show that the mode-seeking property of reverse KL reduces generation diversity and yields unstable learning signals when the teacher distribution has high entropy. To address this, we introduce Entropy-Aware On-Policy Distillation. Our key idea is augmenting the standard reverse KL objective with forward KL when teacher entropy is high, capturing the full range of plausible outputs while retaining precise imitation elsewhere. It balances mode-seeking precision with mode-covering robustness without sacrificing on-policy training efficiency. Experiments show that our method maintains generation diversity (sustained token-level entropy) and improves student–teacher alignment (lower forward KL on high-entropy tokens). Across six math reasoning benchmarks, this yields Pass@8 accuracy gains of $+1.37$ for Qwen3-0.6B-Base, $+2.39$ for Qwen3-1.7B-Base, and $+5.05$ for Qwen3-4B-Base compared to baseline on-policy distillation methods. These results demonstrate that accounting for teacher uncertainty is essential for maintaining diversity and achieving effective knowledge transfer.","absKo":"On-policy distillation은 language model 간 지식 이전을 위한 유망한 접근으로, student가 자신의 trajectory를 따라 dense token-level signal로부터 학습한다. 이 framework는 일반적으로 reverse KL divergence를 사용하여 student가 teacher의 높은 confidence prediction을 맞추도록 유도한다. 그러나 우리는 reverse KL의 mode-seeking 성질이 generation diversity를 감소시키고, teacher distribution의 entropy가 높을 때 불안정한 learning signal을 초래함을 보인다. 이를 해결하기 위해 우리는 Entropy-Aware On-Policy Distillation을 도입한다. 핵심 아이디어는 teacher entropy가 높을 때 표준 reverse KL objective에 forward KL을 보강하여, 다른 부분에서는 정밀한 imitation을 유지하면서 plausible output의 전체 범위를 포착하는 것이다. 이는 on-policy training efficiency를 희생하지 않으면서 mode-seeking의 정밀성과 mode-covering의 견고성을 균형 있게 결합한다. 실험 결과, 우리의 방법은 generation diversity(token-level entropy 유지)를 보존하고 student-teacher alignment(high-entropy token에서 더 낮은 forward KL)를 개선한다. 여섯 개의 math reasoning benchmark 전반에서, 이는 baseline on-policy distillation 방법과 비교하여 Qwen3-0.6B-Base에서 Pass@8 정확도 +1.37, Qwen3-1.7B-Base에서 +2.39, Qwen3-4B-Base에서 +5.05의 향상을 가져온다. 이러한 결과는 teacher uncertainty를 고려하는 것이 다양성을 유지하고 효과적인 지식 이전을 달성하는 데 필수적임을 보여준다."},{"id":"65508","en":"Factored Causal Representation Learning for Robust Reward Modeling in RLHF","ko":"RLHF에서 강건한 보상 모델링을 위한 인수분해 인과 표현 학습","authors":"Yupei Yang, Lin Yang, Wanxi Deng, Lin Qu, Fan Feng, Biwei Huang, Shikui Tu, Lei Xu","pos":"#2413","link":"https://openreview.net/forum?id=CwNItJ07ew","abs":"

A reliable reward model is essential for aligning large language models (LLMs) with human preferences through reinforcement learning from human feedback (RLHF). However, standard reward models are susceptible to spurious features that are not causally related to human labels. This can lead to reward hacking, where high predicted reward does not translate into better behavior. In this work, we address this problem from a causal perspective by proposing a factored representation learning framework that decomposes the model’s contextual embedding into (1) causal factors that are sufficient for reward prediction and (2) non-causal factors that capture reward-irrelevant attributes such as length or sycophantic bias. The reward head is then constrained to depend only on the causal component. In addition, we introduce an adversarial head trained to predict reward from the non-causal factors, while applying gradient reversal to discourage them from encoding reward-relevant information. Experiments on both mathematical and dialogue tasks demonstrate that our method learns more robust reward models and consistently improves downstream RLHF performance over state-of-the-art baselines. Analyses on length and sycophantic bias further validate the effectiveness of our method in mitigating reward hacking behaviors.

","absKo":"

신뢰할 수 있는 reward model은 reinforcement learning from human feedback (RLHF)를 통해 large language model (LLM)을 human preference에 정렬하는 데 필수적이다. 그러나 표준 reward model은 human label과 인과적으로 관련되지 않은 spurious feature에 취약하다. 이는 예측된 reward가 높아도 더 나은 행동으로 이어지지 않는 reward hacking으로 이어질 수 있다. 본 연구에서는 모델의 contextual embedding을 (1) reward 예측에 충분한 causal factor와 (2) length나 sycophantic bias와 같이 reward와 무관한 속성을 포착하는 non-causal factor로 분해하는 factored representation learning framework를 제안함으로써 이 문제를 causal perspective에서 다룬다. 이어서 reward head는 오직 causal component에만 의존하도록 제약된다. 추가로, non-causal factor로부터 reward를 예측하도록 학습되는 adversarial head를 도입하고, gradient reversal을 적용하여 이들이 reward 관련 정보를 인코딩하지 못하도록 억제한다. 수학 및 대화 task 모두에서의 실험은 우리의 방법이 더 견고한 reward model을 학습하고, 최신 state-of-the-art baseline보다 downstream RLHF 성능을 일관되게 향상시킨다는 것을 보여준다. length와 sycophantic bias에 대한 분석은 reward hacking 행동을 완화하는 데 있어 우리의 방법의 효과를 추가로 검증한다.

"},{"id":"63887","en":"Anchor-Final Self-Supervision Drives Hallucination-Aware Optimization in Large Vision-Language Models","ko":"Anchor-Final Self-Supervision이 대규모 비전-언어 model의 hallucination 인지 최적화를 이끈다","authors":"Jiaxi Liu, Yifeng Yang, Xinbing Wang, Qinying Gu, Nanyang Ye","pos":"#2600","link":"https://openreview.net/forum?id=Smm2KRLqx5","abs":"

Hallucinations in large vision-language models (LVLMs) remain a critical challenge, with models often generate tokens that fail to align with visual evidence. To address this issue, we propose AFS: Anchor-Final Self-Supervision, a novel framework for hallucination-aware optimization in LVLMs. By leveraging discrepancies between intermediate and final layer predictions, AFS selectively applies self-supervision to visually descriptive tokens, incorporates hallucination-aware token classification, and encourages consistency between intermediate and final layer distributions. Unlike traditional methods that rely on explicit supervision or post-hoc interventions, AFS optimizes the model via Group Relative Policy Optimization (GRPO), using token-specific rewards derived solely from internal model signals. Experiments demonstrate that AFS significantly reduces hallucinations without compromising recall in caption generation. Beyond captioning, AFS excels in discriminative tasks, improving the reliability of object existence predictions and multimodal reasoning. Furthermore, AFS demonstrates strong cross-dataset generalization, transferring effectively across diverse visual domains.

","absKo":"

large vision-language model (LVLM)에서의 hallucination은 여전히 중요한 과제로 남아 있으며, 모델은 종종 visual evidence와 맞지 않는 token을 생성한다. 이 문제를 해결하기 위해, 우리는 LVLM을 위한 hallucination-aware optimization의 새로운 framework인 AFS: Anchor-Final Self-Supervision을 제안한다. AFS는 중간 layer와 최종 layer 예측 간의 차이를 활용하여, visually descriptive token에 선택적으로 self-supervision을 적용하고, hallucination-aware token classification을 통합하며, 중간 layer와 최종 layer 분포 간의 일관성을 강화한다. 명시적 supervision이나 post-hoc intervention에 의존하는 전통적 방법과 달리, AFS는 Group Relative Policy Optimization (GRPO)을 사용해 model 내부 신호에서만 도출된 token-specific reward로 모델을 최적화한다. 실험 결과 AFS는 caption generation에서 recall을 저해하지 않으면서 hallucination을 크게 줄인다. captioning을 넘어, AFS는 discriminative task에서도 뛰어난 성능을 보이며 object existence prediction과 multimodal reasoning의 신뢰성을 향상시킨다. 더 나아가 AFS는 다양한 visual domain을 가로지르는 강한 cross-dataset generalization을 보여주며 효과적으로 transfer된다.

"},{"id":"65588","en":"ABCD: All Biases Come Disguised","ko":"ABCD: 모든 편향은 위장하여 온다","authors":"Mateusz Nowak, Xavier Cadet, Peter Chin","pos":"#2703","link":"https://openreview.net/forum?id=C7reVNwYsF","abs":"Multiple-choice question (MCQ) benchmarks have been a standard evaluation practice for measuring LLMs' ability to reason and answer knowledge-based questions. Through a synthetic NonsenseQA benchmark, we observe that different LLMs exhibit varying degrees of label-position-few-shot-prompt bias, where the model either uses the answer position, the label in front of the answer, the distributions of correct answers present in the few-shot prompt, or a combination of all to answer each MCQ question. We propose a simple bias-reduced evaluation protocol that replaces the labels of each question with uniform, unordered labels and prompts the LLM to use the whole answer presented. With a simple sentence similarity model, we demonstrate improved robustness and lower standard deviation between different permutations of answers with a minimal drop in LLM's performance, exposing the LLM's capabilities under reduced evaluation artifacts, without any help from the prompt examples or the option labels. Across multiple benchmarks and models, this protocol substantially improves the robustness to answer permutations, reducing mean accuracy variance $3\\times$ with only a minimal decrease in the mean model's performance. Through ablation studies on various embedding models and similarity functions, we show that the method is more robust than the standard ones.","absKo":"다중선택형 질문(MCQ) benchmark는 LLM이 추론하고 지식 기반 질문에 답하는 능력을 측정하기 위한 표준 평가 관행이었다. 합성 NonsenseQA benchmark를 통해, 서로 다른 LLM이 label-position-few-shot-prompt bias를 서로 다른 정도로 보인다는 것을 관찰했다. 여기서 모델은 답의 위치, 답 앞의 label, few-shot prompt에 나타난 정답 분포, 또는 이들의 조합을 사용해 각 MCQ 질문에 답한다. 우리는 각 질문의 label을 균일하고 순서가 없는 label로 바꾸고, LLM이 제시된 전체 답을 사용하도록 유도하는 간단한 bias-reduced 평가 프로토콜을 제안한다. 간단한 sentence similarity model과 함께, 우리는 LLM 성능의 최소한의 저하만으로 서로 다른 답 순열 간의 분산을 줄이고 강건성을 향상시킴을 보인다. 이는 prompt 예시나 option label의 도움 없이, 평가 인공물(evaluation artifacts)을 줄인 상태에서 LLM의 역량을 드러낸다. 여러 benchmark와 model 전반에서, 이 프로토콜은 답 순열에 대한 강건성을 크게 향상시켜 mean accuracy variance를 $3\\times$ 줄이면서 mean model 성능은 최소한만 감소시킨다. 다양한 embedding model과 similarity function에 대한 ablation study를 통해, 이 방법이 표준 방법들보다 더 강건함을 보인다."},{"id":"62133","en":"A Narrowing Geometry in Contaminated Reasoning","ko":"오염된 추론에서의 좁아지는 기하","authors":"Jiakuan Xie, Pengfei Cao, Kang Liu, Jun Zhao","pos":"#2704","link":"https://openreview.net/forum?id=jXhXr53hdG","abs":"

Despite the advancing reasoning capabilities of large language models (LLMs), many reasoning evaluations are increasingly compromised by data contamination, which induces unreliable contaminated reasoning on leaked inputs. While this phenomenon is widely observed, its underlying mechanism remains poorly understood, hindering the ability to distinguish generalization from memorization and to develop effective solutions. In this work, we first identify a distinctive signal of contaminated reasoning, namely the mutual information decay between representations and gradients. Our mechanistic analysis reveals that contaminated models exhibit pronounced eigenspectrum concentration in their representations, leading to a low-dimensional computation regime. Under leaked inputs, this mechanism weakens the linear coupling between representations and gradients, manifested as a structural decay of the singular values in the whitened space. We show that this narrowing geometry mathematically implies a reduction in mutual information, and further demonstrate the practical utility of our analysis by successfully restoring the reasoning behavior of contaminated models, achieving an 11.03% improvement in average consistency with the base model over the strongest baseline.

","absKo":"

large language model (LLM)의 reasoning 능력이 향상되고 있음에도 불구하고, 많은 reasoning evaluation은 data contamination에 의해 점점 더 손상되고 있으며, 이는 유출된 입력에서 신뢰할 수 없는 contaminated reasoning을 유발합니다. 이 현상은 널리 관찰되지만, 그 근본 메커니즘은 여전히 충분히 이해되지 않아 generalization과 memorization을 구분하고 효과적인 해결책을 개발하기 어렵게 만듭니다. 본 연구에서는 먼저 contaminated reasoning의 특징적 신호로서 representation과 gradient 사이의 mutual information decay를 식별합니다. 우리의 mechanistic analysis는 contaminated model이 representation에서 두드러진 eigenspectrum concentration을 보이며, 이로 인해 low-dimensional computation regime에 놓인다는 점을 드러냅니다. 유출된 입력 아래에서는 이 메커니즘이 representation과 gradient 사이의 linear coupling을 약화시키며, 이는 whitened space에서 singular value의 structural decay로 나타납니다. 우리는 이러한 좁아진 geometry가 수학적으로 mutual information의 감소를 의미함을 보이고, 더 나아가 contaminated model의 reasoning behavior를 성공적으로 복원함으로써 우리의 분석의 실용적 유용성을 입증합니다. 그 결과, 가장 강력한 baseline 대비 base model과의 평균 consistency에서 11.03% 향상을 달성했습니다.

"},{"id":"61379","en":"MemEvolve: Meta-Evolution of Agent Memory Systems","ko":"MemEvolve: Agent Memory Systems의 Meta-Evolution","authors":"Guibin Zhang, Haotian Ren, Chong Zhan, Junhao Wang, He Zhu, Wangchunshu Zhou, Shuicheng YAN","pos":"#300","link":"https://openreview.net/forum?id=qpkG0eKx4v","abs":"Self-evolving memory systems are rapidly reshaping the evolutionary paradigm of large language model (LLM)-based agents. Prior work has predominantly relied on manually engineered memory architectures to store trajectories, distill experience, and synthesize reusable tools, enabling agents to evolve on the fly within environment interactions. However, this paradigm is fundamentally constrained by the \\textit{staticity} of the memory system itself: while memory facilitates agent-level evolving, the underlying memory architecture cannot be meta-adapted to diverse task contexts. To address this gap, we propose MemEvolve, a meta-evolutionary framework that jointly evolves agents’ experiential knowledge and their memory architecture, allowing agent systems not only to accumulate experience but also to progressively refine how they learn from it. To ground MemEvolve in prior work and promote openness in future self-evolving systems, we introduce EvolveLab, a unified memory codebase that distills twelve representative memory systems into a modular design space (\\textit{encode}, \\textit{store}, \\textit{retrieve}, \\textit{manage}), providing a standardized implementation substrate and a fair experimental arena. Extensive evaluations on four challenging agentic benchmarks show that MemEvolve delivers (i) substantial performance gains, improving frameworks such as SmolAgent and Flash-Searcher by up to $17.06\\%$, and (ii) strong cross-task and cross-LLM generalization, yielding memory architectures that transfer effectively across diverse benchmarks and backbones.","absKo":"자기 진화하는 memory system은 대규모 language model(LLM) 기반 agent의 진화 패러다임을 빠르게 재편하고 있다. 기존 연구는 주로 trajectory를 저장하고, 경험을 증류하며, 재사용 가능한 tool을 합성하기 위해 수작업으로 설계된 memory architecture에 의존해 왔고, 이를 통해 agent가 environment interaction 내부에서 즉시 진화할 수 있게 했다. 그러나 이러한 패러다임은 memory system 자체의 \\textit{staticity}에 근본적으로 제약된다. 즉, memory는 agent 수준의 evolving을 가능하게 하지만, underlying memory architecture는 다양한 task context에 맞게 meta-adapt될 수 없다. 이 간극을 해결하기 위해 우리는 agent의 experiential knowledge와 memory architecture를 공동으로 진화시키는 meta-evolutionary framework인 MemEvolve를 제안하며, 이를 통해 agent system이 경험을 축적하는 것에 더해 그 경험으로부터 학습하는 방식을 점진적으로 정교화할 수 있도록 한다. MemEvolve를 기존 연구에 기반하여 정립하고 향후 self-evolving system의 개방성을 높이기 위해, 우리는 12개의 대표적 memory system을 \\textit{encode}, \\textit{store}, \\textit{retrieve}, \\textit{manage}라는 모듈형 design space로 정제한 통합 memory codebase인 EvolveLab을 도입하여, 표준화된 구현 기반과 공정한 실험 장을 제공한다. 네 개의 도전적인 agentic benchmark에 대한 광범위한 평가는 MemEvolve가 (i) SmolAgent와 Flash-Searcher 같은 framework의 성능을 최대 $17.06\\%$까지 향상시키는 상당한 성능 향상과, (ii) 다양한 benchmark와 backbone 전반에 걸쳐 효과적으로 이전되는 memory architecture를 산출하는 강한 cross-task 및 cross-LLM generalization을 제공함을 보여준다."},{"id":"61298","en":"Real-Time Monitoring and Calibration of Chain-of-Thought Sycophancy in Large Reasoning Models","ko":"Large Reasoning Model의 Chain-of-Thought 아첨성 실시간 모니터링과 보정","authors":"Jingyu Hu, Shu Yang, Xilin Gong, Hongming Wang, Weiru Liu, Di Wang","pos":"#3114","link":"https://openreview.net/forum?id=rbJsrAxryz","abs":"

Large Reasoning Models (LRMs) suffer from sycophantic behavior, where models tend to agree with users' incorrect beliefs and follow misinformation rather than maintain independent reasoning. This behavior undermines model reliability and poses societal risks. Mitigating LRM sycophancy requires monitoring how this sycophancy emerges during the reasoning trajectory; however, current methods mainly focus on judging based on final answers and correcting them, without understanding how sycophancy develops during reasoning processes. To address this limitation, we propose MONICA, a novel Monitor-guided Calibration framework that monitors and mitigates sycophancy during model inference at the level of reasoning steps, without requiring the model to finish generating its complete answer. MONICA integrates a sycophantic monitor that provides real-time monitoring of sycophantic drift scores during response generation with a calibrator that dynamically suppresses sycophantic behavior when scores exceed predefined thresholds. Extensive experiments across 12 datasets and 3 LRMs demonstrate that our method effectively reduces sycophantic behavior in both intermediate reasoning steps and final answers, yielding robust performance improvements.

","absKo":"

Large Reasoning Model (LRM)은 sycophantic behavior를 보이는데, 이는 모델이 사용자의 잘못된 믿음에 동의하고 독립적인 reasoning을 유지하기보다 misinformation을 따르는 경향을 뜻합니다. 이러한 행동은 모델의 신뢰성을 저해하고 사회적 위험을 초래합니다. LRM의 sycophancy를 완화하려면 reasoning trajectory 동안 이러한 sycophancy가 어떻게 나타나는지 모니터링해야 하지만, 현재의 방법들은 주로 최종 답변을 기준으로 판단해 이를 교정하는 데 집중하며, reasoning 과정에서 sycophancy가 어떻게 전개되는지는 이해하지 못합니다. 이 한계를 해결하기 위해, 우리는 MONICA를 제안합니다. MONICA는 Monitor-guided Calibration framework로서, 모델이 완전한 답변 생성을 마치지 않아도 reasoning step 수준에서 inference 중 sycophancy를 모니터링하고 완화합니다. MONICA는 응답 생성 동안 sycophantic drift score를 실시간으로 모니터링하는 sycophantic monitor와, score가 미리 정의된 threshold를 초과할 때 sycophantic behavior를 동적으로 억제하는 calibrator를 통합합니다. 12개 데이터셋과 3개의 LRM에 걸친 광범위한 실험은, 우리의 방법이 중간 reasoning step과 최종 답변 모두에서 sycophantic behavior를 효과적으로 줄이며 견고한 성능 향상을 제공함을 보여줍니다.

"},{"id":"62574","en":"InfoLaw: Information Scaling Laws for Large Language Models with Quality-Weighted Mixture Data and Repetition","ko":"InfoLaw: Quality-Weighted Mixture Data와 Repetition을 고려한 LLM의 Information Scaling Laws","authors":"Fengze Liu, Weidong Zhou, LIU, Ping Guo, Zijun Wang, Bingni Zhang, Yifan Zhang, Yifeng Yu, Xiaohuan ZHOU, Taifeng Wang","pos":"#315","link":"https://openreview.net/forum?id=fQaVptMRCY","abs":"

Upweighting high-quality data in LLM pretraining often improves performance, but in data-limited regimes, especially under overtraining, stronger upweighting increases repetition and can degrade performance. However, standard scaling laws do not reliably extrapolate across mixture recipes or under repetitions, making the selection for optimal data recipes at scaling underdetermined. To solve this, we introduce \\textbf{InfoLaw} (Information Scaling Laws), a data-aware scaling framework that predicts loss from consumed tokens, model size, data mixture weights, and repetition. The key idea is to model pretraining as information accumulation, where quality controls information density and repetition induces scale-dependent diminishing returns. We first collect the model performance after training on datasets that vary in scale, quality distribution, and repetition level. Then we build up the modeling for information so that information accurately predicts those model performance. InfoLaw predicts performance on unseen data recipes and larger-scale runs (up to 7B, 425B tokens) with 0.15\\% mean and 0.96\\% max absolute error in loss, and it extrapolates reliably across overtraining levels, enabling efficient data-recipe selection under varying compute budgets.

","absKo":"

LLM pretraining에서 고품질 data의 가중치를 높이면 종종 성능이 향상되지만, data-limited regime, 특히 overtraining 하에서는 더 강한 upweighting이 repetition을 증가시키고 성능을 저하시킬 수 있다. 그러나 표준 scaling law는 mixture recipe 간이나 repetition이 존재하는 경우에 신뢰성 있게 extrapolate하지 못하므로, 스케일링에서 최적 data recipe를 선택하는 문제가 미정 상태로 남는다. 이를 해결하기 위해 우리는 소비된 token, model size, data mixture weight, repetition으로부터 loss를 예측하는 data-aware scaling framework인 \\textbf{InfoLaw} (Information Scaling Laws)를 도입한다. 핵심 아이디어는 pretraining을 information accumulation으로 모델링하는 것으로, quality는 information density를 제어하고 repetition은 scale-dependent diminishing return을 유발한다. 우리는 먼저 규모, quality distribution, repetition level이 다른 dataset들로 학습한 뒤의 model performance를 수집한다. 그런 다음 이러한 model performance를 정확히 예측하도록 information 모델링을 구축한다. InfoLaw는 보이지 않은 data recipe와 더 큰 규모의 run(최대 7B, 425B tokens)에서 loss에 대해 평균 0.15\\%, 최대 0.96\\%의 절대 오차로 성능을 예측하며, overtraining 수준 전반에 걸쳐 신뢰성 있게 extrapolate하여 다양한 compute budget 하에서 효율적인 data-recipe 선택을 가능하게 한다.

"},{"id":"60606","en":"Mitigating Translationese Bias in Multilingual LLM-as-a-Judge via Disentangled Information Bottleneck","ko":"분리된 정보 병목을 통한 다국어 LLM-as-a-Judge의 번역투 편향 완화","authors":"Hongbin Zhang, Kehai Chen, Xuefeng Bai, Youcheng Pan, Yang Xiang, Jinpeng Wang, Min zhang","pos":"#108","link":"https://openreview.net/forum?id=yW5LkzZN0T","abs":"

Large language models (LLMs) have emerged as a standard paradigm for automated multilingual evaluation, yet exhibit systematic biases. In this paper, we identify ``translationese bias'', in which LLMs systematically favor machine-translated text over human-authored references, and this bias is particularly pronounced in low-resource languages. We attribute this bias to spurious correlations with (a) strong latent manifold isomorphism with English and (b) high predictive confidence. To mitigate these issues, we present DIBJudge, a robust fine-tuning framework that decouples robust features and bias representations by explicitly isolating spurious attributes into a dedicated bias branch and penalizing mutual dependence to enforce disentanglement. In particular, we present a vector-quantized compression that ensures the robust representation retains minimal and sufficient judgment-critical information. Extensive evaluations on multilingual reward modeling benchmarks and a specially designed translationese bias evaluation suite demonstrate that DIBJudge outperforms strong baselines and effectively mitigates translationese bias.

","absKo":"

대규모 언어 모델(LLM)은 자동화된 다국어 평가의 표준 패러다임으로 부상했지만, 체계적인 편향을 보인다. 본 논문에서는 LLM이 기계 번역된 텍스트를 인간이 작성한 reference보다 체계적으로 선호하는 ``translationese bias''를 식별하며, 이러한 편향은 특히 low-resource language에서 두드러진다는 점을 보인다. 우리는 이 편향이 (a) English와의 강한 latent manifold isomorphism과 (b) 높은 predictive confidence와의 허위 상관관계에서 비롯된다고 설명한다. 이러한 문제를 완화하기 위해, 우리는 허위 속성을 전용 bias branch에 명시적으로 격리하고 mutual dependence를 페널티로 주어 disentanglement를 강제함으로써 robust feature와 bias representation을 분리하는 강건한 fine-tuning framework인 DIBJudge를 제안한다. 특히, 우리는 judgment에 필수적인 정보를 최소한이면서도 충분하게 유지하도록 보장하는 vector-quantized compression을 제시한다. 다국어 reward modeling benchmark와 translationese bias 평가를 위해 특별히 설계된 suite에 대한 광범위한 실험은 DIBJudge가 강력한 baseline을 능가하고 translationese bias를 효과적으로 완화함을 보여준다.

"},{"id":"62768","en":"When Labelers Stay Silent: The Power of Ties in Cost-Effective Preference Learning","ko":"라벨러가 침묵할 때: 비용 효율적 Preference Learning에서 동점(Tie)의 힘","authors":"JIAQI LYU, Zihan Zhang, CJ Y, Shiyu Xia, Ning Xu, Xin Geng","pos":"#1706","link":"https://openreview.net/forum?id=dXY80mjA7D","abs":"

Standard preference alignment relies on a binary forced-choice paradigm, assuming definitive preferences for all pairs. However, we find that indistinguishable pairs are prevalent even in standard benchmarks, where quality differences of two responses often fall below the labeler's discriminative resolution limit. Forcing a choice in such cases could inject significant noise that undermines policy optimization. In this work, we propose a silent-aware framework that introduces a principled way to allow annotators to stay silent (i.e., express ties) and then explicitly model these ties during optimization. Our findings reveal a compelling phenomenon: when ties are properly modeled, supervision from small models yields alignment surpassing that of forced-choice LLMs or human experts. This discovery highlights a cost-effective path for alignment: respecting a labeler’s resolution limit is more critical than increasing its capability, while simultaneously unlocking the latent value in existing benchmarks by properly modeling inherent tie signals without requiring any re-labeling effort. To leverage these signals, we propose several optimization objectives to drive the policy toward high-reward regions while mitigating unreliable updates that lead to arbitrary distribution shifts. Our approaches significantly surpass conventional alignment performance, consistently outperforming the strongest available baselines across diverse benchmarks.

","absKo":"

표준 preference alignment는 모든 쌍에 대해 명확한 선호가 존재한다고 가정하는 binary forced-choice paradigm에 의존한다. 그러나 우리는 표준 benchmark에서도 구분 불가능한 쌍이 널리 존재함을 발견했는데, 두 response의 품질 차이가 종종 labeler의 discriminative resolution limit보다 낮다. 이러한 경우에 선택을 강제하면 policy optimization을 훼손하는 상당한 noise가 유입될 수 있다. 본 연구에서는 annotator가 침묵할 수 있도록, 즉 tie를 표현할 수 있도록 허용하는 원칙적인 방법을 도입하고, optimization 과정에서 이러한 tie를 명시적으로 모델링하는 silent-aware framework를 제안한다. 우리의 결과는 인상적인 현상을 보여준다. tie를 제대로 모델링하면, 작은 model로부터의 supervision이 forced-choice LLM이나 human expert의 supervision을 능가하는 alignment를 달성한다. 이 발견은 비용 효율적인 alignment 경로를 시사한다. 즉, labeler의 resolution limit을 존중하는 것이 그 능력을 높이는 것보다 더 중요하며, 동시에 재라벨링 없이도 inherent tie signal을 적절히 모델링함으로써 기존 benchmark에 숨어 있는 가치를 활용할 수 있다. 이러한 signal을 활용하기 위해, 우리는 policy를 high-reward region으로 유도하면서 임의의 distribution shift를 초래하는 불안정한 update를 완화하는 여러 optimization objective를 제안한다. 우리의 방법은 다양한 benchmark에서 강력한 baseline을 일관되게 능가하며 conventional alignment 성능을 크게 상회한다.

"},{"id":"66160","en":"Tailoring the Training: Difficulty-Aware Learning Strategy Allocation for Large Language Models","ko":"훈련의 맞춤화: 대규모 언어 모델을 위한 난이도 인지 학습 전략 할당","authors":"Xiaoling Zhou, Shuaiyu Zhou, Zhemg Lee, Tao Chen, Xirui Li, Peng Chen, Jie Jiang, Wei Ye, Shikun Zhang","pos":"#1805","link":"https://openreview.net/forum?id=6EX4e6kSd0","abs":"

Although reinforcement learning (RL) enhances the reasoning capabilities of large language models (LLMs), it is primarily learned from the model's self-generated distribution, limiting its ability to acquire reasoning skills beyond its initial knowledge. To overcome this, we propose a Difficulty-Aware Learning Strategy Allocation (DALSA) framework, which adaptively assigns appropriate learning strategies to samples based on their difficulty signals. DALSA is built on the key insight that samples beyond models' knowledge scope are better addressed through supervised fine-tuning (SFT), while those within the boundary but insufficiently mastered benefit more from RL, and well-learned samples are discarded to avoid redundant updates. To realize this principle, we extract a series of difficulty-aware training characteristics and employ a learnable strategy allocator to dynamically determine the optimal learning strategy for each sample based on its training dynamics. The allocator and the LLM are alternately optimized, enabling adaptive strategy allocation. Furthermore, two regularization techniques, anti-curriculum weighting and adversarial label smoothing, are integrated to alleviate the inherent limitations of RL and SFT, backed by comprehensive theoretical analyses. Extensive experiments on ten LLMs ranging from 1.5B to 70B across various tasks indicate that DALSA consistently outperforms baselines under both full and parameter-efficient fine-tuning settings.

","absKo":"

reinforcement learning (RL)은 대규모 language model (LLM)의 reasoning 능력을 향상시키지만, 주로 model이 스스로 생성한 distribution으로부터 학습되므로 초기 지식 범위를 넘어서는 reasoning skill을 습득하는 데 한계가 있다. 이를 극복하기 위해 우리는 Difficulty-Aware Learning Strategy Allocation (DALSA) framework를 제안한다. DALSA는 sample의 difficulty signal에 따라 적절한 learning strategy를 적응적으로 할당한다. DALSA의 핵심 통찰은 model의 지식 범위를 벗어난 sample은 supervised fine-tuning (SFT)으로 다루는 것이 더 적합하고, 경계 안에 있지만 충분히 숙련되지 않은 sample은 RL의 이점을 더 크게 보며, 이미 잘 학습된 sample은 중복 갱신을 피하기 위해 제거해야 한다는 점이다. 이 원리를 구현하기 위해 우리는 일련의 difficulty-aware training characteristic을 추출하고, 학습 가능한 strategy allocator를 사용하여 training dynamics를 기반으로 각 sample의 최적 learning strategy를 동적으로 결정한다. allocator와 LLM은 교대로 최적화되어 적응적 strategy allocation이 가능해진다. 또한 anti-curriculum weighting과 adversarial label smoothing이라는 두 가지 regularization 기법을 통합하여 RL과 SFT의 내재적 한계를 완화하며, 이에 대한 포괄적인 이론적 분석도 제시한다. 1.5B에서 70B에 이르는 10개의 LLM과 다양한 task에 대한 광범위한 실험은 DALSA가 full fine-tuning과 parameter-efficient fine-tuning 설정 모두에서 일관되게 baseline을 능가함을 보여준다.

"},{"id":"64888","en":"Target-Oriented Pretraining Data Selection via Neuron-Activated Graph","ko":"Neuron-Activated Graph를 통한 표적 사전학습 데이터 선택","authors":"Zijun Wang, Haoqin Tu, Weidong Zhou, Yiyang Zhou, Xiaohuan ZHOU, Bingni Zhang, Weiguo Feng, Taifeng Wang, Cihang Xie, Fengze Liu","pos":"#1806","link":"https://openreview.net/forum?id=Ioj1rlc9f7","abs":"

Everyday tasks come with a target, and pretraining models around this target is what turns them into experts. In this paper, we study target-oriented language model (LM) pretraining by introducing ***N**euron-Activated Graph Ranking* (NAG-based Ranking), a training-free and interpretable framework for target pretraining data selection. Rather than using black-box representations, our approach directly characterizes each target input by a sparse set of high-impact neurons in any off-the-shelf LLMs. Concretely, we quantify neuron impact and select the most influential neurons across layers into a compact Neuron-Activated Graph (NAG), and rank candidate data by NAG similarity to target examples. We conduct experiments across six benchmarks, where our NAG-based Ranking improves target-oriented pretraining by 4.9\\% on average over random sampling, and also outperforms state-of-the-art baselines by 5.3\\% accuracy on HellaSwag. It also remains effective under a more applicable multi-target setting, where our best setup surpasses two baselines by 1.1\\% and 4.1\\%, respectively. Furthermore, we provide a comprehensive analysis on why and how our NAG works, e.g., deactivating NAG-selected neurons (only 0.12\\% of all) causes a 23.5\\% performance collapse, and restricting NAG to the final layer incurs a 4.1\\% average drop, indicating that NAG captures a sparse ``functional backbone'' for learning target features.

","absKo":"

일상적인 작업에는 목표가 있으며, 이 목표를 중심으로 model pretraining을 수행해야 비로소 전문가로 변한다. 본 논문에서는 target-oriented language model (LM) pretraining을 연구하기 위해 ***N**euron-Activated Graph Ranking* (NAG-based Ranking)을 소개한다. 이는 target pretraining data selection을 위한 training-free이면서 interpretable한 framework이다. Black-box representation을 사용하는 대신, 우리의 접근법은 off-the-shelf LLMs에서 높은 영향력을 가진 sparse neuron 집합으로 각 target input을 직접 특성화한다. 구체적으로, neuron impact를 정량화하고 layer 전반에서 가장 영향력 있는 neuron들을 선택해 compact한 Neuron-Activated Graph (NAG)을 구성한 뒤, NAG similarity를 기준으로 후보 데이터를 target example과 비교해 순위를 매긴다. 우리는 여섯 개 benchmark에서 실험을 수행했으며, NAG-based Ranking은 random sampling 대비 target-oriented pretraining을 평균 4.9\\% 향상시키고, HellaSwag에서는 state-of-the-art baseline보다 5.3\\% 정확도를 더 높였다. 또한 더 실용적인 multi-target setting에서도 효과가 유지되어, 우리의 최적 구성은 두 baseline보다 각각 1.1\\%와 4.1\\% 더 높은 성능을 보였다. 더 나아가, 왜 그리고 어떻게 NAG가 작동하는지에 대한 포괄적인 분석을 제공한다. 예를 들어 NAG가 선택한 neuron들(전체의 0.12\\%에 불과)을 비활성화하면 성능이 23.5\\% 붕괴하고, NAG를 마지막 layer로 제한하면 평균 4.1\\% 하락이 발생한다. 이는 NAG가 target feature 학습을 위한 희소한 ``functional backbone''을 포착함을 시사한다.

"},{"id":"64445","en":"Training Data Efficiency in Multimodal Process Reward Models","ko":"멀티모달 과정 보상 모델에서의 학습 데이터 효율성","authors":"Jinyuan Li, Chengsong Huang, Langlin Huang, Shaoyang Xu, Haolin Liu, Wenxuan Zhang, Jiaxin Huang","pos":"#1814","link":"https://openreview.net/forum?id=NWRoNoJ9QX","abs":"

Multimodal Process Reward Models (MPRMs) are central to step-level supervision for visual reasoning in MLLMs. Training MPRMs typically requires large-scale Monte Carlo (MC)-annotated corpora, incurring substantial training cost. This paper studies the data efficiency for MPRM training. Our preliminary experiments reveal that MPRM training quickly saturates under random subsampling of the training data, indicating substantial redundancy within existing MC-annotated corpora. To explain this, we formalize a theoretical framework and reveal that informative gradient updates depend on two factors: label mixtures of positive/negative steps and label reliability (average MC scores of positive steps). Guided by these insights, we propose the Balanced-Information Score (BIS), which prioritizes both mixture and reliability based on existing MC signals at the rollout level, without incurring any additional cost. Across two backbones (InternVL2.5-8B and Qwen2.5-VL-7B) on VisualProcessBench, BIS-selected subsets consistently match and even surpass the full-data performance at small fractions. Notably, the BIS subset reaches full-data performance using only 10% of the training data, improving over random subsampling by a relative 4.1%.

","absKo":"

Multimodal Process Reward Model (MPRM)은 MLLM에서 visual reasoning을 위한 step-level supervision의 핵심이다. MPRM을 학습하려면 일반적으로 대규모 Monte Carlo (MC) annotation corpus가 필요하며, 상당한 training cost가 발생한다. 본 논문은 MPRM 학습을 위한 data efficiency를 연구한다. 예비 실험 결과, training data를 무작위로 subsampling하면 MPRM 학습이 빠르게 포화되며, 이는 기존 MC-annotated corpus 내부에 상당한 redundancy가 있음을 시사한다. 이를 설명하기 위해 우리는 이론적 framework를 formalize하고, 정보가 있는 gradient update가 두 가지 요인, 즉 positive/negative step의 label mixture와 label reliability(positive step의 average MC score)에 의존함을 밝혔다. 이러한 통찰에 따라, 우리는 Balanced-Information Score (BIS)를 제안한다. BIS는 추가 비용 없이 rollout level에서 기존 MC signal을 바탕으로 mixture와 reliability를 모두 우선시한다. VisualProcessBench에서 두 backbone (InternVL2.5-8B 및 Qwen2.5-VL-7B)에 대해, BIS로 선택한 subset은 작은 비율의 데이터만으로도 전체 데이터 성능과 일치하거나 심지어 능가함을 일관되게 보였다. 특히 BIS subset은 학습 데이터의 10%만으로 전체 데이터 성능에 도달했으며, random subsampling 대비 상대적으로 4.1% 향상되었다.

"},{"id":"66295","en":"SPA: A Simple but Tough-to-Beat Baseline for Knowledge Injection","ko":"SPA: 지식 주입을 위한 간단하지만 이기기 어려운 베이스라인","authors":"Kexian Tang, Jiani Wang, Shaowen Wang, Kaifeng Lyu","pos":"#1911","link":"https://openreview.net/forum?id=4zGe2J2eJU","abs":"

While large language models (LLMs) are pretrained on massive amounts of data, their knowledge coverage remains incomplete in specialized, data-scarce domains, motivating extensive efforts to study synthetic data generation for knowledge injection. We propose SPA (Scaling Prompt-engineered Augmentation), a simple but tough-to-beat baseline that uses a small set of carefully designed prompts to generate large-scale synthetic data for knowledge injection. Through systematic comparisons, we find that SPAoutperforms several strong baselines. Furthermore, we identify two key limitations of prior approaches: (1) while RL-based methods may improve the token efficiency of LLM-based data augmentation at small scale, they suffer from diversity collapse as data scales, leading to diminishing returns; and (2) while multi-stage prompting may outperform simple augmentation methods, their advantages can disappear after careful prompt tuning. Our results suggest that, for knowledge injection, careful prompt design combined with straightforward large-scale augmentation can be surprisingly effective, and we hope SPA can serve as a strong baseline for future studies in this area.

","absKo":"

large language model (LLM)은 방대한 양의 데이터로 사전 학습되지만, 전문적이고 데이터가 부족한 도메인에서는 knowledge coverage가 여전히 불완전하며, 이로 인해 knowledge injection을 위한 synthetic data generation 연구가 활발히 이루어지고 있다. 우리는 지식 주입을 위해 신중하게 설계된 소수의 prompt를 사용해 대규모 synthetic data를 생성하는 단순하지만 이기기 어려운 baseline인 SPA (Scaling Prompt-engineered Augmentation)를 제안한다. 체계적인 비교를 통해, SPA가 여러 강력한 baseline보다 우수함을 확인했다. 또한 우리는 기존 접근법의 두 가지 핵심 한계를 규명했다. (1) RL-based method는 소규모에서는 LLM-based data augmentation의 token efficiency를 향상시킬 수 있지만, 데이터 규모가 커지면 diversity collapse를 겪어 수익이 감소한다. (2) multi-stage prompting은 단순한 augmentation 방법보다 더 나을 수 있지만, 신중한 prompt tuning 후에는 그 장점이 사라질 수 있다. 우리의 결과는 knowledge injection에서 세심한 prompt design과 단순한 대규모 augmentation의 결합이 놀라울 정도로 효과적일 수 있음을 시사하며, SPA가 이 분야 향후 연구를 위한 강력한 baseline으로 기능하길 기대한다.

"},{"id":"61272","en":"Revisiting Parameter-Based Knowledge Editing in Large Language Models: Theoretical Limits and Empirical Evidence","ko":"Large Language Model의 파라미터 기반 지식 편집 재고: 이론적 한계와 실증적 증거","authors":"Wanying Ren, Xin Song, Futing Wang, Guoxiu He, Aixin Sun","pos":"#1913","link":"https://openreview.net/forum?id=rxWzzqvHYZ","abs":"

Parameter-based knowledge editing updates the internal knowledge of large language models (LLMs) via localized weight modifications and has attracted significant attention. However, most existing methods overlook fundamental theoretical limitations and are rarely evaluated under realistic, practice-oriented settings. In this paper, we first present a theoretical analysis based on the Representation Space Collapse Hypothesis, explaining how localized parameter edits can propagate along fragile directions in the representation space, inducing global interference and ultimately causing reasoning collapse. Building on this insight, we conduct a comprehensive empirical evaluation by systematically varying knowledge complexity, number of edits, evaluation dimensions, and baseline methods. Our results show that parameter-based editing methods consistently damage core LLM capabilities. In contrast, a simple retrieval-based baseline reliably outperforms all parameter-editing methods across all evaluated conditions. These findings highlight that preserving the fundamental capabilities of LLMs after knowledge editing should be a central concern for future research. Data and code are provided in the supplementary material.

","absKo":"

Parameter-based knowledge editing은 localized weight modification을 통해 large language models (LLMs)의 내부 지식을 업데이트하며 큰 주목을 받아왔다. 그러나 기존의 대부분 방법은 근본적인 이론적 한계를 간과하고 있으며, 현실적이고 실무 지향적인 설정에서 거의 평가되지 않았다. 본 논문에서는 먼저 Representation Space Collapse Hypothesis에 기반한 이론적 분석을 제시하여, localized parameter edit가 representation space의 취약한 방향을 따라 어떻게 전파될 수 있는지 설명하고, 그 결과 global interference를 유발하여 궁극적으로 reasoning collapse를 초래하는지를 보인다. 이러한 통찰을 바탕으로, 우리는 knowledge complexity, edit 수, evaluation dimension, baseline method를 체계적으로 변화시키는 포괄적 실증 평가를 수행한다. 우리의 결과는 parameter-based editing 방법이 일관되게 LLM의 핵심 능력을 손상시킨다는 것을 보여준다. 반면, 단순한 retrieval-based baseline은 평가한 모든 조건에서 모든 parameter-editing 방법보다 안정적으로 우수한 성능을 보인다. 이러한 결과는 knowledge editing 이후에도 LLM의 근본적 능력을 보존하는 것이 향후 연구의 중심 과제가 되어야 함을 강조한다. 데이터와 코드는 부록 자료에 제공된다.

"},{"id":"61196","en":"ProactiveLLM: Learning Active Interaction for Streaming Large Language Models","ko":"ProactiveLLM: Streaming Large Language Model을 위한 능동 상호작용 학습","authors":"Junlong Tong, Yao Zhang, Anhao Zhao, Yingqi Fan, Yunpu Ma, Xiaoyu Shen","pos":"#201","link":"https://openreview.net/forum?id=sr1qgowJQ1","abs":"

Standard Large Language Models (LLMs) operate on a ''read-then-generate'' paradigm, incurring avoidable latency and computational redundancy. Recently, streaming LLMs have attempted to overcome these bottlenecks by allowing input and output to unfold synchronously. However, this introduces a critical challenge: how should the LLM determine the optimal timing to interact with the input and output stream? Existing approaches remain confined to passive adaptation, relying on static or content-irrelevant interaction rules. In this paper, we propose ProactiveLLM, which achieves active interaction by treating ''when to generate\" and ''what to generate\" as decoupled objectives. Through masked streaming modeling and self-distillation, the model actively learns to perceive semantic sufficiency from partial inputs. This yields endogenous cues serving as a versatile interface for the plug-and-play integration of diverse decision heads customized for specific latency-accuracy trade-offs. Extensive evaluation across text and speech streaming tasks confirms that ProactiveLLM significantly reduces interaction latency while maintaining quality, validating its capacity for dynamic and active interaction.

","absKo":"

표준 Large Language Models (LLMs)은 ''read-then-generate'' paradigm으로 동작하여, 피할 수 있는 latency와 계산 중복을 초래한다. 최근 streaming LLM은 입력과 출력을 동기적으로 전개하도록 허용함으로써 이러한 병목을 극복하려고 시도하고 있다. 그러나 이는 중요한 과제를 제기한다. LLM이 입력 및 출력 stream과 상호작용할 최적의 시점을 어떻게 결정해야 하는가? 기존 접근법은 여전히 passive adaptation에 머물러 있으며, 정적이거나 내용과 무관한 상호작용 규칙에 의존한다. 본 논문에서는 ''when to generate\"와 ''what to generate\"를 분리된 objective로 다루어 능동적 상호작용을 달성하는 ProactiveLLM을 제안한다. masked streaming modeling과 self-distillation을 통해, 모델은 부분 입력만으로 의미적 충분성을 능동적으로 인식하는 법을 학습한다. 이로부터 다양한 지연-정확도 trade-off에 맞춰 설계된 여러 decision head를 plug-and-play 방식으로 통합할 수 있는 다목적 interface 역할을 하는 endogenous cue가 생성된다. 텍스트 및 음성 streaming task 전반에 걸친 광범위한 평가는 ProactiveLLM이 품질을 유지하면서 상호작용 latency를 크게 줄임을 확인해 주며, 이는 동적이고 능동적인 상호작용 능력을 입증한다.

"},{"id":"60748","en":"Real-Time Aligned Reward Model beyond Semantics","ko":"의미를 넘어서는 실시간 정렬 Reward Model","authors":"Zixuan Huang, Xin Xia, Yuxi Ren, Jianbin Zheng, Xuefeng Xiao, Hongyan Xie, Huaqiu Li, Songshi Liang, Zhongxiang Dai, Fuzhen Zhuang, Jianxin Li, Yikun Ban, deqing wang","pos":"#2011","link":"https://openreview.net/forum?id=wz2zK4l3YJ","abs":"

Reinforcement Learning from Human Feedback (RLHF) is a pivotal technique for aligning large language models (LLMs) with human preferences, yet it is susceptible to reward overoptimization, in which policy models overfit to the reward model, exploit spurious reward patterns instead of faithfully capturing human intent. Prior mitigations primarily relies on surface semantic information and fails to efficiently address the misalignment between the reward model (RM) and the policy model caused by continuous policy distribution shifts. This inevitably leads to an increasing reward discrepancy, exacerbating reward overoptimization. To address these limitations, we introduce R2M (Real-Time Aligned Reward Model), a novel lightweight RLHF framework. R2M goes beyond vanilla reward models that solely depend on the semantic representations of a pretrained LLM. Instead, it leverages the evolving hidden states of the policy (namely policy feedback) to align with the real-time distribution shift of the policy during the RL process. This work points to a promising new direction for improving the performance of reward models through real-time utilization of feedback from policy models.

","absKo":"

Reinforcement Learning from Human Feedback (RLHF)는 대규모 언어 모델(LLM)을 인간 선호에 맞추는 핵심 기술이지만, reward overoptimization에 취약하다. 이 현상은 policy model이 reward model에 과도하게 적합하여, 인간의 의도를 충실히 반영하기보다 허상적인 reward pattern을 악용할 때 발생한다. 기존 완화 방법은 주로 표면적 semantic information에 의존하며, 지속적인 policy distribution shift로 인해 reward model (RM)과 policy model 사이에 발생하는 misalignment를 효율적으로 해결하지 못한다. 이는 필연적으로 reward discrepancy를 증가시켜 reward overoptimization을 악화시킨다. 이러한 한계를 해결하기 위해, 우리는 새로운 경량 RLHF framework인 R2M (Real-Time Aligned Reward Model)을 도입한다. R2M은 사전 학습된 LLM의 semantic representation에만 의존하는 기존 reward model을 넘어, policy의 진화하는 hidden state, 즉 policy feedback을 활용하여 RL 과정에서 policy의 real-time distribution shift와 정렬한다. 본 연구는 policy model로부터의 feedback을 실시간으로 활용하여 reward model의 성능을 향상시키는 유망한 새로운 방향을 제시한다.

"},{"id":"65246","en":"Regret Pre-training: Bridging Prior and Posterior Views for Enhanced Knowledge Grounding","ko":"Regret Pre-training: 향상된 지식 그라운딩을 위한 사전 및 사후 관점 연결","authors":"Mingkuan Zhao, Xiayu Sun, Wentao Hu, Suquan Chen, Jiaxuan Li, Xiaoyan Zhu, Xin Lai, Jiayin Wang","pos":"#2015","link":"https://openreview.net/forum?id=FNZs7SLEEI","abs":"

Causal language models factorize sequence probabilities using only preceding context, leaving future information unexploited during training despite its availability in the training data. This paper introduces Regret Pre-training, a self-supervised framework grounded in the Learning Using Privileged Information (LUPI) paradigm. The framework employs a dual-view architecture in which a single model generates both a causal Student distribution and a future-conditioned Teacher distribution. The training objective augments standard language modeling with a regret loss that minimizes the KL divergence from teacher to student, transferring future-aware signals to the causal representations. We investigate two teacher configurations on the OLMoE-1B-7B architecture:LocalRegret, which extends attention by one future token, andGlobalRegret, which conditions on bidirectional context with the target position masked. Experiments on nine downstream tasks following 4 billion tokens of training demonstrate that both configurations consistently outperform the baseline. On average,GlobalRegret andLocalRegret achieve 33.9% and 32.2% accuracy respectively, surpassing the baseline's 30.2%. Most notably,GlobalRegret improves BoolQ performance by 18.1 percentage points (61.0% vs 42.9%). The framework introduces no additional parameters and requires only one extra inference-mode forward pass per training step. The source code for this paper is publicly available at https://anonymous.4open.science/r/ICML13655/ to facilitate reproducibility.

","absKo":"

Causal language model은 선행 context만을 사용해 sequence probability를 factorize하므로, training data에서 future information이 사용 가능함에도 training 동안 이를 활용하지 못한다. 이 논문은 Learning Using Privileged Information (LUPI) 패러다임에 기반한 self-supervised framework인 Regret Pre-training을 소개한다. 이 framework는 단일 model이 causal Student distribution과 future-conditioned Teacher distribution을 모두 생성하는 dual-view architecture를 사용한다. training objective는 standard language modeling에 teacher에서 student로의 KL divergence를 최소화하는 regret loss를 더해, future-aware signal을 causal representation으로 전달한다. 우리는 OLMoE-1B-7B architecture에서 두 가지 teacher 구성을 조사한다. LocalRegret은 attention을 미래의 토큰 하나까지 확장하고, GlobalRegret은 target position을 mask한 상태에서 bidirectional context에 조건을 건다. 40억 token의 training 뒤 9개의 downstream task에서 수행한 실험은 두 구성 모두 baseline보다 일관되게 우수함을 보여 준다. 평균적으로 GlobalRegret과 LocalRegret은 각각 33.9%와 32.2%의 accuracy를 달성하여 baseline의 30.2%를 능가했다. 특히 GlobalRegret은 BoolQ 성능을 18.1 percentage points 향상시켰다(61.0% 대 42.9%). 이 framework는 추가 parameter를 전혀 도입하지 않으며, training step마다 inference-mode forward pass를 한 번만 추가로 요구한다. 이 논문의 source code는 재현성을 높이기 위해 https://anonymous.4open.science/r/ICML13655/ 에서 공개되어 있다.

"},{"id":"60938","en":"Learning Rate Scaling across LoRA Ranks and Transfer to Full Finetuning","ko":"LoRA Rank 전반의 Learning Rate Scaling과 Full Finetuning으로의 전이","authors":"Nan Chen, Soledad Villar, Soufiane Hayou","pos":"#204","link":"https://openreview.net/forum?id=vCAIwEJZQ6","abs":"Low-Rank Adaptation (LoRA) is a standard tool for parameter-efficient finetuning of large models. While it induces a small memory footprint, its training dynamics can be surprisingly complex as they depend on several hyperparameters such as initialization, adapter rank, and learning rate. In particular, it is unclear *how the optimal learning rate scales with adapter rank*, which forces practitioners to re-tune the learning rate whenever the rank is changed. In this paper, we introduce *Maximal-Update Adaptation* ($\\mu$A), a theoretical framework that characterizes how the \"optimal\" learning rate should scale with model width and adapter rank to produce stable, non-vanishing feature updates under standard configurations. Our analysis leverages techniques from hyperparameter transfer and reveals that the optimal learning rate exhibits different scaling patterns depending on initialization and LoRA scaling factor. Specifically, we identify two regimes: one where the optimal learning rate remains roughly invariant across ranks, and another where it scales inversely with rank. We further identify a configuration that allows learning rate transfer from LoRA to full finetuning, drastically reducing the cost of learning rate tuning for full finetuning. Experiments across language, vision, vision–language, image generation, and reinforcement-learning tasks validate our scaling rules and show that learning rates tuned on LoRA transfer reliably to full finetuning.","absKo":"Low-Rank Adaptation(LoRA)는 large model의 parameter-efficient finetuning을 위한 표준 도구이다. 이는 작은 memory footprint를 유도하지만, 초기화, adapter rank, learning rate와 같은 여러 hyperparameter에 의존하기 때문에 training dynamics는 놀라울 정도로 복잡할 수 있다. 특히, *adapter rank에 따라 optimal learning rate가 어떻게 scaling되는지*가 불분명하여, rank를 바꿀 때마다 practitioner가 learning rate를 다시 조정해야 하는 문제가 있다. 본 논문에서는 표준 configuration에서 안정적이고 소실되지 않는 feature update를 생성하기 위해 \"optimal\" learning rate가 model width와 adapter rank에 따라 어떻게 scaling되어야 하는지를 특징짓는 이론적 framework인 *Maximal-Update Adaptation*($\\mu$A)을 소개한다. 우리의 분석은 hyperparameter transfer의 기법을 활용하며, optimal learning rate가 initialization과 LoRA scaling factor에 따라 서로 다른 scaling pattern을 보인다는 점을 드러낸다. 구체적으로, 두 가지 regime를 식별한다. 하나는 optimal learning rate가 rank 전반에 걸쳐 대체로 불변인 경우이고, 다른 하나는 rank에 반비례하여 scaling되는 경우이다. 또한 LoRA에서 full finetuning으로 learning rate transfer를 가능하게 하는 configuration을 찾아내어, full finetuning의 learning rate 튜닝 비용을 획기적으로 줄인다. 언어, vision, vision–language, image generation, reinforcement-learning task 전반의 실험은 우리의 scaling rule을 검증하며, LoRA에서 튜닝한 learning rate가 full finetuning으로도 안정적으로 transfer됨을 보여준다."},{"id":"66105","en":"OCNR: Stabilizing Self-Play by Mitigating Iteration-Collapse With One-Class Novelty Rewards","ko":"OCNR: One-Class 참신성 보상으로 반복 붕괴를 완화하여 셀프플레이 안정화하기","authors":"Seungyoo Lee, Giung Nam, Hyungi Lee, Juho Lee","pos":"#2110","link":"https://openreview.net/forum?id=70lhVsTZ62","abs":"

Training large language models via self-play often suffers from a persistent iteration-collapse, where performance initially improves but subsequently regresses as training iterations increase. We analyze this phenomenon as arising from cross-iteration degeneration, where the task-generation distribution becomes increasingly confined to a narrow subset of familiar (seen) problems, weakening the effective learning signal and destabilizing training. To address this issue, we propose a plug-in approach that augments existing self-play pipelines with a one-class novelty reward. A Seen Detector trained on a historical buffer of previously used training problems identifies in-support instances and discourages redundant generation by the questioner, thereby steering exploration toward under-explored yet learnable regions. Experimental results show that the proposed method mitigates iteration-collapse during iterative training and yields consistent improvements.

","absKo":"

self-play를 통해 large language model을 training할 때는 종종 지속적인 iteration-collapse가 발생하는데, 이는 처음에는 성능이 향상되다가 training iteration이 늘어날수록 다시 저하되는 현상이다. 우리는 이 현상을 cross-iteration degeneration의 결과로 분석한다. 즉, task-generation distribution이 점점 익숙한(seen) problem의 좁은 부분집합에 갇히게 되어, 유효한 learning signal이 약화되고 training이 불안정해진다. 이 문제를 해결하기 위해, 기존 self-play pipeline에 one-class novelty reward를 추가하는 plug-in approach를 제안한다. 과거에 사용된 training problem의 historical buffer로 학습된 Seen Detector가 in-support instance를 식별하고 questioner의 redundant generation을 억제함으로써, 탐색을 아직 충분히 탐색되지 않았지만 학습 가능한 영역으로 유도한다. 실험 결과, 제안 방법은 iterative training 동안 iteration-collapse를 완화하고 일관된 성능 향상을 가져옴을 보여준다.

"},{"id":"65361","en":"Know More, Know Clearer: A Meta-Cognitive Framework for Knowledge Augmentation in Large Language Models","ko":"더 많이 알고, 더 명확히 알기: Large Language Model의 지식 증강을 위한 메타 인지 프레임워크","authors":"Hao Chen, Ye He, Yuchun Fan, Yukun Yan, Zhenghao Liu, Qingfu Zhu, Maosong Sun, Wanxiang Che","pos":"#2203","link":"https://openreview.net/forum?id=ENuMNYCiV6","abs":"

Knowledge augmentation has significantly enhanced the performance of Large Language Models (LLMs) in knowledge-intensive tasks. However, existing methods typically operate on the simplistic premise that model performance equates with internal knowledge, overlooking the knowledge-confidence gaps that lead to overconfident errors or uncertain truths. To bridge this gap, we propose a novel meta-cognitive framework for reliable knowledge augmentation via differentiated intervention and alignment. Our approach leverages internal cognitive signals to partition the knowledge space into mastered, confused, and missing regions, guiding targeted knowledge expansion. Furthermore, we introduce a cognitive consistency mechanism to synchronize subjective certainty with objective accuracy, ensuring calibrated knowledge boundaries. Extensive experiments demonstrate the our framework consistently outperforms strong baselines, validating its rationality in not only enhancing knowledge capabilities but also fostering cognitive behaviors that better distinguish knowns from unknowns.

","absKo":"

Knowledge augmentation은 knowledge-intensive task에서 Large Language Model(LLM)의 성능을 크게 향상시켜 왔다. 그러나 기존 방법은 대개 모델 성능이 내부 지식과 일치한다는 단순한 전제하에 작동하며, 과신된 오류나 불확실한 진실로 이어지는 knowledge-confidence gap을 간과한다. 이러한 간극을 메우기 위해, 우리는 differentiated intervention과 alignment를 통한 신뢰할 수 있는 knowledge augmentation을 위한 새로운 meta-cognitive framework를 제안한다. 우리의 접근법은 내부 cognitive signal을 활용해 knowledge space를 숙련됨(masters), 혼란됨(confused), 결여됨(missing)의 영역으로 분할하고, 표적화된 knowledge expansion을 안내한다. 더 나아가, 주관적 확신과 객관적 정확도를 동기화하는 cognitive consistency mechanism을 도입하여, 보정된 knowledge boundary를 보장한다. 광범위한 실험은 우리의 framework가 강력한 baseline을 일관되게 능가함을 보여주며, 지식 능력을 향상시킬 뿐 아니라 아는 것과 모르는 것을 더 잘 구분하는 cognitive behavior를 촉진한다는 점에서 그 합리성을 검증한다.

"},{"id":"65293","en":"HugRAG: Hierarchical Causal Knowledge Graph Design for RAG","ko":"HugRAG: RAG를 위한 계층적 인과 지식 그래프 설계","authors":"Nengbo Wang, Tuo Liang, Vikash Singh, Chaoda Song, Van Yang, Yu Yin, Jing Ma, JAGDIP SINGH, Vipin Chaudhary","pos":"#2304","link":"https://openreview.net/forum?id=F2zHYjBAEv","abs":"

Retrieval augmented generation (RAG) has enhanced large language models by enabling access to external knowledge, with graph-based RAG emerging as a powerful paradigm for structured retrieval and reasoning. However, existing graph-based methods often over-rely on surface-level node matching and lack explicit causal modeling, leading to unfaithful or spurious answers. Prior attempts to incorporate causality are typically limited to local or single-document contexts and also suffer from information isolation that arises from modular graph structures, which hinders scalability and cross-module causal reasoning. To address these challenges, we propose HugRAG, a framework that rethinks knowledge organization for graph-based RAG through causal gating across hierarchical modules. HugRAG explicitly models causal relationships to suppress spurious correlations while enabling scalable reasoning over large-scale knowledge graphs. Extensive experiments demonstrate that HugRAG consistently outperforms competitive graph-based RAG baselines across multiple datasets and evaluation metrics. Our work establishes a principled foundation for structured, scalable, and causally grounded RAG systems.

","absKo":"

retrieval augmented generation(RAG)은 외부 지식에 대한 접근을 가능하게 함으로써 large language model을 향상시켜 왔으며, 그중 graph-based RAG는 구조화된 retrieval과 reasoning을 위한 강력한 paradigm으로 부상하고 있다. 그러나 기존의 graph-based 방법은 surface-level node matching에 과도하게 의존하고 explicit causal modeling이 부족하여, 사실과 어긋나거나(spurious) 부정확한 답변을 초래하는 경우가 많다. causality를 통합하려는 기존 시도는 대체로 local 또는 single-document context에 국한되며, modular graph structure에서 비롯되는 information isolation 문제도 겪고 있어 scalability와 cross-module causal reasoning을 저해한다. 이러한 문제를 해결하기 위해, 우리는 hierarchical module 전반의 causal gating을 통해 graph-based RAG의 knowledge organization을 재고하는 framework인 HugRAG를 제안한다. HugRAG는 causal relationship을 명시적으로 모델링하여 spurious correlation을 억제하는 동시에, 대규모 knowledge graph에 대한 scalable reasoning을 가능하게 한다. 광범위한 실험은 HugRAG가 여러 dataset과 평가 metric 전반에서 경쟁력 있는 graph-based RAG baseline을 일관되게 능가함을 보여준다. 우리의 연구는 구조화되고, scalable하며, causal grounding을 갖춘 RAG system을 위한 원리적 기반을 확립한다.

"},{"id":"65759","en":"From Retrieval to Translation: Translating Query into Graph-level Clues for Retrieval-Augmented Generation","ko":"검색에서 번역으로: 검색 증강 생성을 위해 질의를 그래프 수준 단서로 번역하기","authors":"Qichuan Liu, Chenfeng Zheng, Yuxuan Hu, Zerui Chen, Chentao Zhang, Qinggang Zhang, Zhihong Zhang","pos":"#2311","link":"https://openreview.net/forum?id=AS14E8aMmL","abs":"

Retrieval-Augmented Generation (RAG) has recently been enhanced with tree or graph structures to match user intent for precise passage retrieval, which facilitates large language models (LLMs) in effectively mitigating hallucinations by leveraging external knowledge. However, we identify that existing structure-augmented RAG systems are experiencing (i) potential retrieval suspension and (ii) cumulative semantic drift, due to low-quality structures and semantic embeddings that often poorly capture textual details. Motivated by this, we propose a novel paradigm named KG-Translator, which is distinct from traditional matching-based paradigms and instead translates user queries into graph-level clues. Specifically, KG-Translator utilizes lightweight models to conduct named entity recognition (NER) and syntactic parsing on the corpus, constructing a reliable knowledge graph (ParseKG). On top of ParseKG, KG-Translator adopts constrained decoding strategies to faithfully translate clues, traces them to original passages, and employs a lightweight ranking model for precise passage retrieval. Extensive experiments on five datasets demonstrate that KG-Translator significantly outperforms baselines.

","absKo":"

Retrieval-Augmented Generation (RAG)은 최근 tree 또는 graph 구조를 활용해 정밀한 passage retrieval을 위한 사용자 의도 정합성을 높였고, 이를 통해 large language model (LLM)이 external knowledge를 활용하여 hallucination을 효과적으로 완화할 수 있게 되었다. 그러나 우리는 기존의 structure-augmented RAG 시스템이 (i) 잠재적인 retrieval suspension과 (ii) 누적되는 semantic drift를 겪고 있음을 확인한다. 이는 저품질 구조와, 텍스트 세부사항을 충분히 포착하지 못하는 semantic embedding 때문인 경우가 많다. 이러한 문제의식에서 우리는 기존의 matching-based paradigm과 구별되는, 사용자 query를 graph-level clue로 변환하는 새로운 패러다임 KG-Translator를 제안한다. 구체적으로, KG-Translator는 경량 모델을 사용해 corpus에 대해 named entity recognition (NER)과 syntactic parsing을 수행하고, 신뢰할 수 있는 knowledge graph (ParseKG)를 구축한다. ParseKG 위에서 KG-Translator는 constrained decoding 전략을 채택해 clue를 충실히 translate하고, 이를 원래 passage에 추적한 뒤, 정밀한 passage retrieval을 위해 경량 ranking model을 사용한다. 다섯 개 데이터셋에 대한 광범위한 실험은 KG-Translator가 baseline보다 유의미하게 우수함을 보여준다.

"},{"id":"62250","en":"From Outcomes to Actions: Leveraging Hindsight for Long-Horizon Language Agent Training","ko":"결과에서 행동으로: Hindsight 활용으로 장기 언어 에이전트 학습","authors":"Zishang Jiang, tingyun li, Jinyi Han, Xinyi Wang, Mengyun Qiao, Yizhou Ying, Xiaojun Meng, Jiansheng Wei, Jiaqing Liang, Yanghua Xiao","pos":"#2312","link":"https://openreview.net/forum?id=iK3yDEvQ4y","abs":"

Reinforcement learning (RL) has become a widely adopted technique for improving large language models (LLMs) on complex tasks. Despite this progress, existing RL methods still face challenges in training agents with longer-horizon interactions. One major bottleneck is distinguishing the contribution of different actions in long-horizon interaction, leading to high optimization variance. To address this, we introduce a novel policy gradient method, Hindsight Policy Optimization (HPO), that projects both the current policy distribution and the hindsight distribution into an intent space and extracts low-variance learning signals from the Wasserstein distance between them. We theoretically and empirically show that aggregating semantically similar states and actions in the intent space yields a bounded-variance estimator and improves policy performance stably. Our code is available online.

","absKo":"

강화학습 (RL)은 복잡한 작업에서 Large Language Models (LLMs)을 향상시키기 위해 널리 채택된 기법이 되었다. 이러한 진전에도 불구하고, 기존 RL 방법은 여전히 더 긴 horizon 상호작용을 가진 agent를 학습시키는 데 어려움을 겪는다. 주요 병목 중 하나는 long-horizon 상호작용에서 서로 다른 action의 기여를 구분하는 것이며, 이는 높은 최적화 분산을 초래한다. 이를 해결하기 위해 우리는 현재 policy distribution과 hindsight distribution을 intent space로 투영하고, 그 사이의 Wasserstein distance로부터 저분산 학습 신호를 추출하는 새로운 policy gradient 방법인 Hindsight Policy Optimization (HPO)을 제안한다. 우리는 의미적으로 유사한 state와 action을 intent space에서 집계하면 분산이 제한된 estimator를 얻을 수 있고 policy 성능이 안정적으로 향상됨을 이론적으로 그리고 실증적으로 보인다. 코드는 온라인에서 제공된다.

"},{"id":"64040","en":"Discounted Beta-Bernoulli Reward Estimation for Sample-Efficient Reinforcement Learning with Verifiable Rewards","ko":"검증 가능한 보상을 활용한 샘플 효율적 강화학습을 위한 할인된 Beta-Bernoulli 보상 추정","authors":"Haechan Kim, Soohyun Ryu, Gyouk Chu, Doohyuk Jang, Eunho Yang","pos":"#2400","link":"https://openreview.net/forum?id=RUheyL9bb9","abs":"

Reinforcement learning with verifiable rewards (RLVR) has emerged as an effective post-training paradigm for improving the reasoning capabilities of large language models. However, existing group-based RLVR methods often suffer from severe sample inefficiency. This inefficiency stems from reliance on point estimation of rewards from a small number of rollouts, leading to high estimation variance, variance collapse, and ineffective utilization of generated responses. In this work, we reformulate RLVR from a statistical estimation perspective by modeling rewards as samples drawn from a policy-induced distribution and casting advantage computation as the problem of estimating the reward distribution from finite data. Building on this view, we propose Discounted Beta-Bernoulli (DBB) reward estimation, which leverages historical reward statistics for the non-stationary distribution. Although biased, the resulting estimator exhibits reduced and stable variance, theoretically avoids estimated variance collapse, and achieves lower mean squared error than standard point estimation. Extensive experiments across six in-distribution and three out-of-distribution reasoning benchmarks demonstrate that GRPO with DBB consistently outperforms naive GRPO, achieving average Acc@8 improvements of 3.22/2.42 points in-distribution and 12.49/6.92 points out-of-distribution on the 1.7B and 8B models, respectively, without additional computational cost or memory usage.

","absKo":"verifiable reward를 사용하는 reinforcement learning (RLVR)은 large language model의 reasoning 능력을 향상시키는 효과적인 post-training paradigm으로 부상했다. 그러나 기존의 group-based RLVR 방법은 종종 심각한 sample inefficiency를 겪는다. 이러한 비효율은 소수의 rollout에서 reward를 point estimation에 의존하는 데서 비롯되며, 그 결과 추정 variance가 크고, variance collapse가 발생하며, 생성된 response의 활용도도 떨어진다. 본 연구에서는 reward를 policy가 유도한 distribution에서 샘플링된 관측치로 모델링하고, advantage computation을 유한한 데이터로부터 reward distribution을 추정하는 문제로 정식화함으로써 RLVR을 statistical estimation 관점에서 재구성한다. 이 관점에 기반하여, 우리는 비정상 분포에 대해 historical reward statistics를 활용하는 Discounted Beta-Bernoulli (DBB) reward estimation을 제안한다. 이 추정기는 bias가 있지만 variance가 더 작고 안정적이며, 이론적으로 estimated variance collapse를 피하고 표준 point estimation보다 낮은 mean squared error를 달성한다. 6개의 in-distribution reasoning benchmark와 3개의 out-of-distribution reasoning benchmark에 대한 광범위한 실험은 DBB를 사용한 GRPO가 naive GRPO보다 일관되게 우수함을 보여주며, 1.7B 및 8B 모델에서 추가적인 계산 비용이나 memory 사용 없이 in-distribution에서는 평균 Acc@8이 각각 3.22/2.42 point, out-of-distribution에서는 12.49/6.92 point 향상됨을 보인다.

"},{"id":"66459","en":"Decoupling Reasoning and Confidence: Resurrecting Calibration in Reinforcement Learning from Verifiable Rewards","ko":"추론과 신뢰도의 분리: 검증 가능한 보상 기반 강화 학습에서 보정의 부활","authors":"Zhengzhao Ma, Xueru Wen, Boxi Cao, Yaojie Lu, Hongyu Lin, Jinglin Yang, Min He, Xianpei Han, Le Sun","pos":"#2503","link":"https://openreview.net/forum?id=3V1p2bJugq","abs":"

Reinforcement Learning from Verifiable Rewards (RLVR) significantly enhances large language models (LLMs) reasoning but severely suffers from calibration degeneration, where models become excessively over-confident in incorrect answers. Previous studies devote to directly incorporating calibration objective into existing optimization target. However, our theoretical analysis demonstrates that there exists a fundamental gradient conflict between the optimization for maximizing policy accuracy and minimizing calibration error. Building on this insight, we propose DCPO, a simple yet effective framework that systematically decouples reasoning and calibration objectives. Extensive experiments demonstrate that our DCPO not only preserves accuracy on par with GRPO but also achieves the best calibration performance and substantially mitigates the over-confidence issue. Our study provides valuable insights and practical solution for more reliable LLM deployment.

","absKo":"

Verifiable Rewards를 통한 Reinforcement Learning(RLVR)은 large language models(LLMs)의 추론 능력을 크게 향상시키지만, calibration degeneration, 즉 모델이 잘못된 답에 지나치게 과도한 confidence를 보이는 문제로 심각하게 고통받는다. 기존 연구들은 calibration objective를 기존 optimization target에 직접 통합하는 데 집중해 왔다. 그러나 우리의 이론적 분석은 policy accuracy를 최대화하는 optimization과 calibration error를 최소화하는 optimization 사이에 근본적인 gradient conflict가 존재함을 보여준다. 이러한 통찰을 바탕으로, 우리는 reasoning objective와 calibration objective를 체계적으로 decouple하는 간단하지만 효과적인 framework인 DCPO를 제안한다. 광범위한 실험은 우리의 DCPO가 GRPO와 동등한 수준의 accuracy를 유지할 뿐만 아니라, 최고의 calibration performance를 달성하고 over-confidence 문제를 실질적으로 완화함을 보여준다. 우리의 연구는 더 신뢰할 수 있는 LLM deployment를 위한 유용한 통찰과 실용적 해결책을 제공한다.

"},{"id":"60972","en":"Decouple Searching from Training: Scaling Data Mixing via Model Merging for Large Language Model Pre-training","ko":"Search와 Training 분리: Large Language Model Pre-training을 위한 Model Merging 기반 Data Mixing 확장","authors":"Shengrui Li, Fei zhao, Kaiyan Zhao, Jieying Ye, Haifeng Liu, Fangcheng Shi, Zheyong Xie, Yao Hu, Shaosheng Cao","pos":"#2504","link":"https://openreview.net/forum?id=uyRIOjFgOn","abs":"

Determining an effective data mixture is a key factor in Large Language Model (LLM) pre-training, where models must balance general competence with proficiency on hard tasks such as math and code. However, identifying an optimal mixture remains an open challenge, as existing approaches either rely on unreliable tiny-scale proxy experiments or require prohibitively expensive large-scale exploration. To address this, we propose Decouple Searching from Training Mix (DeMix), a novel framework that leverages model merging to predict optimal data ratios. Instead of training proxy models for every sampled mixture, DeMix trains component models on candidate datasets at scale and derives data mixture proxies via weighted model merging. This paradigm decouples search from training costs, enabling evaluation of unlimited sampled mixtures without extra training burden and thus facilitating better mixture discovery through more search trials. Extensive experiments demonstrate that DeMix breaks the trade-off between sufficiency, accuracy and efficiency, obtaining the optimal mixture with higher benchmark performance at lower search cost. Additionally, we release the DeMix Corpora, a comprehensive 22T-token dataset comprising high-quality pre-training data with validated mixtures to facilitate open research.

","absKo":"

효과적인 data mixture를 결정하는 것은 Large Language Model(LLM) pre-training의 핵심 요소로, 모델은 general competence와 math, code 같은 어려운 task에서의 성능 사이를 균형 있게 맞춰야 한다. 그러나 최적의 mixture를 찾는 일은 여전히 열린 문제인데, 기존 접근법은 신뢰하기 어려운 소규모 proxy experiment에 의존하거나, 지나치게 비싼 대규모 탐색을 요구하기 때문이다. 이를 해결하기 위해 우리는 model merging을 활용하여 최적의 data ratio를 예측하는 새로운 framework인 Decouple Searching from Training Mix(DeMix)를 제안한다. 샘플링된 모든 mixture마다 proxy model을 학습하는 대신, DeMix는 후보 dataset 각각에 대해 component model을 대규모로 학습한 뒤 weighted model merging을 통해 data mixture proxy를 도출한다. 이 패러다임은 search와 training cost를 분리하여, 추가 학습 부담 없이 무제한의 샘플링된 mixture를 평가할 수 있게 하며, 더 많은 search trial을 통해 더 나은 mixture 발견을 가능하게 한다. 광범위한 실험은 DeMix가 sufficiency, accuracy, efficiency 사이의 trade-off를 깨뜨려, 더 낮은 search cost로 더 높은 benchmark performance를 갖는 최적의 mixture를 얻음을 보여준다. 추가로, 우리는 open research를 촉진하기 위해 검증된 mixture를 포함한 고품질 pre-training data로 구성된 포괄적 22T-token dataset인 DeMix Corpora를 공개한다.

"},{"id":"66145","en":"Controllable and explainable personality sliders for LLMs at inference time","ko":"추론 시점에서 LLM을 위한 제어 가능하고 설명 가능한 성격 슬라이더","authors":"Florian Hoppe, David Khachaturov, Robert Mullins, Mark Huasong Meng","pos":"#2509","link":"https://openreview.net/forum?id=6TuaAw1DkF","abs":"

Aligning Large Language Models (LLMs) with specific personas typically relies on Supervised Fine-Tuning (SFT) or Reinforcement Learning from Human Feedback (RLHF); however, these methods are resource-intensive, requiring expensive data collection and distinct model training for each target personality. In this work, we propose a parameter-efficient framework for continuous, multi-dimensional personality control via inference-time activation steering. Our approach addresses the challenge of combining multiple interventions by iteratively retraining probes on the residual stream modified by prior traits, ensuring compatibility. Once established, these steering vectors function as modular, reusable primitives; users can instantly synthesize novel, complex personality profiles by simply adjusting steering coefficients (α) without any additional training. To support this, we introduce an automated pipeline that identifies optimal intervention layers via activation separation analysis and calibrates coefficients via hyperparameter optimization to maximize alignment while constraining perplexity. Empirical evaluations validate individual trait shifts using an LLM-as-a-judge framework and demonstrate, via the Big Five inventory, that our method effectively modulates the model's holistic personality profile without updating base model parameters.

","absKo":"

특정 persona에 Large Language Models(LLMs)을 정렬시키는 작업은 일반적으로 Supervised Fine-Tuning(SFT) 또는 Reinforcement Learning from Human Feedback(RLHF)에 의존한다. 그러나 이러한 방법은 자원 소모가 크며, 각 목표 personality마다 비용이 큰 데이터 수집과 별도의 model training을 요구한다. 본 연구에서는 inference-time activation steering을 통해 연속적이고 다차원적인 personality control을 위한 parameter-efficient framework를 제안한다. 우리의 접근은 prior traits에 의해 수정된 residual stream 위에서 probes를 반복적으로 재학습함으로써 여러 intervention을 결합하는 어려움을 해결하고, 호환성을 보장한다. 일단 구축되면 이러한 steering vectors는 modular하고 reusable한 primitive로 기능하며, 사용자는 추가 학습 없이 steering coefficient(α)만 조정하여 새롭고 복잡한 personality profile을 즉시 합성할 수 있다. 이를 지원하기 위해 activation separation analysis를 통해 최적의 intervention layer를 식별하고, perplexity를 제약하면서 alignment를 최대화하도록 hyperparameter optimization으로 coefficient를 보정하는 automated pipeline을 도입한다. 실험 평가는 LLM-as-a-judge framework를 사용해 개별 trait shift를 검증하고, Big Five inventory를 통해 base model parameters를 업데이트하지 않고도 우리의 방법이 모델의 holistic personality profile을 효과적으로 조절함을 보여준다.

"},{"id":"66042","en":"Can Large Language Models Generalize Procedures Across Representations?","ko":"대형 언어 모델은 표현에 걸쳐 절차를 일반화할 수 있는가?","authors":"Fangru Lin, Valentin Hofmann, Xingchen Wan, Weixing Wang, Zifeng Ding, Anthony Cohn, Janet Pierrehumbert","pos":"#2513","link":"https://openreview.net/forum?id=7cMEmnBAQX","abs":"

Large language models (LLMs) are trained and tested extensively on symbolic representations such as code and graphs, yet real-world user tasks are often specified in natural language. To what extent can LLMs generalize across these representations? Here, we approach this question by studying isomorphic tasks involving procedures represented in code, graphs, and natural language (e.g., scheduling steps in planning). We find that training LLMs with popular post-training methods on graphs or code data alone does not reliably generalize to corresponding natural language tasks, while training solely on natural language can lead to inefficient performance gains. To address this gap, we propose a two-stage data curriculum that first trains on symbolic, then natural language data. The curriculum substantially improves model performance across model families and tasks. Remarkably, a 1.5B Qwen model trained by our method can closely match zero-shot GPT-4o in naturalistic planning. Finally, our analysis suggests that successful cross-representation generalization can be interpreted as a form of generative analogy, which our curriculum effectively encourages.

","absKo":"

Large language models (LLMs)는 코드와 graph 같은 symbolic representation으로 대규모로 학습되고 테스트되지만, 실제 사용자 task는 종종 natural language로 명시된다. LLM은 이러한 representation들 사이에서 어느 정도까지 일반화할 수 있을까? 여기서는 code, graph, natural language로 표현된 procedure를 포함하는 isomorphic task를 연구함으로써 이 질문에 접근한다(예: planning에서의 scheduling step). 우리는 graph나 code 데이터에만 popular post-training method를 사용해 LLM을 학습시키는 것이 대응되는 natural language task로 신뢰성 있게 generalize되지 않으며, 반대로 natural language로만 학습하면 성능 향상이 비효율적으로 이루어질 수 있음을 발견했다. 이 간극을 해결하기 위해 우리는 먼저 symbolic 데이터로 학습한 뒤 natural language 데이터로 학습하는 two-stage data curriculum을 제안한다. 이 curriculum은 model family와 task 전반에서 모델 성능을 크게 향상시킨다. 특히, 우리의 방법으로 학습한 1.5B Qwen model은 naturalistic planning에서 zero-shot GPT-4o와 매우 근접한 성능을 낼 수 있다. 마지막으로, 우리의 분석은 성공적인 cross-representation generalization이 generative analogy의 한 형태로 해석될 수 있으며, 우리의 curriculum이 이를 효과적으로 촉진함을 시사한다.

"},{"id":"62051","en":"Anchoring Self-Play for Code Repair","ko":"코드 복구를 위한 앵커링 Self-Play","authors":"Caroline Choi, Zeyneb Kaya, Shirley Wu, Tengyu Ma, Tatsunori Hashimoto, Ludwig Schmidt","pos":"#2601","link":"https://openreview.net/forum?id=kJ6dvGnV2f","abs":"Code repair is an important capability for language models (LMs): given a buggy program and unit tests, an LM must produce a fixed program that passes the tests. We aim to scale supervision for code repair by having an LM generate bug--fix tasks with unconstrained edits, using unit tests as the only verifier. We propose generator-fixer self-play, in which a single model is trained with reinforcement learning to alternate between generating bugs and fixing them. As the fixer improves, the generator adapts to produce increasingly difficult bugs, yielding an automatic curriculum. However, because unit tests certify correctness but not realism, we find that the generator can drift from bugs encountered in practice, improving repair on self-generated bugs while degrading on real-world bugs. We propose Anchored Self Play (ASP), which anchors self-play with a small reference set by (i) adding a code-embedding similarity reward to guide generation and (ii) mixing reference bugs into fixer training to prevent drift. To reflect LM-assisted programming, where bugs come from humans, LMs, and human edits of LM code, we introduce BugSourceBench, a code repair benchmark spanning human-authored bugs, human-edited buggy LM code, and errors in LM-generated code. Across bug sources, ASP achieves the best fix rates, improving average fix rate by $+25$% (relative) / $+7.2$ pp (absolute) over standard self-play, with gains on both LM-error bugs ($+100$% relative / $+11$ pp absolute) and human-authored bugs ($+7.1$% relative / $+3.4$ pp absolute).","absKo":"Code repair는 language model (LM)의 중요한 능력이다. buggy program과 unit test가 주어졌을 때, LM은 테스트를 통과하는 수정된 program을 생성해야 한다. 우리는 unit test를 유일한 verifier로 사용하여, 제약 없는 edit를 가진 bug--fix task를 LM이 생성하도록 함으로써 code repair에 대한 supervision을 확장하고자 한다. 우리는 generator-fixer self-play를 제안하는데, 여기서 단일 model이 reinforcement learning으로 bug를 생성하고 이를 수정하는 역할을 번갈아 수행하도록 학습된다. fixer가 향상됨에 따라 generator는 점점 더 어려운 bug를 생성하도록 적응하며, 자동 curriculum을 형성한다. 그러나 unit test는 correctness는 검증하지만 realism은 보장하지 않으므로, generator가 실제에서 마주치는 bug에서 벗어나 자체 생성 bug에 대한 repair는 향상되지만 실제 bug에 대한 성능은 저하될 수 있음을 발견했다. 이를 해결하기 위해 Anchored Self Play (ASP)를 제안한다. ASP는 작은 reference set으로 self-play를 고정시키는데, (i) code-embedding similarity reward를 추가하여 generation을 유도하고, (ii) reference bug를 fixer training에 혼합하여 drift를 방지한다. 인간, LM, 그리고 LM 코드에 대한 인간의 편집에서 비롯된 bug를 모두 포함하는 code repair benchmark인 BugSourceBench를 도입하여, LM-assisted programming에서 bug의 출처를 반영한다. 다양한 bug source 전반에서 ASP는 최고의 fix rate를 달성하며, 표준 self-play 대비 평균 fix rate를 $+25$% (상대) / $+7.2$ pp (절대) 향상시킨다. 이 향상은 LM-error bug에서의 $+100$% 상대 / $+11$ pp 절대 향상과 human-authored bug에서의 $+7.1$% 상대 / $+3.4$ pp 절대 향상을 모두 포함한다."},{"id":"62424","en":"CAMEL: Confidence-Gated Reflection for Reward Modeling","ko":"CAMEL: Reward Modeling을 위한 Confidence-Gated Reflection","authors":"Zirui Zhu, Yong Liu, Yang Luo, Hailun Xu, Kanchan Sarkar, Kun Xu, Yang You","pos":"#2615","link":"https://openreview.net/forum?id=gl5xoOJJ22","abs":"

Reward models play a fundamental role in aligning large language models with human preferences. Existing methods predominantly follow two paradigms: scalar discriminative preference models, which are efficient but lack interpretability, and generative judging models, which offer richer reasoning at the cost of higher computational overhead. We observe that the log-probability margin between verdict tokens strongly correlates with prediction correctness, providing a reliable proxy for instance difficulty without additional inference cost. Building on this insight, we propose CAMEL, a confidence-gated reflection framework that performs a lightweight single-token preference decision first and selectively invokes reflection only for low-confidence instances. To induce effective self-correction, we train the model via reinforcement learning with counterfactual prefix augmentation, which exposes the model to diverse initial verdicts and encourages genuine revision. Empirically, CAMEL achieves state-of-the-art performance on three widely used reward-model benchmarks with 82.9\\% average accuracy, surpassing the best prior model by 3.2\\% and outperforming 70B-parameter models using only 14B parameters, while establishing a strictly better accuracy-efficiency Pareto frontier.

","absKo":"Reward model은 large language model을 human preference와 정렬하는 데 근본적인 역할을 한다. 기존 방법은 주로 두 가지 패러다임을 따른다: 효율적이지만 해석 가능성이 부족한 scalar discriminative preference model, 그리고 더 풍부한 reasoning을 제공하지만 계산 오버헤드가 높은 generative judging model이다. 우리는 verdict token 간 log-probability margin이 prediction correctness와 강하게 상관된다는 점을 관찰했으며, 이는 추가 inference cost 없이 instance difficulty를 신뢰성 있게 대변하는 proxy가 된다. 이 통찰을 바탕으로, 우리는 먼저 가벼운 single-token preference decision을 수행하고 confidence가 낮은 instance에 대해서만 선택적으로 reflection을 호출하는 confidence-gated reflection framework인 CAMEL을 제안한다. 효과적인 self-correction을 유도하기 위해, 우리는 counterfactual prefix augmentation을 사용한 reinforcement learning으로 model을 학습시키며, 이를 통해 model이 다양한 initial verdict를 접하고 진정한 revision을 하도록 장려한다. 실험적으로 CAMEL은 널리 사용되는 세 가지 reward-model benchmark에서 평균 82.9\\% accuracy로 state-of-the-art 성능을 달성했으며, 이전 최고 model보다 3.2\\%p 앞서고 14B parameter만으로 70B parameter model을 능가했으며, 더 나아가 strictly better한 accuracy-efficiency Pareto frontier를 확립했다."},{"id":"66140","en":"Training Prompt Matters: State-Adaptive Optimization for Robust Fine-Tuning","ko":"훈련 프롬프트가 중요하다: 견고한 Fine-Tuning을 위한 상태 적응형 최적화","authors":"Wenhang Shi, Yiren Chen, Shuqing Bian, Zhe Zhao, Pengfei Hu, Jinhao Dong, WEI LU, Xiaoyong Du","pos":"#2912","link":"https://openreview.net/forum?id=6XXz4OPAXA","abs":"

While prompt engineering is instrumental in maximizing the capabilities of Large Language Models (LLMs) during inference, the role of prompts during training remains critically underexplored. Prevailing fine-tuning paradigms typically treat training prompts as mere surface forms, assuming that semantically equivalent instructions yield identical learning outcomes. However, we reveal that this equivalence is deceptive: while paraphrased prompts often lead to comparable in-task performance, they induce drastically different cross-task impacts regarding catastrophic forgetting and generalization. Crucially, these impacts are positively correlated across tasks, indicating the existence of superior prompts that consistently yield better performance. Furthermore, we discover that these superior prompts can be robustly identified by task loss prior to learning. Leveraging these insights, we introduce State-Adaptive Prompt Optimization (SAPO), a lightweight yet effective training strategy that shifts task formulation from a static input to a dynamic, state-adaptive variable. Comprehensive experiments on diverse benchmarks confirm its effectiveness, which significantly mitigates forgetting while improving generalization, achieving substantial performance gains over state-of-the-art methods. These results provide insights into how training prompts shape learning dynamics and offer a practical recipe for robust fine-tuning.

","absKo":"

prompt engineering은 inference 동안 Large Language Models (LLMs)의 역량을 극대화하는 데 중요하지만, training 중 prompt의 역할은 여전히 충분히 탐구되지 않았다. 기존 fine-tuning 패러다임은 대체로 training prompt를 단순한 표면 형식으로 취급하며, 의미적으로 동등한 지시문은 동일한 학습 결과를 낳는다고 가정한다. 그러나 우리는 이러한 동등성이 착시임을 밝힌다. 바꿔 쓴 prompt는 종종 과제 내 성능은 비슷하게 만들지만, catastrophic forgetting과 generalization 측면에서는 과제 간 영향이 극적으로 다르다. 결정적으로, 이러한 영향은 과제 전반에서 양의 상관을 보이며, 이는 지속적으로 더 나은 성능을 내는 superior prompt가 존재함을 시사한다. 더 나아가 우리는 이러한 superior prompt가 학습 이전의 task loss만으로도 견고하게 식별될 수 있음을 발견했다. 이러한 통찰을 활용하여 우리는 State-Adaptive Prompt Optimization (SAPO)를 제안한다. 이는 task formulation을 정적 입력이 아니라 동적으로 상태에 적응하는 변수로 전환하는, 가볍지만 효과적인 training strategy다. 다양한 benchmark에 대한 포괄적 실험은 SAPO의 효과를 확인해 주며, forgetting을 크게 줄이는 동시에 generalization을 향상시켜 state-of-the-art method 대비 상당한 성능 향상을 달성한다. 이 결과는 training prompt가 학습 역학을 어떻게 형성하는지에 대한 통찰을 제공하며, 견고한 fine-tuning을 위한 실용적인 처방을 제시한다.

"},{"id":"61656","en":"Margin-Adaptive Confidence Ranking for Reliable LLM Judgement","ko":"신뢰할 수 있는 LLM 판단을 위한 Margin-Adaptive Confidence Ranking","authors":"Gaojie Jin, Yong Tao, Lijia Yu, Tianjin Huang","pos":"#3211","link":"https://openreview.net/forum?id=o3JP1sHHSY","abs":"

Jung et al. (2025) introduce a hypothesis testing framework for guaranteeing agreement between large language models (LLMs) and human judgments, relying on the assumption that the model’s estimated confidence is monotonic with respect to human-disagreement risk. In practice, however, this assumption may be violated, and the generalization behavior of the confidence estimator is not explicitly analyzed. We mitigate these issues by learning a dedicated confidence estimator instead of relying on heuristic confidence signals. Our approach leverages simulated annotator diversity and a margin-based ranking formulation to explicitly model how confidently an LLM distinguishes between human-agreement and human-disagreement cases. We further derive generalization guarantees for this estimator, revealing a margin-dependent trade-off that informs the design of an adaptive estimator training procedure. When integrated into fixed-sequence testing, the learned confidence estimator yields improved ranking accuracy and empirically strengthens the monotonic relationship between confidence and disagreement risk, leading to higher success rates in satisfying target agreement levels across multiple datasets and judge models.

","absKo":"

Jung et al. (2025)은 large language model(LLM)과 human judgment 간의 합의를 보장하기 위한 hypothesis testing framework를 제안했으며, 이는 모델의 추정 confidence가 human-disagreement risk에 대해 monotonic하다는 가정에 의존합니다. 그러나 실제로는 이 가정이 위배될 수 있고, confidence estimator의 generalization behavior는 명시적으로 분석되지 않았습니다. 우리는 heuristic confidence signal에 의존하는 대신 전용 confidence estimator를 학습함으로써 이러한 문제를 완화합니다. 우리의 접근은 simulated annotator diversity와 margin-based ranking formulation을 활용하여, LLM이 human-agreement 사례와 human-disagreement 사례를 얼마나 확신 있게 구별하는지 명시적으로 모델링합니다. 우리는 또한 이 estimator에 대한 generalization guarantee를 도출하여, margin에 의존하는 trade-off를 밝혀내고, 이는 adaptive estimator training procedure의 설계를 안내합니다. fixed-sequence testing에 통합했을 때, 학습된 confidence estimator는 더 나은 ranking accuracy를 제공하고 confidence와 disagreement risk 사이의 monotonic 관계를 실증적으로 강화하여, 여러 dataset과 judge model에서 목표 agreement level을 만족하는 success rate를 높입니다.

"},{"id":"64893","en":"Optimal Transport for Reward Modeling from Noisy Feedback","ko":"Noisy Feedback으로부터 Reward Modeling을 위한 Optimal Transport","authors":"Eric Wang, Licheng Pan, Haocheng Yang, Yunsheng Lu, Yongqi Tong, Yinuo Wang, Shijian Wang, Zhixuan Chu, Lei Shen, Haoxuan Li, Yuan Lu","pos":"#3711","link":"https://openreview.net/forum?id=InoePQ7HNI","abs":"

Reward models are fundamental to Reinforcement Learning from Human Feedback (RLHF), yet real-world datasets are inevitably corrupted by noisy preference. Conventional training objectives tend to overfit these errors, while existing denoising approaches often rely on homogeneous noise assumptions that fail to capture the complexity of linguistic preferences. To handle these challenges, we propose SelectiveRM, a framework grounded in optimal transport. We first devise a Joint Consistency Discrepancy to align the distribution of model predictions with preference data. Furthermore, to address the limitation of strict mass conservation which compels the model to fit outliers, we incorporate a Mass Relaxation mechanism via partial transport. This enables the autonomous exclusion of samples with noisy preference that contradict semantic consistency. Theoretically, we demonstrate that SelectiveRM optimizes a tighter upper bound on the unobserved clean risk. Extensive experiments validate that our approach significantly outperforms state-of-the-art baselines across diverse benchmarks. Code is available at https://anonymous.4open.science/r/SelectiveRM-33F1.

","absKo":"

Reward model은 Reinforcement Learning from Human Feedback (RLHF)의 핵심이지만, 실제 데이터셋은 필연적으로 noisy preference에 의해 오염된다. 전통적인 training objective는 이러한 오류에 과적합하는 경향이 있으며, 기존 denoising 접근법은 언어적 preference의 복잡성을 포착하지 못하는 homogeneous noise 가정에 의존하는 경우가 많다. 이러한 문제를 다루기 위해, 우리는 optimal transport에 기반한 framework인 SelectiveRM을 제안한다. 먼저 Joint Consistency Discrepancy를 설계해 model prediction의 분포를 preference data와 정렬한다. 더 나아가, 모델이 outlier에 맞추도록 강제하는 strict mass conservation의 한계를 해결하기 위해 partial transport를 통한 Mass Relaxation mechanism을 도입한다. 이를 통해 semantic consistency와 모순되는 noisy preference sample을 자율적으로 배제할 수 있다. 이론적으로, SelectiveRM이 관측되지 않은 clean risk에 대한 더 타이트한 upper bound를 최적화함을 보인다. 광범위한 실험은 우리의 방법이 다양한 benchmark에서 state-of-the-art baseline을 크게 능가함을 검증한다. 코드: https://anonymous.4open.science/r/SelectiveRM-33F1.

"},{"id":"66660","en":"Confidence is Not Universal: Task-Dependent Calibration and Emergent Behavior in LLMs","ko":"신뢰도는 보편적이지 않다: LLM에서의 태스크 의존적 Calibration과 창발적 행동","authors":"Chaeyun Jang, Moonseok Choi, Yegon Kim, Seungyoo Lee, Juho Lee, Hyungi Lee","pos":"#3900","link":"https://openreview.net/forum?id=1dvDUYc8nm","abs":"

Large language models (LLMs) increasingly support human decision-making, rendering human-interpretable confidence essential. However, it remains unclear whether verbalized confidence calibration generalizes across heterogeneous tasks without degrading accuracy. We show that universal confidence calibration fails. Across diverse benchmarks, we identify two incompatible task families with distinct confidence semantics. In reasoning-centric tasks, confidence supervision transfers within the family, often improving calibration while preserving or even improving accuracy, and induces emergent behaviors such as confidence-dependent reasoning length and self-verification. Retrieval- and copy-oriented tasks also exhibit within-family transfer, but fail to generalize to reasoning tasks, with cross-family supervision degrading both calibration and accuracy. Motivated by this finding, we disentangle confidence into reasoning uncertainty and evidence localization uncertainty. This simple decomposition restores cross-family generalization using supervised fine-tuning alone, suggesting that effective confidence alignment requires task-aware semantics rather than a universal scalar notion.

","absKo":"

대규모 언어 모델(LLM)은 점점 더 인간의 의사결정을 지원하고 있으며, 따라서 사람이 해석할 수 있는 confidence는 필수적이다. 그러나 verbalized confidence calibration이 정확도를 저해하지 않으면서 이질적인 task 전반으로 일반화되는지는 아직 분명하지 않다. 우리는 universal confidence calibration이 실패함을 보인다. 다양한 benchmark 전반에서, 우리는 confidence semantics가 서로 다른 두 개의 양립할 수 없는 task family를 식별한다. reasoning 중심 task에서는 confidence supervision이 같은 family 내에서는 전이되며, 종종 정확도를 유지하거나 오히려 향상시키면서 calibration을 개선하고, confidence-dependent reasoning length 및 self-verification과 같은 emergent behavior를 유도한다. retrieval 및 copy 지향 task 역시 같은 family 내에서는 전이를 보이지만 reasoning task로는 일반화되지 못하며, cross-family supervision은 calibration과 정확도 모두를 저하시킨다. 이러한 발견에 동기부여를 받아, 우리는 confidence를 reasoning uncertainty와 evidence localization uncertainty로 분해한다. 이 단순한 분해는 supervised fine-tuning만으로 cross-family generalization을 회복시키며, 효과적인 confidence alignment에는 보편적인 단일 scalar 개념이 아니라 task-aware semantics가 필요함을 시사한다.

"},{"id":"63350","en":"Dynamic Multimodal Evaluation via Knowledge-Enhanced Benchmark Evolution","ko":"지식 강화 벤치마크 진화를 통한 동적 멀티모달 평가","authors":"Junzhe Zhang, Huixuan Zhang, Xiaojun Wan","pos":"#4105","link":"https://openreview.net/forum?id=XrqMNac6Ft","abs":"

The rapid development of multimodal large language models (MLLMs) has created an urgent demand for more reliable and robust evaluation protocols. However, existing static benchmarks are prone to data contamination and performance saturation, which can result in inflated or misleading evaluation results. To address these limitations, we first introduce a graph formulation to represent both static and dynamic visual question answering (VQA) samples. Building upon this frmulation, we propose Knowledge-Enhanced Benchmark Evolution (KBE), a dynamic multimodal evaluation framework that first analyzes the original static benchmark, then expands it by integrating multimodal knowledge, transforming the static benchmark into a controllable, dynamic evolving version. Crucially, KBE can both reconstruct questions by Re-selecting visual information in the original image and expand existing questions with external textual knowledge. By explicitly controlling the degree of question exploration, KBE enables difficulty-controllable evaluation across a wide range of model capabilities. Extensive experimental results demonstrate that KBE effectively mitigates data contamination and benchmark saturation, while providing a more comprehensive and flexible assessment of MLLM performance.

","absKo":"

multimodal large language models (MLLMs)의 빠른 발전은 더욱 신뢰할 수 있고 robust한 evaluation protocol에 대한 긴급한 요구를 만들어냈다. 그러나 기존의 static benchmark는 data contamination과 performance saturation에 취약하여, 과장되거나 오해를 부를 수 있는 evaluation result를 초래할 수 있다. 이러한 한계를 해결하기 위해, 우리는 먼저 static 및 dynamic visual question answering (VQA) sample을 모두 표현하는 graph formulation을 도입한다. 이 formulation을 바탕으로, 우리는 Knowledge-Enhanced Benchmark Evolution (KBE)을 제안한다. 이는 원래 static benchmark를 먼저 분석한 뒤, multimodal knowledge를 통합하여 이를 controllable하고 dynamic하게 진화하는 버전으로 변환하는 dynamic multimodal evaluation framework이다. 특히 KBE는 원본 image에서 visual information을 Re-select하여 question을 재구성할 수도 있고, external textual knowledge를 통해 기존 question을 확장할 수도 있다. question exploration의 정도를 명시적으로 제어함으로써, KBE는 다양한 model capability 전반에 걸쳐 difficulty-controllable evaluation을 가능하게 한다. 광범위한 실험 결과는 KBE가 data contamination과 benchmark saturation을 효과적으로 완화하는 동시에, MLLM performance에 대해 더 포괄적이고 유연한 평가를 제공함을 보여준다.

"},{"id":"61159","en":"Zero Sum SVD: Balancing Loss Sensitivity for Low Rank LLM Compression","ko":"Zero Sum SVD: Low Rank LLM Compression을 위한 Loss Sensitivity 균형화","authors":"Ali Abbasi, Chayne Thrash, Haoran Qin, Shansita Sharma, Sepehr Seifi, Soheil Kolouri","pos":"#1702","link":"https://openreview.net/forum?id=tAdsDOkRBw","abs":"

Advances in large language models have driven strong performance across many tasks, but their memory and compute costs still hinder deployment. SVD-based compression reduces storage and can speed up inference via low-rank factors, yet performance depends on how rank is allocated under a global compression ratio. Prior methods often use homogeneous ranks for similarly sized matrices, despite large differences in loss sensitivity, or rely on expensive iterative pre-truncation optimization to determine per matrix ranks. We propose Zero Sum SVD (ZS-SVD), a post-training method that performs global singular component selection using activation whitening and first-order calibration loss estimates in whitened coordinates. ZS-SVD prunes components across the whole model with a zero sum rule that keeps the cumulative predicted loss change near zero, automatically yielding heterogeneous ranks without solving a rank allocation optimization. Motivated by evidence that gradients near pretrained solutions exhibit low rank structure, we also introduce an optional lightweight correction that applies a single projected gradient update after truncation, followed by re-truncation. Extensive experiments across multiple LLM architectures show consistent gains across diverse benchmarks and compression ratios.

","absKo":"

large language models의 발전은 다양한 task에서 강력한 성능을 이끌어냈지만, memory와 compute 비용은 여전히 배포를 가로막는다. SVD 기반 compression은 storage를 줄이고 low-rank factor를 통해 inference를 가속할 수 있지만, 성능은 global compression ratio 아래에서 rank를 어떻게 배분하느냐에 따라 달라진다. 기존 방법들은 비슷한 크기의 matrix에 대해 대체로 homogeneous rank를 사용하곤 하는데, loss sensitivity에는 큰 차이가 있음에도 그러하며, 혹은 matrix별 rank를 정하기 위해 비용이 큰 iterative pre-truncation optimization에 의존한다. 우리는 Zero Sum SVD (ZS-SVD)를 제안한다. 이는 activation whitening과 whitened coordinate에서의 first-order calibration loss estimate를 사용해 global singular component selection을 수행하는 post-training 방법이다. ZS-SVD는 전체 model에 걸쳐 component를 제거하면서 누적된 예측 loss change가 0에 가깝게 유지되도록 하는 zero sum 규칙을 사용하여, rank allocation optimization을 풀지 않고도 자동으로 heterogeneous rank를 산출한다. pretrained solution 주변의 gradient가 low rank structure를 보인다는 증거에 착안하여, 우리는 truncation 후 single projected gradient update를 적용한 다음 다시 re-truncation하는 선택적 경량 correction도 도입한다. 여러 LLM architecture와 다양한 benchmark 및 compression ratio에 걸친 광범위한 실험은 일관된 성능 향상을 보여준다.

"},{"id":"65453","en":"WarmServe: Enabling One-for-Many GPU Prewarming for Multi-LLM Serving","ko":"WarmServe: 다중 LLM 서빙을 위한 One-for-Many GPU 사전 예열 지원","authors":"Chiheng Lou, Sheng Qi, Rui Kang, Yong Zhang, Chen Sun, pengcheng wang, Xuanzhe Liu, Xin Jin","pos":"#1708","link":"https://openreview.net/forum?id=DVHpvumD60","abs":"Deploying multiple models within shared GPU clusters is a key strategy to improve resource efficiency in large language model (LLM) serving. Existing multi-LLM serving systems improve GPU utilization at the cost of degraded inference performance, particularly time-to-first-token (TTFT). We attribute this degradation to the lack of awareness regarding future workload characteristics. In contrast, recent analyses have shown the strong periodicity and long-term predictability of real-world LLM serving workloads. In this paper, we propose *one-for-many GPU prewarming*, which proactively loads parameters from multiple models onto GPUs based on workload forecasts. These prewarmed weights enable the system to promptly instantiate serving instances upon encountering request bursts. We design and implement WarmServe, a multi-LLM serving system incorporating three key techniques: (1) a model placement algorithm that optimizes prewarming decisions to minimize cross-model prewarming interference, (2) a KV cache reservation strategy that repurposes idle KV cache space on running GPUs for prewarming new models, and (3) an efficient GPU memory switching mechanism for tensor management. Evaluation on real-world datasets shows that WarmServe reduces TTFT by up to 50.8$\\times$ compared to the state-of-the-art autoscaling-based system, while supporting up to 2.5$\\times$ higher request throughput than the GPU-sharing system.","absKo":"공유 GPU cluster에서 여러 model을 함께 배포하는 것은 large language model(LLM) serving의 자원 효율을 높이는 핵심 전략이다. 기존 multi-LLM serving system은 GPU 활용률을 높이는 대신 inference performance, 특히 time-to-first-token(TTFT)을 저하시킨다. 우리는 이러한 저하가 미래 workload 특성에 대한 인식 부족에서 비롯된다고 본다. 반면, 최근 분석은 실제 LLM serving workload가 강한 주기성과 장기 예측 가능성을 보인다는 점을 보여주었다. 본 논문에서는 workload forecast를 바탕으로 여러 model의 parameters를 GPU에 선제적으로 적재하는 *one-for-many GPU prewarming*을 제안한다. 이렇게 prewarm된 weight는 request burst가 발생했을 때 시스템이 serving instance를 즉시 인스턴스화할 수 있게 한다. 우리는 WarmServe를 설계하고 구현했으며, 이는 세 가지 핵심 기법을 포함하는 multi-LLM serving system이다: (1) cross-model prewarming interference를 최소화하도록 prewarming 결정을 최적화하는 model placement algorithm, (2) 실행 중인 GPU의 유휴 KV cache 공간을 새 model prewarming에 재활용하는 KV cache reservation strategy, (3) tensor 관리를 위한 효율적인 GPU memory switching mechanism. 실제 데이터셋 평가에서 WarmServe는 state-of-the-art autoscaling-based system과 비교해 TTFT를 최대 50.8$\\times$ 줄이면서, GPU-sharing system보다 최대 2.5$\\times$ 높은 request throughput을 지원한다."},{"id":"62373","en":"WAVE: Window-Aware Vocabulary-Efficient Early-Exit for Training-Free LLM Acceleration","ko":"WAVE: Training-Free LLM 가속화를 위한 Window-Aware Vocabulary-Efficient Early-Exit","authors":"Seonggeun Kim, Gilha lee, Hyun Kim","pos":"#1709","link":"https://openreview.net/forum?id=hCY1J253Kf","abs":"Large language models (LLMs) incur substantial inference latency due to autoregressive decoding, in which each token requires a full forward pass through all transformer layers. Early-exit methods that terminate computation at intermediate layers offer a promising remedy, yet existing approaches suffer from fundamental limitations. Confidence-based methods rely on evaluating the full LM head at every layer, introducing considerable overhead that can negate the expected speedup. Schedule-based methods avoid this cost through predetermined exit schedules, but their monotonically decreasing layer allocation collapses to shallow layers, thereby constraining the maximum generation length. Learned exit predictors further require costly task-specific training and are vulnerable to distribution shifts in unseen domains. We propose Window-Aware Vocabulary-Efficient Early-Exit (WAVE), a training-free framework that addresses these challenges through two key innovations. First, exit window scheduling identifies an optimal layer range for early-exit decisions via offline calibration, preventing premature convergence to shallow layers while substantially reducing the number of exit checks. Second, a proxy LM head constructs a lightweight vocabulary subset at the window’s starting layer, reducing per-layer exit overhead by 87\\% relative to full LM head. WAVE requires no gradient-based training and enables immediate deployment with only a brief calibration phase. Experiments on Llama-2 7B demonstrate up to 1.4$\\times$ average speedup while preserving output quality, with full compatibility with W4A16 quantization, establishing WAVE as a practical early-exit framework for accelerating LLMs inference without retraining.","absKo":"Large language models (LLMs)는 각 토큰이 모든 transformer layer를 통과하는 전체 forward pass를 필요로 하는 autoregressive decoding 때문에 상당한 inference latency를 초래한다. 중간 layer에서 계산을 종료하는 early-exit 방법은 유망한 해결책이지만, 기존 접근법에는 근본적인 한계가 있다. confidence-based 방법은 각 layer마다 전체 LM head를 평가해야 하므로 상당한 오버헤드가 발생해 기대했던 speedup을 상쇄할 수 있다. schedule-based 방법은 미리 정해진 exit schedule을 통해 이 비용을 피하지만, layer allocation이 단조롭게 감소하여 얕은 layer에 수렴함으로써 최대 생성 길이를 제한한다. learned exit predictor는 추가로 비용이 큰 task-specific training을 요구하고, 보지 못한 도메인에서 distribution shift에 취약하다. 우리는 이러한 문제를 해결하는 training-free framework인 Window-Aware Vocabulary-Efficient Early-Exit (WAVE)를 제안한다. WAVE는 두 가지 핵심 혁신을 통해 동작한다. 첫째, exit window scheduling은 offline calibration을 통해 early-exit 결정을 위한 최적의 layer 범위를 식별하여, 얕은 layer로의 조기 수렴을 막으면서 exit check 수를 크게 줄인다. 둘째, proxy LM head는 window의 시작 layer에서 경량 vocabulary subset을 구성하여, full LM head 대비 layer당 exit overhead를 87\\% 줄인다. WAVE는 gradient-based training이 전혀 필요 없으며, 짧은 calibration 단계만으로 즉시 배포할 수 있다. Llama-2 7B 실험에서는 출력 품질을 유지하면서 최대 1.4$\\times$의 평균 speedup을 보였고, W4A16 quantization과도 완전히 호환되어, retraining 없이 LLM inference를 가속하는 실용적인 early-exit framework로서 WAVE의 가능성을 입증한다."},{"id":"62303","en":"The Optimal Token Baseline: Variance Reduction for Long-Horizon LLM-RL","ko":"최적 토큰 베이스라인: 장기 지평선 LLM-RL에서 분산 감소","authors":"Yingru Li, Jiawei Xu, Ziniu Li, Jiacai Liu, Wei Liu, Yuxuan Tong, Longtao Zheng, Zhenghai Xue, Yaxiang Zhang, Tianle Cai, Ge Zhang, Qian Liu, Baoxiang Wang","pos":"#1808","link":"https://openreview.net/forum?id=hpxI4Mndt4","abs":"Reinforcement Learning for Large Language Models (LLMs) often suffers from training collapse in long-horizon tasks due to exploding gradient variance. To mitigate this, baseline is commonly introduced for advantage computation; however, traditional value models remain difficult to optimize, and standard group-based baselines overlook sequence heterogeneity. Although classic optimal baseline theory can achieve global variance reduction, it neglects token heterogeneity and requires prohibitive gradient-based computation. In this work, we derive the Optimal Token Baseline (OTB) from first principles, proving that gradient updates should be weighted inversely to their cumulative gradient norm. To ensure efficiency, we propose the Logit-Gradient Proxy that approximates the gradient norm using only forward-pass probabilities. Our method achieves training stability and matches the performance of large group sizes ($N=32$) with only $N=4$, reducing token consumption by over 65\\% across single-turn and tool-integrated reasoning tasks.","absKo":"Large Language Models (LLMs)에 대한 Reinforcement Learning은 exploding gradient variance 때문에 long-horizon tasks에서 training collapse를 자주 겪는다. 이를 완화하기 위해 advantage computation에 baseline이 흔히 도입되지만, 전통적인 value models는 최적화가 어렵고 standard group-based baselines는 sequence heterogeneity를 간과한다. classic optimal baseline theory는 전역 variance reduction을 달성할 수 있지만 token heterogeneity를 무시하며, gradient-based computation이 지나치게 비싸다. 본 연구에서는 원리부터 Optimal Token Baseline (OTB)을 유도하고, gradient updates는 누적 gradient norm에 반비례하도록 가중되어야 함을 증명한다. 효율성을 보장하기 위해, 우리는 forward-pass probabilities만을 사용해 gradient norm을 근사하는 Logit-Gradient Proxy를 제안한다. 우리의 방법은 training stability를 달성하며, single-turn 및 tool-integrated reasoning tasks 전반에서 token consumption을 65\\% 이상 줄이면서도 큰 group size($N=32$)의 성능을 $N=4$만으로도 맞춘다."},{"id":"60933","en":"SoftMoE: Soft Differentiable Routing for Mixture-of-Experts in LLMs","ko":"SoftMoE: LLM의 Mixture-of-Experts를 위한 Soft Differentiable Routing","authors":"Mikołaj Zasada, Łukasz Struski, Jacek Tabor, Marcin Kurdziel","pos":"#1900","link":"https://openreview.net/forum?id=vGTFOp3jLO","abs":"

Sparse Mixture-of-Experts (MoE) architectures enable scaling LLM parameters under a fixed inference budget by activating only a small subset of experts via top-k routing. While this preserves causality and suits autoregressive language models, the discrete top-k operator is not differentiable, forcing a fixed number of active experts per input and resulting in inefficient use of computation. We propose SoftMoE, which replaces discrete routing with a truncated soft top-k LapSum relaxation, allowing gradient-based optimization of expert routing. We further parameterize the mean number of active experts per layer and impose a global budget constraint, enabling the model to learn how to allocate expert capacity across layers. SoftMoE remains fully compatible with autoregressive modeling and achieves performance comparable to or better than sparse MoE on language modeling and downstream tasks, while activating significantly fewer experts. Notably, the learned allocation is highly non-uniform, with later layers activating more experts.

","absKo":"

Sparse Mixture-of-Experts(MoE) architecture는 top-k routing을 통해 소수의 expert만 활성화함으로써 고정된 inference budget 하에서 LLM parameter를 확장할 수 있게 한다. 이는 causality를 보존하고 autoregressive language model에 적합하지만, discrete top-k operator는 differentiable하지 않기 때문에 입력마다 활성 expert 수가 고정되고 계산 자원이 비효율적으로 사용된다. 우리는 discrete routing을 truncated soft top-k LapSum relaxation으로 대체하여 expert routing에 대한 gradient-based optimization을 가능하게 하는 SoftMoE를 제안한다. 또한 layer당 활성 expert의 평균 개수를 parameterize하고 global budget constraint를 부과함으로써, 모델이 layer 전반에 expert capacity를 어떻게 배분할지 학습할 수 있게 한다. SoftMoE는 autoregressive modeling과 완전히 호환되며, language modeling과 downstream task에서 sparse MoE와 동등하거나 더 나은 성능을 달성하면서도 훨씬 적은 수의 expert를 활성화한다. 특히, 학습된 배분은 매우 비균일하며, 후반 layer일수록 더 많은 expert를 활성화한다.

"},{"id":"63793","en":"Rethinking Pretraining Data Detection for LLMs: From Local to Global","ko":"LLM의 pretraining 데이터 탐지 재고: 로컬에서 글로벌로","authors":"Chenye Ke, Yan Zhuang, Zirui Liu, Qi Liu","pos":"#1914","link":"https://openreview.net/forum?id=ThlRUXqFZ0","abs":"

The advancements of Large Language Models (LLMs) are primarily attributed to massive pretraining data, which also introduces risks like privacy leakage and data contamination. Therefore, it is crucial to determine whether an LLM has been trained on a given target text. Existing detection methods primarily rely on local statistics of isolated tokens (e.g., those with the lowest probabilities), neglecting the probability dynamics during the token generation process. In this paper, we shift the detection paradigm from a local token to a global sequence perspective, grounded in the core intuition that memorized sequences exhibit volatility patterns distinct from those generated via inference. We propose Adaptive Entropic Convolutional Analysis (AECA), a framework that conceptualizes the probability sequence as a dynamic signal, integrating calibration with convolutional filtering to effectively capture memorization signals. Extensive experiments demonstrate that AECA surpasses state-of-the-art baseline methods, achieving an average AUC improvement of up to 1.5\\%, with its advantage being particularly pronounced in long-text scenarios.

","absKo":"

Large Language Models (LLMs)의 발전은 주로 막대한 pretraining data에 기인하지만, 이는 privacy leakage와 data contamination 같은 위험도 함께 초래한다. 따라서 LLM이 주어진 target text로 학습되었는지 판별하는 것은 매우 중요하다. 기존 detection 방법은 주로 고립된 token의 local statistic, 예를 들어 가장 낮은 probability를 갖는 token에 의존하며, token generation 과정에서의 probability dynamics는 간과한다. 본 논문에서는 detection paradigm을 local token 관점에서 global sequence 관점으로 전환한다. 이는 memorized sequence가 inference를 통해 생성된 sequence와 구별되는 volatility pattern을 보인다는 핵심 직관에 기반한다. 우리는 probability sequence를 동적인 signal로 개념화하고, calibration과 convolutional filtering을 결합하여 memorization signal을 효과적으로 포착하는 Adaptive Entropic Convolutional Analysis (AECA)를 제안한다. 광범위한 실험은 AECA가 state-of-the-art baseline을 능가하며, 평균 AUC를 최대 1.5\\%까지 향상시킨다는 것을 보여주고, 특히 장문 텍스트 시나리오에서 그 우위가 두드러진다.

"},{"id":"66638","en":"RelayCaching: Accelerating LLM Collaboration via Decoding KV Cache Reuse","ko":"RelayCaching: 디코딩 KV Cache 재사용을 통한 LLM 협업 가속","authors":"Yingsheng Geng, Yuchong Gao, Weihong Wu, Guyue Liu, Jiang liu","pos":"#1915","link":"https://openreview.net/forum?id=1tbhBSXcyX","abs":"The increasing complexity of AI tasks has shifted the paradigm from monolithic models toward multi-agent large language model (LLM) systems. However, these collaborative architectures introduce a critical bottleneck: redundant prefill computation for shared content generated by previous agents, which significantly increases KV cache memory usage and time-to-first-token (TTFT). While various KV cache methods have been proposed to mitigate prefill redundancy, they either fail to maintain accuracy on agent-generated outputs or exhibit low reuse rates due to rigid constraints. We present RelayCaching, a training-free inference method that directly reuses decoding phase KV caches from previous agents in subsequent prefill phases. Our key insight is that KV caches for identical content are highly consistent across phases, while prefix-induced deviations are sparse and localized within a limited range of layers and token positions. By selectively recomputing KV caches at these positions, RelayCaching preserves model accuracy with minimal overhead, yielding a superior accuracy–efficiency trade-off over existing methods. Experiments on diverse collaborative LLM tasks spanning mathematical reasoning, general knowledge, and code generation demonstrate that RelayCaching achieves over $80$\\% KV cache reuse, reduces TTFT by up to $4.7\\times$ compared to the standard pipeline, all with negligible accuracy degradation.","absKo":"AI task의 복잡성이 증가함에 따라 패러다임은 단일체 모델에서 multi-agent large language model (LLM) system으로 이동하고 있다. 그러나 이러한 협업 architecture는 중요한 병목을 도입한다. 이전 agent가 생성한 shared content에 대해 redundant prefill computation이 발생하여 KV cache memory usage와 time-to-first-token (TTFT)이 크게 증가한다. prefill redundancy를 완화하기 위해 다양한 KV cache 방법이 제안되었지만, 이들은 agent-generated output에서 정확도를 유지하지 못하거나 엄격한 제약 때문에 reuse rate가 낮다. 우리는 RelayCaching을 제안하는데, 이는 이전 agent의 decoding phase KV cache를 이후 prefill phase에서 직접 재사용하는 training-free inference method이다. 우리의 핵심 통찰은 동일한 content에 대한 KV cache가 phase 간 매우 일관적이라는 점이며, prefix에 의해 유도된 편차는 희소하고 제한된 범위의 layer와 token position에 국소화되어 있다는 점이다. 이러한 위치에서 KV cache를 선택적으로 recompute함으로써, RelayCaching은 최소한의 오버헤드로 model accuracy를 보존하며, 기존 방법보다 우수한 accuracy–efficiency trade-off를 제공한다. 수학적 추론, 일반 지식, 코드 생성에 걸친 다양한 협업 LLM task에서 수행한 실험은 RelayCaching이 80%를 넘는 KV cache reuse를 달성하고, 표준 pipeline 대비 TTFT를 최대 4.7× 줄이며, 정확도 저하는 거의 없음을 보여준다."},{"id":"60769","en":"ProjQ: Project-and-Quantize for Adapter-Aware LLM Compression","ko":"ProjQ: Adapter-Aware LLM Compression을 위한 Project-and-Quantize","authors":"Wenya Yu, Chao Zhang, Li Wang, Samson Lasaulce, Merouane DEBBAH","pos":"#2005","link":"https://openreview.net/forum?id=wqhWZhfaWK","abs":"Post-Training Quantization (PTQ) and Low-Rank Adaptation (LoRA) constitute the standard pipeline for efficient Large Language Model (LLM) deployment. However, applying them sequentially poses a problem: PTQ often leaves behind random noise that is spread out (across the model's weights) in a way LoRA can't easily fix, meaning that LoRA ends up wasting its limited capacity trying to fix uncorrectable noise instead of improving task performance. In this paper, we propose \\textbf{ProjQ}, a novel framework for constraining quantization noise to the low-rank manifold via orthogonal subspace projection. We derive an efficient alternating algorithm that shapes the quantization noise into a low-rank structure, effectively offloading dominant error components to the subsequent adapter while minimizing the residual error in the orthogonal \"uncorrectable\" subspace. Our theoretical analysis demonstrates that ProjQ preserves strictly greater model plasticity for downstream tasks compared to standard PTQ. Extensive experiments on LLaMA-2 and Qwen2.5 confirm that ProjQ consistently outperforms existing methods in both quantization error compensation and downstream task fine-tuning, achieving up to $2\\times$ lower evaluation loss for compensation and matching the performance of standard 4-bit baselines on language modeling tasks with only 3 bits. The code is available on \\url{https://github.com/yy9301/ProjQ}.","absKo":"Post-Training Quantization (PTQ)와 Low-Rank Adaptation (LoRA)는 efficient Large Language Model (LLM) deployment를 위한 표준 파이프라인을 이룬다. 그러나 이를 순차적으로 적용하면 문제가 생긴다. PTQ는 종종 모델의 weight 전반에 걸쳐 퍼져 있는 random noise를 남기는데, LoRA는 이를 쉽게 보정하지 못한다. 그 결과 LoRA는 제한된 capacity를 task performance 개선이 아니라 보정 불가능한 noise를 억지로 수정하는 데 낭비하게 된다. 본 논문에서는 orthogonal subspace projection을 통해 quantization noise를 low-rank manifold에 제약하는 새로운 프레임워크 \\textbf{ProjQ}를 제안한다. 우리는 quantization noise를 low-rank 구조로 형성하는 효율적인 alternating algorithm을 유도하며, 이를 통해 다음 adapter로 dominant error components를 효과적으로 넘기는 동시에 orthogonal한 \"uncorrectable\" subspace에서 residual error를 최소화한다. 이론적 분석은 ProjQ가 표준 PTQ에 비해 downstream task에 대해 strictly greater model plasticity를 보존함을 보여준다. LLaMA-2와 Qwen2.5에서의 광범위한 실험은 ProjQ가 quantization error compensation과 downstream task fine-tuning 모두에서 기존 방법을 일관되게 능가하며, compensation에서는 최대 $2\\times$ 낮은 evaluation loss를 달성하고 language modeling task에서 3 bits만으로 standard 4-bit baseline의 성능을 맞춘다는 점을 확인한다. 코드는 \\url{https://github.com/yy9301/ProjQ}에 공개되어 있다."},{"id":"62519","en":"One LR Doesn’t Fit All: Heavy-Tail Guided Layerwise Learning Rates for LLMs","ko":"One LR은 모든 곳에 적용되지 않는다: LLM을 위한 Heavy-Tail 기반 계층별 학습률","authors":"Di He, Songjun Tu, Keyu Wang, Lu Yin, Shiwei Liu","pos":"#2104","link":"https://openreview.net/forum?id=fs9KaJyhRO","abs":"

Learning rate configuration is a fundamental aspect of modern deep learning. The prevailing practice of applying a uniform learning rate across all layers overlooks the structural heterogeneity of Transformers, potentially limiting their effectiveness as the backbone of Large Language Models (LLMs). In this paper, we introduce \\textbf{Layerwise Learning Rate (LLR)}, an adaptive scheme that assigns distinct learning rates to individual Transformer layers. Our method is grounded in Heavy-Tailed Self-Regularization (HT-SR) theory, which characterizes the empirical spectral density (ESD) of weight correlation matrices to quantify heavy-tailedness. Layers with weaker heavy-tailedness are assigned larger learning rates to accelerate their training, while layers with stronger heavy-tailedness receive smaller learning rates. By tailoring learning rates in this manner, LLR promotes balanced training across layers, leading to faster convergence and improved generalization. Extensive experiments across architectures (from LLaMA to GPT-nano), optimizers (AdamW and Muon), and parameter scales (60M–1B) demonstrate that LLR achieves up to 1.5× training speedup and outperforms baselines, notably raising average zero-shot accuracy from 47.09% to 49.02%. A key advantage of LLR is its low tuning overhead: it transfers nearly optimal LR settings directly from the uniform baseline. Our code is submitted.

","absKo":"

learning rate 설정은 현대 deep learning의 근본적인 측면이다. 모든 layer에 동일한 learning rate를 적용하는 현재의 관행은 Transformer의 구조적 이질성을 간과하며, Large Language Models(LLMs)의 backbone으로서의 효과를 제한할 수 있다. 본 논문에서는 개별 Transformer layer에 서로 다른 learning rate를 할당하는 adaptive scheme인 \\textbf{Layerwise Learning Rate (LLR)}를 제안한다. 우리의 방법은 Heavy-Tailed Self-Regularization(HT-SR) 이론에 기반하며, 이는 weight correlation matrix의 empirical spectral density(ESD)를 특성화하여 heavy-tailedness를 정량화한다. heavy-tailedness가 약한 layer에는 더 큰 learning rate를 부여해 학습을 가속하고, heavy-tailedness가 강한 layer에는 더 작은 learning rate를 부여한다. 이처럼 learning rate를 세밀하게 조정함으로써, LLR은 layer 전반의 균형 잡힌 학습을 촉진하여 더 빠른 수렴과 향상된 generalization을 이끈다. LLaMA부터 GPT-nano까지의 architecture, AdamW와 Muon의 optimizer, 그리고 60M–1B parameter scale 전반에 걸친 광범위한 실험은 LLR이 최대 1.5×의 training speedup을 달성하고 baseline을 능가하며, 특히 average zero-shot accuracy를 47.09%에서 49.02%로 끌어올림을 보여준다. LLR의 핵심 장점은 낮은 tuning overhead로, 거의 최적에 가까운 LR 설정을 uniform baseline에서 직접 전이할 수 있다는 점이다. 우리의 코드는 제출되었다.

"},{"id":"65926","en":"Olmix: A Framework for Data Mixing Throughout LM Development","ko":"Olmix: 언어 모델 개발 전반의 데이터 혼합 프레임워크","authors":"Mayee Chen, Tyler Murray, David Heineman, Matt Jordan, Hannaneh Hajishirzi, Christopher Re, Luca Soldaini, Kyle Lo","pos":"#2108","link":"https://openreview.net/forum?id=8pOl3azhbL","abs":"

Data mixing---determining the ratios of data from different domains---is a first-order concern for training language models (LMs), but existing mixing methods have poorly understood design choices and assume that the set of domains remain fixed throughout development. We present Olmix, a framework that addresses two challenges encountered during LM development. First, the configuration space for developing a mixing method is not well understood---design choices across existing methods lack justification or consensus and overlook practical issues like data constraints. We conduct a comprehensive empirical study of this space, identifying which design choices lead to a strong mixing method. Second, the domain set evolves throughout LM development as datasets are revised and expanded---a problem setting largely unaddressed by existing works. We study how to efficiently recompute the mixture after the domain set is updated, given an existing mix from before the update. We introduce mixture reuse, a mechanism that reuses existing relative ratios and recomputes ratios only for domains affected by an update. Over a sequence of five domain-set updates mirroring real-world LM development, mixture reuse matches the performance of fully recomputing the mix after each update with 74% less compute and improves over training without mixing by 11.6% on downstream tasks.

","absKo":"data mixing, 즉 서로 다른 domain의 data 비율을 결정하는 일은 language model(LM) 학습에서 일차적으로 중요한 문제지만, 기존 mixing 방법들은 설계 선택에 대한 이해가 부족하고 개발 기간 동안 domain 집합이 고정되어 있다고 가정한다. 우리는 LM 개발 중 마주치는 두 가지 문제를 해결하는 프레임워크 Olmix를 제시한다. 첫째, mixing method를 개발하기 위한 configuration space는 잘 이해되어 있지 않다. 기존 방법들 사이의 설계 선택은 정당화나 합의가 부족하며, data constraint와 같은 실제적인 문제를 간과한다. 우리는 이 공간에 대한 포괄적인 실증 연구를 수행하여, 어떤 설계 선택이 강력한 mixing method로 이어지는지 식별한다. 둘째, domain 집합은 dataset이 개정되고 확장됨에 따라 LM 개발 과정에서 진화한다. 이는 기존 연구가 대체로 다루지 않은 문제 설정이다. 우리는 domain set이 업데이트된 뒤, 업데이트 이전의 기존 mix가 주어졌을 때 mixture를 얼마나 효율적으로 재계산할 수 있는지 연구한다. 우리는 mixture reuse를 도입하는데, 이는 기존 상대 비율을 재사용하고 업데이트의 영향을 받는 domain에 대해서만 비율을 다시 계산하는 메커니즘이다. 실제 LM 개발을 모사한 다섯 번의 domain-set update 연속 과정에서, mixture reuse는 각 업데이트마다 mix를 완전히 재계산하는 것과 동일한 성능을 74% 더 적은 compute로 달성하며, downstream task에서 mixing 없이 학습하는 것보다 11.6% 향상된다.

"},{"id":"63577","en":"LEAP: Zone-Aware MCTS for LLM Self-Speculative Decoding","ko":"LEAP: LLM 자기 추측 디코딩을 위한 zone-aware MCTS","authors":"LeiQuan Zheng, Yuan Liu","pos":"#2205","link":"https://openreview.net/forum?id=VrI9sWkJB1","abs":"Self-speculative decoding accelerates LLM inference by using a lightweight draft model for generation and a target model for verification, where the draft model is constructed by a subset of the target model’s layers, and the key challenge lies in layer configuration strategies. To address this challenge, we propose LEAP, a plug-and-play approach that formulates and optimizes the draft model construction problem as a sequential decision-making process by Monte Carlo Tree Search (MCTS). To navigate the prohibitive search space of deep LLMs, we leverage two empirical observations: (i) the prefilling-derived redundancy information remains informative during decoding, and (ii) the layer redundancy exhibits zone-wise characteristics. These observations enable a structured search space through zone partitioning and layer grouping, which serves as an inductive bias to facilitate efficiency of MCTS. Experimental results show that LEAP achieves a speedup of $1.7\\times\\sim2.0\\times$ for LLM inference.","absKo":"Self-speculative decoding은 generation에 lightweight draft model을, verification에 target model을 사용하여 LLM inference를 가속하는데, draft model은 target model의 layer 일부로 구성되며 핵심 과제는 layer configuration strategy에 있다. 이 과제를 해결하기 위해 우리는 LEAP을 제안한다. LEAP은 draft model 구성 문제를 Monte Carlo Tree Search(MCTS)를 통해 sequential decision-making process로 정식화하고 최적화하는 plug-and-play 접근법이다. 깊은 LLM의 막대한 search space를 탐색하기 위해 우리는 두 가지 경험적 관찰을 활용한다. (i) prefilling에서 유도된 redundancy information은 decoding 동안에도 유의미하고, (ii) layer redundancy는 zone-wise한 특성을 보인다는 점이다. 이러한 관찰은 zone partitioning과 layer grouping을 통해 structured search space를 가능하게 하며, 이는 MCTS의 효율성을 높이는 inductive bias로 작동한다. 실험 결과 LEAP은 LLM inference에서 $1.7\\times\\sim2.0\\times$의 speedup을 달성한다."},{"id":"61654","en":"Long-Context Modeling with Dynamic Hierarchical Sparse Attention for Memory-Constrained LLM Inference","ko":"메모리 제약 LLM Inference를 위한 Dynamic Hierarchical Sparse Attention 기반 Long-Context Modeling","authors":"Siheng Xiong, Joe Zou, Faramarz Fekri, Yae Jee Cho","pos":"#2209","link":"https://openreview.net/forum?id=o3gN27ITWV","abs":"The quadratic cost of attention limits the scalability of long-context LLMs, especially under limited hardware memory budgets. While attention is often sparse, existing static sparse methods cannot adapt to task- or input-dependent variations, and recent dynamic approaches rely on predefined templates or heuristics that may sacrifice generality. We propose Dynamic Hierarchical Sparse Attention (DHSA), a data-driven framework that predicts attention sparsity online while keeping the LLM backbone frozen. DHSA performs hierarchical routing by estimating importance at the chunk level and propagating it to token-level interactions, preserving causally important dependencies while enabling efficient sparsification. Across Needle-in-a-Haystack and LongBench, DHSA maintains near-dense accuracy in highly sparse regimes, achieving 12-20% relative accuracy gains over Block Sparse Attention at comparable prefill cost. With a memory-efficient tiled backend, DHSA delivers up to $10\\times$ prefill speedup at 128K context length. On LLaMA-3.1-8B (4-bit), DHSA scales to 100K context on a single 24GB GPU, where dense attention fails. We provide complementary GPU and CPU backends, enabling DHSA to run across diverse hardware environments and multiple open-weight model families. These results demonstrate DHSA as an efficient and adaptable solution for memory-constrained long-context LLM inference.","absKo":"attention의 quadratic cost는 long-context LLM의 확장성을 제한하며, 특히 제한된 hardware memory budget 하에서 더욱 그렇습니다. attention은 종종 sparse하지만, 기존의 static sparse method는 task 또는 input에 따른 변화를 적응적으로 반영할 수 없고, 최근의 dynamic 접근은 일반성을 희생할 수 있는 사전 정의된 template이나 heuristic에 의존합니다. 우리는 LLM backbone을 freeze한 채 attention sparsity를 온라인으로 예측하는 data-driven framework인 Dynamic Hierarchical Sparse Attention(DHSA)을 제안합니다. DHSA는 chunk 수준에서 중요도를 추정하고 이를 token-level interaction으로 전파하는 hierarchical routing을 수행함으로써, 인과적으로 중요한 dependency를 보존하면서 효율적인 sparsification을 가능하게 합니다. Needle-in-a-Haystack과 LongBench 전반에서 DHSA는 매우 sparse한 regime에서도 dense에 가까운 정확도를 유지하며, 비슷한 prefill cost에서 Block Sparse Attention 대비 12-20%의 상대 정확도 향상을 달성합니다. memory-efficient tiled backend와 결합하면, DHSA는 128K context length에서 최대 10배의 prefill speedup을 제공합니다. LLaMA-3.1-8B(4-bit)에서는 DHSA가 단일 24GB GPU에서 100K context까지 확장되며, dense attention은 실패합니다. 우리는 상보적인 GPU 및 CPU backend를 제공하여, 다양한 hardware 환경과 여러 open-weight model family에서 DHSA를 실행할 수 있게 합니다. 이러한 결과는 DHSA가 memory-constrained long-context LLM inference를 위한 효율적이고 적응적인 해법임을 보여줍니다."},{"id":"66437","en":"MechVQA: Benchmarking and Enhancing Multimodal LLMs on Comprehensive Mechanical Drawing Understanding","ko":"MechVQA: 종합적 기계 도면 이해에 대한 멀티모달 LLM 벤치마킹 및 향상","authors":"Qian Kou, Xiaofeng Shi, Yulin Li, Xiaosong Qiu, XinyangWang, Hua Zhou, Cao Dongxing","pos":"#2212","link":"https://openreview.net/forum?id=3jsh1kSoYk","abs":"

Multimodal Large Language Models (MLLMs) have demonstrated significant achievements in general visual question answering (VQA) tasks. However, they remain brittle on mechanical engineering drawings, where high annotation density and weak domain knowledge, compounded by unreliable spatial relation reasoning under strict projection rules and geometric constraints, make decisive cues easy to miss and frequently lead to wrong answers. To bridge this gap, we introduce the first comprehensive mechanical drawing understading dataset MechVQA created through an semi-automated construction and quality-control pipeline. MechVQA contains 3.3k high-density pictures with 21K question–answer pairs, spanning 10 different fine-grained tasks across three capability levels: Recognition, Reasoning, and Judging, providing a testbed to evaluate and improve MLLMs understanding on real world mechanical drawings. On top of MechVQA, we then develop the MechVL model through a multi-stage training paradigm, building a strong domain specialized baseline. Extensive experimental results demonstrate that MechVL outperforms strong closed-source MLLMs by 6\\% on MechVQA total score, significantly enhancing mechanical drawing understanding ability and providing a reusable foundation for deploying MLLMs in mechanical design and inspection scenarios.

","absKo":"

Multimodal Large Language Model (MLLM)은 일반 visual question answering (VQA) task에서 큰 성과를 보여 왔다. 그러나 기계공학 도면에서는 annotation density가 높고 domain knowledge가 약하며, 엄격한 projection rule와 geometric constraint 아래에서 신뢰할 수 없는 spatial relation reasoning이 겹치면서 결정적인 단서를 놓치기 쉬워 오답으로 이어지는 경우가 많다. 이 간극을 메우기 위해, 우리는 semi-automated construction 및 quality-control pipeline을 통해 구축한 최초의 종합적인 기계 도면 이해 데이터셋 MechVQA를 소개한다. MechVQA는 3개 capability level, 즉 Recognition, Reasoning, Judging 전반에 걸친 10개의 세분화된 task를 포함하는 3.3k장의 고밀도 이미지와 21K개의 question-answer pair로 구성되어 있으며, 실제 기계 도면에 대한 MLLM의 이해를 평가하고 개선하기 위한 testbed를 제공한다. 이어서 MechVQA 위에서 multi-stage training paradigm을 통해 MechVL model을 개발하여 강력한 domain-specialized baseline을 구축한다. 광범위한 실험 결과는 MechVL이 MechVQA total score에서 강력한 closed-source MLLM보다 6\\% 앞서며, 기계 도면 이해 능력을 크게 향상시키고 기계 설계 및 검사 시나리오에 MLLM을 배치하기 위한 재사용 가능한 기반을 제공함을 보여준다.

"},{"id":"66014","en":"MergeMix: Optimizing Mid-Training Data Mixtures via Learnable Model Merging","ko":"MergeMix: 학습 가능한 모델 병합을 통한 중간 학습 데이터 혼합 최적화","authors":"Jiapeng Wang, Changxin Tian, Kunlong Chen, ziqi liu, Jiaxin Mao, Xin Zhao, Zhiqiang Zhang, JUN ZHOU","pos":"#2213","link":"https://openreview.net/forum?id=7ucFlopjQu","abs":"Optimizing data mixtures is is essential for unlocking the full potential of of large language models (LLMs), yet identifying the optimal composition remains computationally prohibitive due to reliance on heuristic trials or expensive proxy training. To address this, we introduce MergeMix, a novel appraoch that efficiently determines optimal data mixing ratios by repurposing model merging weights as a high-fidelity, low-cost performance proxy. By training domain-specific experts on minimal tokens and optimizing their merging weights against downstream benchmarks, MergeMix effectively optimizes the performance of data mixtures without incurring the cost of full-scale training. Extensive experiments on models with 8B and 16B parameters validate that MergeMix achieves performance comparable to or surpassing exhaustive manual tuning while drastically reducing search costs. Furthermore, MergeMix exhibits high rank consistency (Spearman $\\rho > 0.9$) and strong cross-scale transferability, offering a scalable, automated solution for data mixture optimization.","absKo":"대규모 언어 모델(LLMs)의 잠재력을 최대한 끌어내기 위해서는 data mixture를 최적화하는 것이 필수적이지만, 최적의 구성을 찾는 일은 heuristic trial이나 비용이 큰 proxy training에 의존해야 하므로 계산적으로 매우 어렵습니다. 이를 해결하기 위해 우리는 모델 merging weights를 고충실도 저비용 성능 proxy로 재활용하여 최적의 data mixing ratio를 효율적으로 결정하는 새로운 접근법인 MergeMix를 소개합니다. 최소한의 token으로 domain-specific expert를 학습하고, downstream benchmark에 대해 이들의 merging weights를 최적화함으로써, MergeMix는 full-scale training 비용 없이 data mixture의 성능을 효과적으로 최적화합니다. 8B 및 16B parameter 모델에서 수행한 광범위한 실험은 MergeMix가 exhaustive manual tuning과 동등하거나 그 이상인 성능을 달성하면서 탐색 비용을 크게 줄임을 검증합니다. 또한 MergeMix는 높은 rank consistency(Spearman $\\rho > 0.9$)와 강한 cross-scale transferability를 보여, data mixture 최적화를 위한 확장 가능하고 자동화된 솔루션을 제공합니다."},{"id":"65242","en":"Mining Tensor/Neuron-Level Sparsity to Maximize Mixture-of-Experts Potential in Post-Training and Inference","ko":"Post-Training 및 추론에서 Mixture-of-Experts 잠재력을 최대화하기 위한 텐서/뉴런 수준 희소성 활용","authors":"Weilin Cai, Le Qin, Shwai He, Junwei Cui, Ang Li, Jiayi Huang","pos":"#2214","link":"https://openreview.net/forum?id=FPUFIOpZW9","abs":"Mixture of Experts (MoE) has emerged as a mainstream architecture for Large Language Models (LLMs), balancing computational efficiency with model scalability. While prior work has explored increasing tensor-level sparsity via finer-grained expert configurations during pre-training, we identify significant unexploited sparsity at both the tensor and neuron levels during post-training and inference. To leverage this, we propose complete expert partition for post-training and threshold-based token-expert dropping for inference. These techniques improve the Mixtral-8$\\times$7B model's average accuracy by 1% across nine downstream benchmarks (notably 4% on GSM8K). To further optimize the accuracy-efficiency trade-off for inference, we introduce dual-threshold token-expert dropping with partial expert partition and reconstruction. Our approach yields a 1.19$\\times$ MoE speedup and a 0.5% accuracy gain on Mixtral-8$\\times$7B when combining post-training and inference optimizations. For inference-only optimization on OLMoE-Instruct and DeepSeek-V2-Lite-Chat, we achieve up to 1.41$\\times$ MoE speedup with a negligible accuracy loss ($<$0.5%).","absKo":"Mixture of Experts (MoE)는 Large Language Models (LLMs)에서 주류 아키텍처로 부상했으며, 연산 효율성과 모델 확장성 사이의 균형을 맞춘다. 기존 연구는 pre-training 동안 더 세분화된 expert 구성으로 tensor-level sparsity를 높이는 것을 탐구해 왔지만, 우리는 post-training과 inference 단계에서 tensor 수준과 neuron 수준 모두에 상당한 미활용 sparsity가 존재함을 확인했다. 이를 활용하기 위해, 우리는 post-training을 위한 complete expert partition과 inference를 위한 threshold-based token-expert dropping을 제안한다. 이러한 기법들은 Mixtral-8$\\times$7B 모델의 average accuracy를 9개 downstream benchmark 전반에서 1% 향상시키며(특히 GSM8K에서 4% 향상), inference의 accuracy-efficiency trade-off를 추가로 최적화하기 위해 partial expert partition과 reconstruction을 결합한 dual-threshold token-expert dropping을 도입한다. 우리의 접근법은 post-training과 inference 최적화를 결합했을 때 Mixtral-8$\\times$7B에서 1.19$\\times$ MoE speedup과 0.5% accuracy 향상을 달성한다. OLMoE-Instruct와 DeepSeek-V2-Lite-Chat에 대한 inference-only 최적화에서는 거의 무시할 수 있는 accuracy 손실($<$0.5%)로 최대 1.41$\\times$ MoE speedup을 달성한다."},{"id":"66545","en":"Heterogeneous Customizable Personalized Federated Fine-Tuning Approach for Large Language Models","ko":"대규모 언어 모델을 위한 이질적 맞춤형 개인화 Federated Fine-Tuning 접근법","authors":"xin tong, Baojiang cui","pos":"#2306","link":"https://openreview.net/forum?id=2jdpJN9OmD","abs":"

Personalized federated LoRA fine tuning has become a key approach to addressing data heterogeneity in distributed fine tuning of large language models (LLMs). Existing methods typically assume homogeneous personalization needs across clients, relying on dual LoRA or personalized calibration schemes. However, they fail to account for the heterogeneity of local personalization requirements and the conflicting optimization objectives in dual LoRA, limiting scalability and performance. To address this, we propose Het-CPFLoRA, a customizable heterogeneous federated LoRA fine tuning algorithm inspired by the decoupling properties of LoRA parameters. We employ a single adapter fine tuning scheme to mitigate conflicts between personalized and generalized optimization, decouple LoRA into generalized and personalized subspaces for local customization, and use SVD compression to integrate cross client generalized knowledge. During inference, we introduce an OOD oriented dynamic mechanism to adjust the weighting between personalized and generalized decoupling knowledge, improving performance on user data. Extensive experiments on two public benchmark datasets show that Het-CPFLoRA outperforms state of the art methods in both personalization and generalization across heterogeneous scenarios. The code will be released as an open-source project.

","absKo":"

Personalized federated LoRA fine tuning은 분산된 large language models (LLMs) fine tuning에서 data heterogeneity를 다루는 핵심 접근법이 되었다. 기존 방법은 일반적으로 모든 client에 걸쳐 동질적인 personalization 요구를 가정하고, dual LoRA 또는 personalized calibration scheme에 의존한다. 그러나 이러한 방법은 local personalization requirement의 heterogeneity와 dual LoRA에서의 상충하는 optimization objective를 충분히 반영하지 못해, 확장성과 성능을 제한한다. 이를 해결하기 위해, 우리는 LoRA parameter의 decoupling property에서 영감을 받은 customizable heterogeneous federated LoRA fine tuning algorithm인 Het-CPFLoRA를 제안한다. 우리는 personalized optimization과 generalized optimization 사이의 충돌을 완화하기 위해 single adapter fine tuning scheme을 사용하고, local customization을 위해 LoRA를 generalized subspace와 personalized subspace로 decouple하며, SVD compression을 사용해 client 간 generalized knowledge를 통합한다. 추론 시에는 OOD 지향 dynamic mechanism을 도입해 personalized knowledge와 generalized decoupling knowledge 사이의 weighting을 조정함으로써 user data에서의 성능을 향상시킨다. 두 개의 공개 benchmark dataset에 대한 광범위한 실험은 Het-CPFLoRA가 heterogeneous scenario 전반에서 personalization과 generalization 모두에서 state-of-the-art method를 능가함을 보여준다. 코드는 오픈소스 프로젝트로 공개될 예정이다.

"},{"id":"66251","en":"GEMQ: Global Expert-Level Mixed-Precision Quantization for MoE LLMs","ko":"GEMQ: MoE LLM을 위한 전역 전문가 수준 혼합 정밀도 양자화","authors":"Jianing Deng, Song Wang, Dongwei Wang, Zijie Liu, Tianlong Chen, Huanrui Yang, Jingtong Hu","pos":"#2308","link":"https://openreview.net/forum?id=5Q4ts0pr9T","abs":"

Mixture-of-Experts Large Language Models (MoE-LLMs) achieve strong performance but incur substantial memory overhead due to massive expert parameters. Mixed-precision quantization mitigates this cost by allocating expert-wise bit-widths based on their importance, approaching the accuracy-memory Pareto frontier and enabling extreme low-bit quantization. However, existing methods rely on layer-wise importance estimation and overlook router shifts induced by quantization, resulting in suboptimal allocation and routing. In this work, we propose Global Expert-level Mixed-precision Quantization (GEMQ) to overcome these limitations via (1) a global linear-programming formulation that captures model-wide expert importance based on quantization error analysis, and (2) efficient router fine-tuning to adapt routing to quantized experts. These components are integrated into a progressive quantization framework that iteratively refines importance estimation and allocation. Experiments demonstrate that GEMQ significantly reduces memory and accelerates inference with minimal accuracy degradation.

","absKo":"Mixture-of-Experts Large Language Model(MoE-LLM)은 강력한 성능을 달성하지만, 대규모 expert parameter로 인해 상당한 메모리 오버헤드를 수반한다. mixed-precision quantization은 expert별 중요도에 따라 bit-width를 할당함으로써 이러한 비용을 완화하며, accuracy-memory Pareto frontier에 접근하고 극단적인 low-bit quantization을 가능하게 한다. 그러나 기존 방법들은 layer-wise 중요도 추정에 의존하고 quantization으로 인한 router shift를 간과하여, 최적 이하의 할당과 routing을 초래한다. 본 연구에서는 이러한 한계를 극복하기 위해 Global Expert-level Mixed-precision Quantization(GEMQ)을 제안한다. 이는 (1) quantization error 분석을 바탕으로 model-wide expert 중요도를 포착하는 global linear programming formulation과, (2) quantized expert에 routing을 적응시키기 위한 효율적인 router fine-tuning으로 구성된다. 이러한 구성요소들은 중요도 추정과 할당을 반복적으로 정제하는 progressive quantization framework에 통합된다. 실험은 GEMQ가 정확도 저하를 최소화하면서 메모리를 크게 절감하고 inference를 가속함을 보여준다."},{"id":"61371","en":"GEM: Geometric Entropy Mixing for Optimal LLM Data Curation","ko":"GEM: 최적 LLM Data Curation을 위한 Geometric Entropy Mixing","authors":"Yue Min, Ziyun Qiao, Ruining Chen, Yujun Li","pos":"#2309","link":"https://openreview.net/forum?id=qubIJ4i1XQ","abs":"

LLM pre-training efficacy increasingly depends on data composition rather than sheer volume. Yet, optimal mixing is hindered by categorization flaws: human taxonomies suffer from ontological misalignment, and Euclidean clustering fails to address embedding anisotropy. We introduce GEM (Geometric Entropy Mixing), a framework reformulating data curation as a variational problem on the hypersphere augmented with a mixing-balance regularizer. By decoupling the generative prior and optimizing the objective via a provable MM (Minorize-Maximize) algorithm, GEM effectively counteracts the cluster collapse to discover balanced semantic structures invisible to Euclidean heuristics. We employ teacher-student distillation to scale this geometric fidelity to web-scale corpora and introduce the Geometric Influence Score (GIS) for interpretable taxonomy generation. Experiments with 1.1B-parameter models demonstrate that GEM establishes a new state-of-the-art when integrated into mixing strategies like DoReMi and RegMix, improving average downstream accuracy by up to 1.2% and offering a robust coordinate system for predictable data mixing.

","absKo":"

LLM pre-training의 효과는 점점 단순한 데이터 양보다 데이터 구성에 더 크게 좌우되고 있다. 그러나 최적의 mixing은 범주화의 결함 때문에 저해된다. 인간이 만든 taxonomy는 ontological misalignment를 겪고, Euclidean clustering은 embedding anisotropy를 해결하지 못한다. 우리는 GEM (Geometric Entropy Mixing)을 제안한다. 이는 데이터 큐레이션을 mixing-balance regularizer가 추가된 hypersphere 위의 변분 문제로 재정식화하는 프레임워크다. 생성 prior를 분리하고, 증명 가능한 MM (Minorize-Maximize) 알고리즘으로 목적함수를 최적화함으로써, GEM은 Euclidean heuristic으로는 보이지 않는 균형 잡힌 semantic structure를 발견하기 위해 cluster collapse에 효과적으로 대응한다. 우리는 teacher-student distillation을 사용해 이러한 geometric fidelity를 web-scale 코퍼스로 확장하고, 해석 가능한 taxonomy 생성을 위해 Geometric Influence Score (GIS)를 도입한다. 11억 파라미터 모델을 사용한 실험은 GEM이 DoReMi와 RegMix 같은 mixing 전략에 통합될 때 새로운 state-of-the-art를 확립하며, 평균 downstream 정확도를 최대 1.2%까지 향상시키고, 예측 가능한 data mixing을 위한 견고한 좌표계를 제공함을 보여준다.

"},{"id":"64143","en":"Enhancing Numerical Prediction in LLMs via Smooth MMD Alignment","ko":"부드러운 MMD 정렬을 통한 LLM의 수치 예측 향상","authors":"Zhuo Zuo, Li Yue, Wenhao Zheng, Chenpeng Wang, Xianggen Liu","pos":"#2406","link":"https://openreview.net/forum?id=QWkdixm455","abs":"

Despite their strong general capabilities, large language models (LLMs) often remain unreliable when outputs must be numerically precise. A key reason is the training objective: standard cross-entropy treats numeric tokens as unstructured categories and ignores the metric structure of their values. We address this mismatch by proposing Smooth Maximum Mean Discrepancy (SMMD), which builds on the classic MMD by incorporating value-distance kernels over numeric tokens and graph-based smoothness. With this kernel defined over a numeric sub-vocabulary, SMMD aligns the predicted numeric distribution to the target via kernel matching and smooths the prediction--target residual over the induced kernel graph to encourage local consistency. We evaluate SMMD on four numeric-target tasks---mathematical reasoning, arithmetic calculation, clock-time recognition, and chart question answering---across multiple open-weight LLM and VLM backbones. SMMD consistently improves accuracy over both cross-entropy and recent numeric-target losses; analyses show complementary effects between MMD and smoothness and underscore the importance of distance-based kernel design.

","absKo":"

강력한 일반 능력에도 불구하고, large language model(LLM)은 출력이 수치적으로 정확해야 할 때 여전히 신뢰성이 떨어지는 경우가 많다. 핵심 이유는 학습 목표에 있다. 표준 cross-entropy는 numeric token을 구조화되지 않은 범주로 취급하고 그 값의 metric structure를 무시한다. 우리는 numeric token 위의 value-distance kernel과 graph-based smoothness를 통합한 고전적 MMD 기반의 Smooth Maximum Mean Discrepancy(SMMD)를 제안하여 이러한 불일치를 해결한다. numeric sub-vocabulary 위에 정의된 이 kernel을 통해, SMMD는 kernel matching으로 예측된 numeric distribution을 target에 정렬하고, 유도된 kernel graph 위에서 prediction-target residual을 smooth하게 만들어 local consistency를 장려한다. 우리는 여러 open-weight LLM 및 VLM backbone에 걸쳐 수학적 추론, 산술 계산, 시각의 시계 인식, 차트 질문 응답의 네 가지 numeric-target task에서 SMMD를 평가한다. SMMD는 cross-entropy와 최근의 numeric-target loss 모두에 비해 일관되게 정확도를 향상시키며, 분석 결과 MMD와 smoothness의 상호보완적 효과와 distance-based kernel design의 중요성을 강조한다.

"},{"id":"64365","en":"Context Distillation Retains Post-Training Capabilities in Continually Trained LMs","ko":"컨텍스트 증류는 지속 학습된 LM에서 사후 학습 능력을 유지한다","authors":"Shankar Padmanabhan, Mustafa Omer Gul, Tanya Goyal","pos":"#2511","link":"https://openreview.net/forum?id=OJsGhlTayF","abs":"

Post-training endows pretrained LLMs with a variety of desirable skills, such as instruction-following, reasoning, and others. However, these post-trained LLMs only encode knowledge up to a cut-off date, necessitating continual adaptation. Unfortunately, existing solutions cannot effectively learn new knowledge from adaptation document corpora and simultaneously mitigate the forgetting of earlier learned capabilities. To address this, we introduce Distillation via Split Contexts (DiSC), a simple context-distillation based approach for continual knowledge adaptation. DiSC derives student and teacher distributions by conditioning on distinct segments of the training example and minimizes the KL divergence between them for the common tokens. This insight allows us to efficiently apply context-distillation without requiring explicit generation steps during training. We run experiments on three post-trained models and two adaptation domains. Compared to prior finetuning and distillation methods for continual adaptation, DiSC consistently reports the best trade-off between learning new knowledge and mitigating forgetting of previously learned skills like instruction-following and reasoning, or factual knowledge.

","absKo":"

Post-training은 pretrained LLM에 instruction-following, reasoning 등 다양한 바람직한 능력을 부여합니다. 그러나 이러한 post-trained LLM은 cut-off date까지의 지식만 인코딩하므로, 지속적인 적응이 필요합니다. 안타깝게도 기존 해결책은 adaptation document corpus로부터 새로운 지식을 효과적으로 학습하는 동시에, 이전에 학습한 능력의 망각을 동시에 완화하지 못합니다. 이를 해결하기 위해 우리는 continual knowledge adaptation을 위한 간단한 context-distillation 기반 접근법인 Distillation via Split Contexts(DiSC)를 도입합니다. DiSC는 학습 예제의 서로 다른 segment를 조건으로 삼아 student와 teacher distribution을 도출하고, 공통 token에 대해 두 distribution 간 KL divergence를 최소화합니다. 이러한 통찰을 통해 우리는 training 동안 명시적인 generation step 없이도 context-distillation을 효율적으로 적용할 수 있습니다. 우리는 세 개의 post-trained model과 두 개의 adaptation domain에서 실험을 수행했습니다. continual adaptation을 위한 기존 finetuning 및 distillation method와 비교했을 때, DiSC는 instruction-following, reasoning, factual knowledge와 같은 이전에 학습한 skill의 망각을 완화하면서 새로운 지식을 학습하는 균형 면에서 일관되게 가장 우수한 결과를 보고합니다.

"},{"id":"62592","en":"Beyond Sunk Costs: Boosting LLM Pre-training Efficiency via Orthogonal Growth of Mixture-of-Experts","ko":"매몰비용을 넘어: Mixture-of-Experts의 직교적 성장으로 LLM Pre-training 효율 향상","authors":"Ruizhe Wang, Yucheng Ding, Xiao Liu, Yaoxiang Wang, Peng CHENG, Baining Guo, Zheng-Jun Zha, Yeyun Gong","pos":"#2610","link":"https://openreview.net/forum?id=fDZZXgdx2F","abs":"

As the computational demands for pre-training Large Language Models (LLMs) continue to surge, the need for efficient training paradigms becomes critical. Despite the vast resources already invested in existing pre-trained checkpoints, these assets often remain under-leveraged due to architectural limitations. We introduce an \"orthogonal growth\" strategy designed to \"recycle\" these checkpoints by strategically expanding their parameters prior to continued training. Our method focuses on optimizing converged Mixture-of-Experts (MoE) models through two dimensions: interpositional layer copying for increased depth and noisy expert duplication for expanded width. Through extensive scaling laws analysis, we demonstrate a strong positive correlation between the \"sunk cost\" (prior investment) and the final model accuracy. Empirical results on models up to 70B parameters and 1T tokens show that our recycling approach yields a 10.6\\% accuracy improvement compared to training from scratch under identical extra compute budgets. This work provides a cost-effective blueprint for sustainable large-scale LLM development.

","absKo":"

대규모 언어 모델(LLM)의 pre-training에 필요한 계산 수요가 계속 급증함에 따라, 효율적인 학습 패러다임의 필요성이 점점 더 중요해지고 있다. 기존 pre-trained checkpoint에 이미 막대한 자원이 투입되었음에도, 이러한 자산은 구조적 한계 때문에 종종 충분히 활용되지 못한다. 우리는 계속된 학습에 앞서 이들 checkpoint의 파라미터를 전략적으로 확장함으로써 이를 “재활용”하는, “orthogonal growth” 전략을 제안한다. 우리의 방법은 수렴된 Mixture-of-Experts(MoE) 모델을 두 가지 차원에서 최적화하는 데 초점을 둔다: 깊이를 늘리기 위한 interpositional layer copying과 폭을 확장하기 위한 noisy expert duplication이다. 광범위한 scaling law 분석을 통해, 우리는 “sunk cost”(이전 투자)와 최종 모델 정확도 사이에 강한 양의 상관관계가 있음을 보인다. 최대 70B 파라미터와 1T tokens 규모의 모델에 대한 실험 결과, 우리의 재활용 접근법은 동일한 추가 compute 예산 하에서 scratch부터 학습하는 것과 비교해 10.6\\%의 정확도 향상을 달성한다. 이 연구는 지속 가능한 대규모 LLM 개발을 위한 비용 효율적 청사진을 제시한다.

"},{"id":"66270","en":"ACON: Optimizing Context Compression for Long-horizon LLM Agents","ko":"ACON: 장기 지평 LLM 에이전트를 위한 컨텍스트 압축 최적화","authors":"Minki Kang, Wei-Ning Chen, Dongge Han, Huseyin Inan, Lukas Wutschitz, Yanzhi Chen, Robert A Sim, Saravanakumar Rajmohan","pos":"#2702","link":"https://openreview.net/forum?id=5EmOOLtH5P","abs":"

Large language models (LLMs) are increasingly deployed as agents in dynamic real-world environments, where success depends on maintaining precise records of actions and observations. However, the resulting unbounded context growth in long-horizon agentic tasks makes two critical bottlenecks: prohibitive inference memory costs and reasoning degradation due to irrelevant information. Existing compression methods fail to fully address this, often relying on brittle heuristics or requiring parameter updates impractical for proprietary or large-scale LLMs. We introduce Agent Context Optimization (ACON), a unified framework that optimally compresses both observations and history into concise, informative representations. Distinct from prior works, ACON employs an optimization in natural language space: it iteratively refines compression guidelines based on failure analysis of the agent, ensuring critical state information is preserved without model fine-tuning. To further minimize computational overhead, we distill the optimized compressor into smaller models. Experiments on AppWorld, OfficeBench, and Multi-objective QA demonstrate that ACON reduces peak token usage by 26–54% while maintaining task performance. Notably, it enables smaller LMs to function effectively as long-horizon agents, achieving up to 46% performance improvement by mitigating context distraction.

","absKo":"

Large language models (LLMs)은 점점 더 dynamic한 real-world environment에서 agent로 배치되고 있으며, 성공은 행동과 관측을 정확하게 기록하는 능력에 달려 있다. 그러나 long-horizon agentic task에서 context가 무한정 커지면 두 가지 핵심 병목이 발생한다. 하나는 과도한 inference memory cost이고, 다른 하나는 irrelevant information으로 인한 reasoning 저하다. 기존 compression 방법은 이를 충분히 해결하지 못하며, 종종 취약한 heuristic에 의존하거나 proprietary 또는 large-scale LLM에는 비현실적인 parameter update를 요구한다. 우리는 관측과 history 모두를 간결하고 정보성 높은 표현으로 최적으로 압축하는 통합 framework인 Agent Context Optimization (ACON)을 소개한다. 기존 연구와 달리, ACON은 natural language space에서의 optimization을 사용한다. 즉, agent의 failure analysis를 바탕으로 compression guideline을 반복적으로 정제하여, model fine-tuning 없이도 중요한 state information이 보존되도록 한다. 계산 오버헤드를 더 줄이기 위해, 최적화된 compressor를 더 작은 model로 distill한다. AppWorld, OfficeBench, Multi-objective QA에서의 실험은 ACON이 task performance를 유지하면서 peak token usage를 26–54% 줄임을 보여준다. 특히 context distraction을 완화함으로써 더 작은 LM이 long-horizon agent로 효과적으로 기능할 수 있게 하며, 최대 46%의 performance improvement를 달성한다.

"},{"id":"61625","en":"NITP: Next Implicit Token Prediction for LLM Pre-training","ko":"NITP: LLM Pre-training을 위한 Next Implicit Token Prediction","authors":"Xiangdong Zhang, Debing Zhang, Shaofeng Zhang, Xiaohan Qin, Yu Cheng, Junchi Yan","pos":"#304","link":"https://openreview.net/forum?id=oPQmCKS1tV","abs":"

Standard Next-Token Prediction (NTP) supervises language models solely through discrete labels in the output logit space. We argue that this sparse, one-hot supervision leaves the latent representation space under-constrained, allowing hidden states to drift into degenerate and anisotropic configurations that limit generalization. To address this issue, we propose Next Implicit Token Prediction (NITP), which augments discrete prediction with dense, continuous supervision directly in the representation space. NITP requires the model to predict the implicit semantic content of the next token, using shallow-layer representations from the same model as stable self-supervised targets. Theoretically, we show that NITP regularizes the optimization landscape by eliminating under-constrained degrees of freedom and enforcing a compact, structured representation geometry. Empirically, across dense and MoE models ranging from 0.5B to 9B parameters, NITP consistently improves downstream performance with negligible computational overhead. Notably, on the 9B MoE model, NITP achieves a 5.7% absolute improvement on MMLU-Pro, along with gains of 6.4% on C3 and 4.3% on CommonsenseQA, with ~2% additional training FLOPs and no additional inference cost.

","absKo":"

Standard Next-Token Prediction(NTP)은 출력 logit space의 discrete label만을 통해 language model을 감독한다. 우리는 이러한 희소한 one-hot supervision이 latent representation space를 충분히 제약하지 못하며, 그 결과 hidden state가 일반화를 제한하는 degenerate하고 anisotropic한 구성으로 drift할 수 있다고 주장한다. 이 문제를 해결하기 위해, 우리는 discrete prediction에 representation space에서 직접 이루어지는 dense하고 연속적인 supervision을 추가하는 Next Implicit Token Prediction(NITP)를 제안한다. NITP는 같은 model의 shallow-layer representation을 안정적인 self-supervised target으로 사용해, 모델이 다음 token의 implicit semantic content를 예측하도록 요구한다. 이론적으로 우리는 NITP가 제약이 부족한 자유도를 제거하고 compact하며 구조화된 representation geometry를 강제함으로써 optimization landscape를 정규화함을 보인다. 경험적으로, 0.5B에서 9B parameter에 이르는 dense 및 MoE model 전반에서 NITP는 일관되게 미미한 계산 오버헤드로 downstream performance를 향상시킨다. 특히 9B MoE model에서는 NITP가 MMLU-Pro에서 5.7%의 절대 향상을 달성했고, C3에서 6.4%, CommonsenseQA에서 4.3%의 향상을 보였으며, 추가 training FLOPs는 ~2%에 불과하고 inference cost는 추가되지 않았다.

"},{"id":"62187","en":"Memory-Efficient LLMs Training with Dynamic Sparsity: From Stability to Practical Scaling","ko":"Dynamic Sparsity로 메모리 효율적 LLM Training: 안정성에서 실용적 확장성으로","authors":"Qiao Xiao, Boqian Wu, Patrik Okanovic, Tomasz Sternal, Maurice Keulen, Elena Mocanu, Mykola Pechenizkiy, Decebal Constantin Mocanu, Torsten Hoefler","pos":"#309","link":"https://openreview.net/forum?id=ivVPgBZewP","abs":"

Dynamic Sparse Training (DST) offers a promising paradigm for improving the training and inference efficiency of deep neural networks; however, we find that in large language model training, DST suffers from optimization instability, manifested as loss spikes following topology updates. In this work, we show that the naive use of standard Adam-based optimizers leads to a cold-start issue for newly regrown parameters, resulting in excessively large updates and disrupted training dynamics. We propose Sparse Memory-Efficient Training (SMET), which stabilizes DST by combining optimizer warm-up for regrown parameters with density-aware learning-rate scaling. SMET further reduces memory consumption by storing gradients and optimizer states only for active parameters. We provide a theoretical analysis of the update behaviors under SMET, showing improved optimization stability. Extensive experiments demonstrate that SMET enables stable, scalable, and memory-efficient sparse pre-training of LLMs, paving the way for sparse training as a practical alternative to dense training.

","absKo":"

Dynamic Sparse Training (DST)는 deep neural network의 학습 및 추론 효율을 향상시키기 위한 유망한 패러다임을 제공하지만, 우리는 large language model 학습에서 DST가 topology update 이후 loss spike로 나타나는 optimization instability를 겪는다는 사실을 발견했다. 본 연구에서는 표준 Adam 기반 optimizer를 단순히 사용할 경우 새로 다시 성장한(regrown) parameter에 cold-start 문제가 발생하여, 지나치게 큰 update와 교란된 training dynamics를 초래함을 보인다. 우리는 regrown parameter에 대한 optimizer warm-up과 density-aware learning-rate scaling을 결합하여 DST를 안정화하는 Sparse Memory-Efficient Training (SMET)을 제안한다. SMET은 또한 활성 parameter에 대해서만 gradient와 optimizer state를 저장함으로써 memory consumption을 추가로 줄인다. 우리는 SMET 하에서의 update 거동에 대한 이론적 분석을 제시하여 optimization stability가 향상됨을 보인다. 광범위한 실험은 SMET이 LLM의 안정적이고 확장 가능하며 memory-efficient한 sparse pre-training을 가능하게 하여, sparse training이 dense training의 실용적인 대안이 될 수 있음을 입증한다.

"},{"id":"62339","en":"Draft-Conditioned Constrained Decoding for Structured Generation in LLMs","ko":"LLM 구조화 생성용 Draft-Conditioned Constrained Decoding","authors":"Avinash Reddy, Thayne Walker, Jaime Ide, Amrit Singh Bedi","pos":"#3209","link":"https://openreview.net/forum?id=hWDUpdwRZp","abs":"Large language models (LLMs) are increasingly used to generate executable outputs, JSON objects, and API calls, where a single syntax error can make the output unusable. Constrained decoding enforces validity token-by-token via masking and renormalization, but it can distort generation when the model assigns low probability mass to valid continuations, pushing decoding toward locally valid yet semantically incorrect trajectories. We propose \\emph{Draft-Conditioned Constrained Decoding (DCCD)}, a simple two-step, training-free inference procedure that decouples semantic planning from structural enforcement: an unconstrained draft is generated first, and constrained decoding is then applied, conditioned on this draft, to guarantee validity. We analyze DCCD through a KL-projection view, showing that draft conditioning increases feasible mass and reduces the cumulative “projection tax” induced by hard constraints, with an optional best-of-$K$ draft selection. Across structured reasoning benchmarks, DCCD improves strict structured accuracy by up to +24 percentage points over standard constrained decoding (e.g., 15.2\\% to 39.0\\% on GSM8K with a 1B model), and enables smaller model pairs to match or exceed much larger constrained baselines, yielding substantial gains in parameter efficiency.","absKo":"대규모 language model (LLM)은 executable output, JSON object, API call을 생성하는 데 점점 더 많이 사용되며, 이때 단 하나의 syntax error만으로도 출력이 사용할 수 없게 될 수 있다. Constrained decoding은 masking과 renormalization을 통해 token-by-token으로 validity를 강제하지만, 모델이 유효한 continuation에 낮은 probability mass를 부여할 경우 generation을 왜곡하여 locally valid하지만 semantic적으로는 잘못된 trajectory로 decoding을 밀어 넣을 수 있다. 우리는 semantic planning과 structural enforcement를 분리하는 간단한 2단계, training-free inference procedure인 \\emph{Draft-Conditioned Constrained Decoding (DCCD)}를 제안한다. 먼저 unconstrained draft를 생성한 뒤, 그 draft를 조건으로 constrained decoding을 적용하여 validity를 보장한다. 우리는 KL-projection 관점에서 DCCD를 분석하여, draft conditioning이 feasible mass를 증가시키고 hard constraint로 인해 유도되는 누적 “projection tax”를 줄임을 보이며, 선택적으로 best-of-$K$ draft selection도 가능함을 보인다. 구조화 추론 benchmark 전반에서 DCCD는 표준 constrained decoding 대비 strict structured accuracy를 최대 +24 percentage points까지 향상시키며(예: 1B model에서 GSM8K가 15.2\\%에서 39.0\\%로 상승), 더 작은 model pair가 훨씬 더 큰 constrained baseline과 같거나 그 이상의 성능을 내도록 하여 parameter efficiency 측면에서 큰 이득을 제공한다."},{"id":"66528","en":"Removing Noise, not Finding Gold: Quality Filtering for Large-Scale Pretraining","ko":"노이즈 제거, 금 찾기가 아니다: 대규모 사전 훈련을 위한 품질 필터링","authors":"Thiziri Nait Saada, Louis Béthune, Michal Klein, David Grangier, Marco Cuturi, Pierre Ablin","pos":"#3801","link":"https://openreview.net/forum?id=2taaKYQR7h","abs":"

Large-scale models are pretrained on massive web-crawled datasets containing documents of mixed quality, making data filtering essential. A popular method is Classifier-based Quality Filtering (CQF), which trains a binary classifier to distinguish between pretraining data and a small, high-quality set. It assigns each pretraining document a quality score defined as the classifier's score and retains only the top-scoring ones. We provide an in-depth analysis of CQF. We show that while CQF improves downstream task performance, it does not necessarily enhance language modeling on the high-quality set. Importantly, we find that training on CQF-selected data can outperform training directly on the high-quality set, even when the latter is sufficiently large. This finding alone is particularly striking, given the substantial effort and cost recently devoted to augmenting high-quality data. We explain this paradox by the fact that CQF implicitly filters the high-quality dataset as well as the low-quality one. Finally, we introduce an optimization-driven notion of data quality and demonstrate that it can be reliably estimated using small-scale proxy experiments. Altogether, our results both elucidate the mechanisms behind CQF and deepen our understanding of data selection methods widely used in practice.

","absKo":"

대규모 모델은 품질이 서로 다른 문서를 포함하는 방대한 웹 크롤링 데이터셋으로 pretrain되므로, 데이터 filtering이 필수적이다. 널리 쓰이는 방법은 Classifier-based Quality Filtering(CQF)로, pretraining data와 소규모의 고품질 집합을 구분하는 binary classifier를 학습시킨다. 이때 각 pretraining document에는 classifier score로 정의된 quality score가 할당되며, 점수가 높은 문서만 보존한다. 우리는 CQF에 대한 심층 분석을 제시한다. CQF가 downstream task 성능은 개선하지만, 반드시 고품질 집합에 대한 language modeling을 향상시키는 것은 아님을 보인다. 특히, 고품질 집합이 충분히 크더라도 CQF로 선택된 data로 학습하는 것이 고품질 집합을 직접 학습하는 것보다 더 나은 성능을 낼 수 있음을 발견했다. 최근 고품질 데이터를 확장하는 데 막대한 노력과 비용이 투입되었다는 점을 고려하면, 이 결과는 그 자체로 매우 놀랍다. 우리는 CQF가 저품질 데이터뿐 아니라 고품질 데이터도 암묵적으로 filtering하기 때문에 이러한 역설이 발생한다고 설명한다. 마지막으로, 우리는 optimization-driven 관점의 data quality 개념을 도입하고, 이를 소규모 proxy experiment으로 신뢰성 있게 추정할 수 있음을 보인다. 종합하면, 우리의 결과는 CQF의 작동 메커니즘을 설명할 뿐 아니라 실제로 널리 사용되는 data selection method에 대한 이해를 심화한다.

"},{"id":"63102","en":"Learn-to-learn on Arbitrary Textual Conditioning: A Hypernetwork-Driven Meta-gated LLM","ko":"임의 텍스트 조건에서의 학습법 학습: 하이퍼네트워크 기반 메타 게이트 LLM","authors":"Luo Ji, Qi Qin, Ningyuan Xi, Teng Chen, Qingqing Gu, Hongyan Li","pos":"#3909","link":"https://openreview.net/forum?id=a7LaDjVpTC","abs":"Conventional LLMs may suffer from heterogeneous corpus and subtle condition changes. While finetuning can create the catastrophe forgetting issue, application of meta-learning on LLMs is also limited due to its complexity and scalability. In this paper, we activate the meta-signal of $\\beta$ within the SwiGLU blocks, resulting a meta-gating mechanism which adaptively adjusts the nonlinearity of FFN. A hypernetwork is employed which dynamically produces $\\beta$ on textual conditions, providing meta-controllability on LLMs. By testing on different condition types such as task, domain, persona, and style, our method outperforms finetuning and meta-learning baselines, and can generalize reasonable on unseen task, condition type or instructions. Our code can be found https://anonymous.4open.science/r/MeGan-CAC0.","absKo":"기존 LLMs는 heterogeneous corpus와 미세한 condition 변화로 인해 성능 저하를 겪을 수 있다. finetuning은 catastrophe forgetting 문제를 일으킬 수 있고, LLMs에 meta-learning을 적용하는 것 역시 복잡성과 확장성의 한계로 제약된다. 본 논문에서는 SwiGLU block 내 $\\beta$의 meta-signal을 활성화하여, FFN의 비선형성을 적응적으로 조절하는 meta-gating mechanism을 구현한다. 또한 textual condition에 따라 $\\beta$를 동적으로 생성하는 hypernetwork를 사용하여, LLMs에 대한 meta-controllability를 제공한다. task, domain, persona, style 등 다양한 condition type에 대해 실험한 결과, 우리의 방법은 finetuning 및 meta-learning baseline을 능가하며, 보지 못한 task, condition type, instruction에도 합리적으로 generalize할 수 있다. 코드는 https://anonymous.4open.science/r/MeGan-CAC0 에서 확인할 수 있다."},{"id":"60763","en":"Capacity without Access: Reinterpreting the Mid-Depth Spectral Plateau in LLMs","ko":"접근 없는 용량: LLM의 중간 깊이 Spectral Plateau 재해석","authors":"Seong-Min Kang, Woo-Seong Yun, Nahyun Lee, Yoon-Sik Cho","pos":"#4008","link":"https://openreview.net/forum?id=wsw7Y085RY","abs":"

Prior probing-based analyses show that individual layers specialize in distinct linguistic and semantic functions. A complementary line of work observes that deeper layers enter an alignment-dominated regime, suggesting that residual updates become largely collinear with the hidden state vectors. These findings appear to conflict: layer-wise functional specialization versus apparent representational stagnation in deeper layers. This raises a key question: Is the observed representational stagnation in deeper layers attributable to suboptimal acquisition or encoding of novel features, or are these features learned appropriately but exhibit minimal marginal contribution to the model’s predictive output? We address this by separating model behavior into (i) representational capacity: the richness and spectral diversity of the encoded features, and (ii) accessibility: the extent to which these features are aligned with, and exploited by, the output-relevant subspace. Analyzing hidden-state covariance across depth shows that intermediate layers maintain a broad representational span and rich spectral diversity, indicating that their representational capacity remains largely intact. Yet they project only weakly onto output-relevant subspace, indicating that their accessibility within task-relevant representational dimensions is tightly constrained. To probe their functional relevance, we add a minimal diagnostic pathway that routes intermediate-layer embeddings to the terminal readout. Analysis of the induced behavioral changes indicates that the apparent stagnation is better accounted for by geometric limitations on accessibility, rather than by a deficit in representational capacity. The code is available at https://anonymous.4open.science/r/Capacity

","absKo":"

이전의 probing 기반 분석들은 개별 layer가 서로 다른 언어적, 의미적 기능에 특화되어 있음을 보여준다. 이에 상응하는 다른 연구 흐름은 deeper layer가 alignment가 지배적인 regime에 진입한다고 관찰하며, residual update가 hidden state vector와 대체로 collinear해진다고 시사한다. 이러한 결과들은 layer-wise functional specialization과 deeper layer에서의 겉보기 representation stagnation이라는 서로 상충하는 듯한 결론을 낳는다. 이는 중요한 질문을 제기한다. deeper layer에서 관찰되는 representation stagnation은 새로운 feature의 비최적 획득 또는 인코딩에 기인하는가, 아니면 이러한 feature는 적절히 학습되었지만 모델의 predictive output에 대한 marginal contribution이 매우 작기 때문인가? 우리는 model behavior를 (i) representational capacity: 인코딩된 feature의 풍부함과 spectral diversity, 그리고 (ii) accessibility: 이러한 feature가 output-relevant subspace와 얼마나 정렬되어 있고, 또한 얼마나 활용되는가로 분리하여 이 문제에 접근한다. depth에 따른 hidden-state covariance를 분석하면 intermediate layer가 넓은 representational span과 풍부한 spectral diversity를 유지하고 있어, representational capacity가 대체로 온전함을 보여준다. 그러나 이들은 output-relevant subspace에는 약하게 투영되므로, task-relevant representational dimension 내에서의 accessibility가 매우 제한되어 있음을 나타낸다. 이들의 기능적 관련성을 탐색하기 위해, intermediate-layer embedding을 terminal readout으로 연결하는 최소한의 diagnostic pathway를 추가한다. 이로 인해 유도된 behavioral change를 분석한 결과, 겉보기 stagnation은 representational capacity의 결함보다 accessibility에 대한 geometric limitation으로 더 잘 설명된다. 코드는 https://anonymous.4open.science/r/Capacity 에서 제공된다.

"},{"id":"61645","en":"Convex Dataset Valuation for Post-Training","ko":"Post-Training을 위한 Convex Dataset Valuation","authors":"Siqi Zeng, Christopher Jung, Rui Li, Zhe Kang, Ming Li, Nima Noorshams, Zhigang Wang, Fuchun Peng, Han Zhao, Xue Feng","pos":"#4107","link":"https://openreview.net/forum?id=oAAL0wZSMF","abs":"

Improving LLM performance on downstream tasks sometimes requires leveraging auxiliary datasets during post-training. In practice, however, developers face constraints on compute, labeling, and licensing costs that preclude using all available data, necessitating principled dataset-level selection. These constraints are increasingly shaped by dataset marketplaces, where data acquisition is governed by budgets and negotiation. We study dataset valuation as a subset selection problem during LLM post-training. Our goal is to identify and weight auxiliary datasets so as to maximize target task performance given constrained budgets. We first show that commonly used gradient alignment scores provide a reasonable yet incomplete valuation signal, as they ignore redundancy among datasets. To address this, we propose a convex scalable dataset-level valuation method based on kernel mean matching (KMM) in gradient space, which jointly accounts for alignment with the target task and redundancy across auxiliary datasets. Through extensive experiments across diverse post-training settings and multilingual reasoning tasks, we show that our approach consistently outperforms existing valuation baselines, achieving stronger performance with low computational overhead. Our results position dataset valuation as a practical decision tool for post-training data selection in market-constrained large language model settings.

","absKo":"

LLM의 downstream task 성능을 향상하려면 post-training 동안 auxiliary dataset을 활용해야 하는 경우가 있다. 그러나 실제로는 compute, labeling, licensing 비용 제약 때문에 사용할 수 있는 모든 데이터를 활용할 수 없으며, 이에 따라 원칙적인 dataset-level selection이 필요하다. 이러한 제약은 점점 dataset marketplace에 의해 형성되고 있으며, 여기서는 data acquisition이 budget과 negotiation에 의해 좌우된다. 우리는 LLM post-training에서 dataset valuation을 subset selection 문제로 다룬다. 우리의 목표는 제한된 budget 하에서 target task 성능을 최대화하도록 auxiliary dataset을 식별하고 가중치를 부여하는 것이다. 먼저, 널리 사용되는 gradient alignment score가 합리적이기는 하지만 불완전한 valuation signal임을 보인다. 이는 dataset 간 redundancy를 무시하기 때문이다. 이를 해결하기 위해, 우리는 gradient space에서 kernel mean matching(KMM)에 기반한 convex하고 scalable한 dataset-level valuation method를 제안하며, 이는 target task와의 alignment와 auxiliary dataset 간 redundancy를 동시에 고려한다. 다양한 post-training 설정과 multilingual reasoning task 전반에 걸친 광범위한 실험에서, 우리의 접근법이 기존 valuation baseline을 일관되게 능가하며 낮은 computational overhead로 더 강한 성능을 달성함을 보인다. 우리의 결과는 dataset valuation을 market-constrained large language model 환경에서 post-training data selection을 위한 실용적 의사결정 도구로 자리매김한다.

"},{"id":"62626","en":"POET-X: Memory-efficient LLM Training by Scaling Orthogonal Transformation","ko":"POET-X: Orthogonal Transformation 스케일링으로 메모리 효율적 LLM Training","authors":"Zeju Qiu, Lixin LIU, Adrian Weller, Han Shi, Weiyang Liu","pos":"#414","link":"https://openreview.net/forum?id=et8jpWLUuD","abs":"

Efficient and stable training of large language models (LLMs) remains a core challenge in modern machine learning systems. We tackle this problem with Reparameterized Orthogonal Equivalence Training (POET), a spectrum-preserving framework that optimizes each weight matrix through orthogonal equivalence transformation. Although POET provides strong training stability, its original implementation incurs high memory consumption and computational overhead due to intensive matrix multiplications. To overcome these limitations, we introduce POET-X, a scalable and memory-efficient variant that performs orthogonal equivalence transformations with significantly reduced computational cost. POET-X maintains the generalization and stability benefits of POET while achieving substantial improvements in throughput and memory efficiency. In experiments, POET-X enables the pretraining of billion-parameter LLMs on a single Nvidia H100 GPU, and in contrast, standard optimizers such as AdamW run out of memory under the same settings.

","absKo":"

대규모 언어 모델(LLM)의 효율적이고 안정적인 학습은 현대 머신러닝 시스템에서 여전히 핵심 과제이다. 우리는 각 가중치 행렬을 orthogonal equivalence transformation을 통해 최적화하는 spectrum-preserving framework인 Reparameterized Orthogonal Equivalence Training(POET)으로 이 문제를 다룬다. POET는 강한 학습 안정성을 제공하지만, 원래 구현은 intensive matrix multiplications 때문에 높은 메모리 사용량과 계산 오버헤드를 유발한다. 이러한 한계를 극복하기 위해, 우리는 계산 비용을 크게 줄인 상태에서 orthogonal equivalence transformations를 수행하는 확장 가능하고 메모리 효율적인 변형인 POET-X를 제안한다. POET-X는 POET의 일반화 및 안정성 이점을 유지하면서 throughput과 memory efficiency를 크게 향상시킨다. 실험에서 POET-X는 단일 Nvidia H100 GPU에서 billion-parameter LLM의 pretraining을 가능하게 했으며, 반대로 AdamW와 같은 표준 optimizer는 동일한 설정에서 메모리 부족으로 실행되지 못했다.

"},{"id":"61384","en":"Scaling-Aware Adapter for Structure-Grounded LLM Reasoning","ko":"Structure-Grounded LLM Reasoning을 위한 Scaling-Aware Adapter","authors":"Zihao Jing, QIUHAO Zeng, Ruiyi Fang, Yan Li, Yan Sun, Boyu Wang, Pingzhao Hu","pos":"#1905","link":"https://openreview.net/forum?id=ql7SeCTnvu","abs":"

Large language models (LLMs) enable reasoning over biomolecular structures, yet existing methods remain modality-specific and typically compress structural inputs via sequence-based tokenization or fixed-length query connectors. Such architectures either omit geometric grounding required to mitigate structural hallucinations or impose inflexible modality-fusion bottlenecks that both over-compress and misallocate structural tokens, impeding generalized all-atom reasoning. We introduce Cuttlefish, a unified all-atom LLM that grounds language reasoning in geometric cues while scaling modality tokens with structural complexity. First, Scaling-Aware Patching uses an instruction-conditioned gating mechanism to generate variable-size patches over structural graphs, adaptively scaling the query-token budget with structural complexity to mitigate fixed-length connector bottlenecks. Second, Geometry Grounding Adapter refines these adaptive tokens via cross-attention to modality embeddings and injects the resulting modality tokens into the LLM, exposing explicit geometric cues to reduce structural hallucination. Experiments across diverse all-atom benchmarks show that Cuttlefish achieves superior performance in heterogeneous structure-grounded reasoning.

","absKo":"

Large language model(LLM)은 biomolecular structure에 대한 추론을 가능하게 하지만, 기존 방법은 여전히 modality-specific하며 일반적으로 sequence-based tokenization 또는 fixed-length query connector를 통해 structural input을 압축한다. 이러한 architecture는 structural hallucination을 완화하는 데 필요한 geometric grounding을 누락하거나, 지나치게 압축하고 structural token을 잘못 배분하는 경직된 modality-fusion bottleneck을 강요하여, 일반화된 all-atom reasoning을 저해한다. 우리는 언어 추론을 geometric cue에 기반시키면서 구조 복잡도에 따라 modality token을 확장하는 통합 all-atom LLM인 Cuttlefish를 소개한다. 첫째, Scaling-Aware Patching은 instruction-conditioned gating mechanism을 사용하여 structural graph 위에 가변 크기 patch를 생성하고, structural complexity에 따라 query-token budget을 적응적으로 확장함으로써 고정 길이 connector bottleneck을 완화한다. 둘째, Geometry Grounding Adapter는 cross-attention을 통해 이러한 adaptive token을 modality embedding으로 정제하고, 그 결과 생성된 modality token을 LLM에 주입하여 명시적 geometric cue를 노출하고 structural hallucination을 줄인다. 다양한 all-atom benchmark 전반의 실험은 Cuttlefish가 heterogeneous structure-grounded reasoning에서 우수한 성능을 달성함을 보여준다.

"},{"id":"60570","en":"DAG-MoE: From Simple Mixture to Structural Aggregation in Mixture-of-Experts","ko":"DAG-MoE: Mixture-of-Experts에서 단순 혼합에서 구조적 집계로","authors":"Jiarui Feng, Hanqing Zeng, Karish Grover, Ruizhong Qiu, Yinglong Xia, Qiang Zhang, Qifan Wang, Ren Chen, Dongqi Fu, Jiayi Liu, Zhuokai Zhao, Xiangjun Fan, Benyu Zhang, Yixin Chen","pos":"#107","link":"https://openreview.net/forum?id=yv1iRquxF9","abs":"

Mixture-of-Experts (MoE) models have become a leading approach for decoupling parameter count from computational cost in large language models. Despite significant progress, effectively scaling MoE performance remains a challenge. Previous work shows that the use of fine-grained experts enlarges the space of expert combinations and can improve flexibility, but it also imposes substantial routing overhead, creating a new scalability bottleneck. In this paper, we explore a complementary axis for scaling --- expert-output mixture. We first analyze the limitations of the standard weighted-summation aggregation in conventional MoE architectures. We then theoretically demonstrate that introducing structural aggregation both expands the expert-combination space without altering the experts or router configuration and enables possible multi-step reasoning within a single MoE layer. To this end, we propose DAG-MoE, a sparse MoE framework that employs a lightweight module to automatically learn the optimal aggregation structure among the selected experts. We evaluate DAG-MoE under standard language modeling settings. Extensive experiments show that DAG-MoE consistently improves performance in both pretraining and fine-tuning, surpassing traditional MoE baselines.

","absKo":"

Mixture-of-Experts(MoE) 모델은 대규모 language model에서 파라미터 수와 계산 비용을 분리하는 선도적 접근으로 자리 잡았다. 상당한 진전에도 불구하고, MoE 성능을 효과적으로 확장하는 일은 여전히 도전 과제다. 기존 연구는 fine-grained expert를 사용하면 expert 조합 공간이 넓어지고 유연성이 향상될 수 있음을 보였지만, 동시에 routing overhead를 크게 증가시켜 새로운 확장성 병목을 만든다. 본 논문에서는 확장의 상보적 축으로 expert-output mixture를 탐구한다. 먼저 우리는 기존 MoE 아키텍처에서 표준 weighted-summation aggregation의 한계를 분석한다. سپس 이론적으로 structural aggregation을 도입하면 expert나 router 구성을 변경하지 않으면서도 expert 조합 공간을 확장할 수 있고, 단일 MoE layer 내에서 다단계 추론도 가능해짐을 보인다. 이를 위해 우리는 선택된 expert들 사이의 최적 aggregation structure를 자동으로 학습하는 경량 모듈을 사용하는 sparse MoE 프레임워크인 DAG-MoE를 제안한다. 우리는 표준 language modeling 설정에서 DAG-MoE를 평가했다. 광범위한 실험은 DAG-MoE가 pretraining과 fine-tuning 모두에서 일관되게 성능을 향상시키며, 기존 MoE baseline을 능가함을 보여준다.

"},{"id":"62300","en":"Towards One-to-Many Temporal Grounding","ko":"일대다 Temporal Grounding을 향해","authors":"Qi Xu, Tan Yue, Shihao Chen, Jiahao Meng, Anran Wang, Shunping Ji, Hao Fei, Xiangtai Li","pos":"#1110","link":"https://openreview.net/forum?id=hrZCqzmsPW","abs":"

Temporal Grounding (TG) aims to localize video segments corresponding to a textual query. Prior research predominantly focuses on single-segment retrieval. Real-world scenarios, however, often require localizing multiple disjoint segments for a single query—a setting we term One-to-Many Temporal Grounding (OMTG). Previous state-of-the-art MLLMs, optimized for one-to-one settings, struggle in this context, often yielding near-zero scores due to a lack of event cardinality perception. To bridge this gap, we present a systematic solution with three key contributions. First, we establish the first comprehensive OMTG benchmark, introducing Count Accuracy (C-Acc) and Effective Temporal F1 (EtF1) as evaluation metrics. Second, we curate a high-quality OMTG dataset comprising 56k samples through a sophisticated construction pipeline. Third, we develop novel temporal and caption reward functions specifically designed for OMTG. In particular, the caption reward leverages Chain-of-Thought reasoning over dense video captions to explicitly guide policy optimization toward both preciseness and completeness. Extensive experiments show our model achieves a new state-of-the-art EtF1 of 43.65\\% on OMTG Bench, outperforming Gemini 2.5 Pro and Seed-1.8 by 15.85\\% and 15.61\\%, respectively.

","absKo":"

Temporal Grounding (TG)은 textual query에 대응하는 video segment를 localize하는 것을 목표로 한다. 기존 연구는 주로 single-segment retrieval에 초점을 맞춰 왔다. 그러나 실제 상황에서는 하나의 query에 대해 여러 개의 분리된 segment를 localize해야 하는 경우가 많으며, 우리는 이를 One-to-Many Temporal Grounding (OMTG)이라 부른다. one-to-one setting에 최적화된 기존 state-of-the-art MLLMs는 event cardinality perception이 부족해 이 맥락에서 잘 작동하지 못하며, 종종 거의 0에 가까운 점수를 낸다. 이 격차를 메우기 위해, 우리는 세 가지 핵심 기여를 포함하는 체계적 해법을 제시한다. 첫째, 우리는 Count Accuracy (C-Acc)와 Effective Temporal F1 (EtF1)을 평가 지표로 도입한 최초의 포괄적 OMTG benchmark를 구축한다. 둘째, 정교한 construction pipeline을 통해 56k samples로 구성된 고품질 OMTG dataset을 수집한다. 셋째, OMTG에 특화된 새로운 temporal 및 caption reward functions를 개발한다. 특히 caption reward는 dense video captions에 대한 Chain-of-Thought reasoning을 활용하여 policy optimization이 preciseness와 completeness 모두를 향하도록 명시적으로 유도한다. 광범위한 실험에서 우리 모델은 OMTG Bench에서 43.65\\%의 새로운 state-of-the-art EtF1을 달성했으며, Gemini 2.5 Pro와 Seed-1.8을 각각 15.85\\%, 15.61\\% 차이로 능가한다.

"},{"id":"65191","en":"Where Signals Are Sparse, We Synthesize: Reinforcing Self-Corrective Reasoning in Vision–Language Models via Rollout Augmentation","ko":"신호가 희소한 곳에서 우리는 합성한다: Rollout 증강을 통해 Vision–Language Model의 자기 교정 추론 강화하기","authors":"Yi Ding, Ziliang Qiu, Bolian Li, Ruqi Zhang","pos":"#1705","link":"https://openreview.net/forum?id=FxiO8RTOsq","abs":"Self-correction is essential for solving complex reasoning problems in vision–language models (VLMs), yet existing reinforcement learning (RL) methods struggle to learn it. Effective self-correction behaviors emerge only rarely during RL, making learning signals sparse. To address this challenge, we propose c**o**rre**ct**i**o**n-s**p**ecific rollo**u**t**s**} (**Octopus**), a rollout-augmentation framework that synthesizes dense self-correction supervision by recombining existing rollouts without computational overhead. This rollout augmentation simultaneously improves sample efficiency and stabilizes RL optimization. Furthermore, we introduce a two-stage RL training strategy that disentangles self-correction and direct reasoning, avoiding signal conflicts and enabling both behaviors to be learned effectively. Building on this, we introduce $\\texttt{Octopus-8B}$, an advanced reasoning VLM with controllable self-correction capabilities. It achieves SoTA performance among open-source VLMs across 7 benchmarks, outperforming the best RLVR baseline by 1.0 score while requiring only $0.72\\times$ training time per step.","absKo":"복잡한 reasoning 문제를 vision–language model (VLM)에서 해결하려면 self-correction이 필수적이지만, 기존 reinforcement learning (RL) 방법은 이를 학습하는 데 어려움을 겪는다. 효과적인 self-correction behavior는 RL 동안 드물게만 나타나므로 학습 신호가 희소하다. 이 문제를 해결하기 위해, 우리는 c**o**rre**ct**i**o**n-s**p**ecific rollo**u**t**s**} (**Octopus**)를 제안한다. 이는 추가 계산 오버헤드 없이 기존 rollout을 재조합하여 dense한 self-correction supervision을 합성하는 rollout-augmentation framework이다. 이 rollout augmentation은 sample efficiency를 동시에 향상시키고 RL 최적화를 안정화한다. 더 나아가, 우리는 self-correction과 direct reasoning을 분리하여 signal conflict를 피하고 두 행동이 효과적으로 학습되도록 하는 two-stage RL training strategy를 도입한다. 이를 바탕으로, 우리는 controllable self-correction capability를 갖춘 고급 reasoning VLM인 $\\texttt{Octopus-8B}$를 소개한다. 이는 7개 benchmark 전반에서 open-source VLM 중 SoTA 성능을 달성했으며, step당 training time은 $0.72\\times$만 사용하면서도 기존 최고의 RLVR baseline보다 1.0 score 더 높다."},{"id":"60578","en":"UniScale: Adaptive Unified Inference Scaling via Online Joint Optimization of Model Routing and Test-Time Scaling","ko":"UniScale: 모델 라우팅과 테스트 타임 스케일링의 온라인 공동 최적화를 통한 적응형 통합 추론 스케일링","authors":"Kaiyu Huang, Xingyu Wang, Mingze Kong, Zhubo Shi, Yuqian Hou, Hong Xu, Zhongxiang Dai, Minchen Yu, Qingjiang Shi","pos":"#1711","link":"https://openreview.net/forum?id=yopy1zxFFM","abs":"

In real-world deployments of large language models (LLMs), balancing inference quality and computational cost has become a central challenge. Existing approaches tackle this trade-off along two largely independent dimensions: model routing, which switches among models of different scales to match request complexity, and test-time scaling (TTS), which adjusts inference-time compute within a fixed model for fine-grained control. However, this decoupled design introduces inherent limitations. Model routing yields coarse-grained, discrete performance changes due to the sparse set of model scales, while single-model TTS often encounters capacity ceilings and exhibits diminishing returns as compute increases. Moreover, treating the two mechanisms separately restricts adaptability in dynamic inference environments. To overcome these limitations, we introduce Unified Inference Scaling (UIS), which unifies model routing and TTS in a single optimization space. Building on this formulation, we propose UniScale, an online framework that models adaptive UIS as a contextual multi-armed bandit problem and learns inference policies via LinUCB. The framework incorporates efficiency-aware learning and cost modeling to ensure stable and scalable optimization over high-dimensional action spaces. Evaluation shows that UniScale effectively exploits the synergy in the UIS space to deliver a fine-grained and consistently better quality-cost trade-off across diverse, dynamic inference scenarios.

","absKo":"

실제 대규모 언어 모델(LLM) 배포에서는 inference quality와 computational cost의 균형을 맞추는 것이 핵심 과제가 되었다. 기존 접근법은 이 trade-off를 대체로 서로 독립적인 두 차원에서 다룬다. 하나는 요청의 복잡도에 맞추어 서로 다른 규모의 model 사이를 전환하는 model routing이고, 다른 하나는 고정된 모델 내부에서 inference-time compute를 조절하여 세밀한 제어를 제공하는 test-time scaling(TTS)이다. 그러나 이러한 분리 설계는 본질적인 한계를 낳는다. Model routing은 희소한 model scale 집합 때문에 거칠고 이산적인 성능 변화를 제공하는 반면, 단일 모델 TTS는 종종 capacity ceiling에 부딪히며 compute가 증가해도 수익이 점차 감소한다. 더 나아가, 두 메커니즘을 분리해서 다루면 동적 inference 환경에서의 적응성이 제한된다. 이러한 한계를 극복하기 위해, 우리는 model routing과 TTS를 단일 optimization space로 통합한 Unified Inference Scaling (UIS)를 제안한다. 이 formulation을 바탕으로, 우리는 적응형 UIS를 contextual multi-armed bandit problem으로 모델링하고 LinUCB를 통해 inference policy를 학습하는 온라인 framework인 UniScale을 제안한다. 이 framework는 고차원 action space 전반에서 안정적이고 확장 가능한 optimization을 보장하기 위해 efficiency-aware learning과 cost modeling을 포함한다. 평가는 UniScale이 UIS space의 시너지를 효과적으로 활용하여 다양한 동적 inference 시나리오에서 더 세밀하고 일관되게 우수한 quality-cost trade-off를 제공함을 보여준다.

"},{"id":"65616","en":"UMEM: Unified Memory Extraction and Management Framework for Generalizable Memory","ko":"UMEM: 일반화 가능한 메모리를 위한 통합 메모리 추출 및 관리 프레임워크","authors":"Yongshi Ye, Hui Jiang, Feihu Jiang, Tian Lan, Yichao Du, Biao Fu, Xiaodong Shi, Qianghuai Jia, Longyue Wang, Weihua Luo","pos":"#1713","link":"https://openreview.net/forum?id=BoiXvrwtdi","abs":"

Self-evolving memory serves as the trainable parameters for Large Language Models (LLMs)-based agents, where extraction (distilling insights from experience) and management (updating the memory bank) must be tightly coordinated. Existing methods predominately optimize memory management while treating memory extraction as a static process, resulting in poor generalization, where agents accumulate instance-specific noise rather than robust memories. To address this, we propose Unified Memory Extraction and Management (UMEM), a self-evolving agent framework that jointly optimizes a Large Language Model to simultaneous extract and manage memories. To mitigate overfitting to specific instances, we introduce Semantic Neighborhood Modeling and optimize the model with a neighborhood-level marginal utility reward via GRPO. This approach ensures memory generalizability by evaluating memory utility across clusters of semantically related queries. Extensive experiments across five benchmarks demonstrate that UMEM significantly outperforms highly competitive baselines, achieving up to a 10.67% improvement in multi-turn interactive tasks. Futhermore, UMEM maintains a monotonic growth curve during continuous evolution. Codes and models will be publicly released.

","absKo":"

Self-evolving memory는 Large Language Models (LLMs)-based agents의 학습 가능한 parameter로 작동하며, 여기서 extraction(경험으로부터 insight를 distill하는 것)과 management(memory bank를 업데이트하는 것)는 긴밀하게 조정되어야 한다. 기존 방법들은 주로 memory management를 최적화하는 데 집중하면서 memory extraction을 정적인 과정으로 취급해 왔고, 그 결과 일반화가 좋지 않아 agent가 견고한 memory가 아니라 instance-specific noise를 축적하게 된다. 이를 해결하기 위해 우리는 Unified Memory Extraction and Management (UMEM)을 제안한다. 이는 Large Language Model이 memory를 동시에 extract하고 manage하도록 공동 최적화하는 self-evolving agent framework이다. 특정 instance에 대한 overfitting을 완화하기 위해, 우리는 Semantic Neighborhood Modeling을 도입하고 GRPO를 통해 neighborhood-level marginal utility reward로 model을 최적화한다. 이 접근법은 의미적으로 관련된 query 클러스터 전반에서 memory utility를 평가함으로써 memory의 generalizability를 보장한다. 다섯 개 benchmark에 걸친 광범위한 실험은 UMEM이 매우 경쟁력 있는 baseline들을 크게 능가하며, multi-turn interactive task에서 최대 10.67%의 향상을 달성함을 보여준다. 더 나아가 UMEM은 continuous evolution 동안 monotonic growth curve를 유지한다. 코드와 model은 공개적으로 release될 예정이다.

"},{"id":"61483","en":"Skip a Layer or Loop It? Learning Program-of-Layers in LLMs","ko":"Layer를 건너뛸까, 반복할까? LLM에서 Program-of-Layers 학습","authors":"Ziyue Li, Yang Li, Tianyi Zhou","pos":"#1902","link":"https://openreview.net/forum?id=pl10b6EQAN","abs":"

Large language models (LLMs) perform inference by following a fixed depth and order, non-recurrent execution of all layers. We reveal the wide existence of training-free, flexible, dynamic \"program-of-layers (PoLar)\", where pretrained layers can be packed as modules and then skipped or looped to form a customized program for each input. For most inputs, substantially shorter program executions can achieve the same or better accuracy, while incorrect predictions of the original LLM can be corrected by alternative programs with fewer layers. These observations indicate that inference admits multiple valid latent computations beyond the standard forward pass. To efficiently achieve PoLar in practice, we propose a lightweight PoLar prediction network, which learns to generate execution programs that dynamically skip or repeat pretrained layers for each input. Experiments on mathematical reasoning benchmarks demonstrate that PoLar consistently improves accuracy over standard inference and prior dynamic-depth methods, often while executing fewer layers, and that these gains persist under out-of-distribution evaluation. Our results suggest that fixed-depth execution captures only a narrow subset of an LLM’s latent reasoning capacity.

","absKo":"

Large language models (LLMs)는 모든 layer를 비순환적으로, 고정된 depth와 순서에 따라 실행함으로써 inference를 수행한다. 우리는 pretrained layer를 module처럼 묶은 뒤, 이를 건너뛰거나 반복하여 각 입력마다 맞춤형 program을 구성하는, 학습이 필요 없는 유연하고 동적인 \"program-of-layers (PoLar)\"의 광범위한 존재를 밝힌다. 대부분의 입력에 대해, 훨씬 더 짧은 program execution만으로도 동일하거나 더 나은 accuracy를 달성할 수 있으며, 원래 LLM의 잘못된 prediction도 더 적은 layer를 사용하는 대체 program으로 교정될 수 있다. 이러한 관찰은 standard forward pass를 넘어 inference가 여러 개의 유효한 latent computation을 허용함을 시사한다. 실제로 PoLar를 효율적으로 구현하기 위해, 우리는 각 입력마다 pretrained layer를 동적으로 건너뛰거나 반복하는 execution program을 생성하도록 학습하는 경량 PoLar prediction network를 제안한다. 수학적 reasoning benchmark에서의 실험은 PoLar가 표준 inference와 이전 dynamic-depth method보다 일관되게 accuracy를 향상시키며, 더 적은 layer를 실행하면서도 이러한 향상이 out-of-distribution evaluation에서도 유지됨을 보여준다. 우리의 결과는 고정 깊이 실행이 LLM의 latent reasoning capacity 중 극히 일부만 포착한다는 점을 시사한다.

"},{"id":"63571","en":"Propose, Solve, Verify: Self-Play Through Formal Verification","ko":"제안하고, 해결하고, 검증하기: 형식 검증을 통한 self-play","authors":"Alex Wilf, Pranjal Aggarwal, Bryan Parno, Daniel Fried, Louis-Philippe Morency, Paul Pu Liang, Sean Welleck","pos":"#2006","link":"https://openreview.net/forum?id=Vuaq1qXCki","abs":"

Training models through self-play alone (without any human data) has been a longstanding goal in AI, but its effectiveness for training large language models remains unclear, particularly in code generation where rewards based on unit tests are brittle and prone to error propagation. We study self-play in the verified code generation setting, where formal verification provides reliable correctness signals. We introduce Propose, Solve, Verify (PSV) a simple self-play framework where formal verification signals are used to create a proposer capable of generating challenging synthetic problems and a solver trained via expert iteration. We use PSV to train PSV-Verus, which across three benchmarks improves pass@1 by up to 9.6x over inference-only and expert-iteration baselines. We show that performance scales with the number of generated questions and training iterations, and through ablations identify formal verification and difficulty-aware proposal as essential ingredients for successful self-play.

","absKo":"

인간 데이터 없이 self-play만으로 모델을 학습하는 것은 AI에서 오랫동안 추구되어 온 목표였지만, 특히 unit test 기반 보상이 취약하고 error propagation에 취약한 code generation에서 large language model(LLM) 학습에 얼마나 효과적인지는 여전히 불분명하다. 우리는 formal verification이 신뢰할 수 있는 정답 신호를 제공하는 verified code generation setting에서 self-play를 연구한다. 우리는 Propose, Solve, Verify (PSV)를 소개하는데, 이는 formal verification 신호를 사용해 도전적인 synthetic problem을 생성할 수 있는 proposer와 expert iteration으로 학습된 solver를 만드는 단순한 self-play framework이다. 우리는 PSV로 PSV-Verus를 학습하며, 세 개의 benchmark에서 inference-only 및 expert-iteration baseline 대비 pass@1을 최대 9.6배 향상시킨다. 우리는 성능이 생성된 question 수와 training iteration 수에 따라 scale함을 보이고, ablation을 통해 successful self-play의 필수 요소로 formal verification과 difficulty-aware proposal을 확인한다.

"},{"id":"60967","en":"Pushing the Boundaries of Natural Reasoning: Interleaved Bonus from Formal-Logic Verification","ko":"Natural Reasoning의 한계 확장: Formal-Logic Verification에서 얻는 Interleaved Bonus","authors":"Chuxue Cao, Jinluan Yang, Haoran Li, Kunhao Pan, Zijian Zhao, Zhengyu Chen, Yuchen Tian, Lijun Wu, Conghui He, Sirui Han, Yike Guo","pos":"#2007","link":"https://openreview.net/forum?id=uzVlKB1C5E","abs":"

Large Language Models (LLMs) show remarkable capabilities, yet their stochastic next-token prediction creates logical inconsistencies and reward hacking that formal symbolic systems avoid. To bridge this gap, we introduce a formal logic verification-guided framework that dynamically interleaves formal symbolic verification with the natural language generation process, providing real-time feedback to detect and rectify errors as they occur. Distinguished from previous neuro-symbolic methods limited by passive post-hoc validation, our approach actively penalizes intermediate fallacies during the reasoning chain. We operationalize this framework via a novel two-stage training pipeline that synergizes formal logic verification-guided supervised fine-tuning and policy optimization. Extensive evaluation on six benchmarks spanning mathematical, logical, and general reasoning demonstrates that our 7B and 14B models outperform state-of-the-art baselines by average margins of 10.4\\% and 14.2\\%, respectively. These results validate that formal verification can serve as a scalable mechanism to significantly push the performance boundaries of advanced LLM reasoning.

","absKo":"

Large Language Model (LLM)은 놀라운 능력을 보이지만, stochastic next-token prediction 때문에 formal symbolic system이 피하는 논리적 불일치와 reward hacking이 발생한다. 이러한 간극을 메우기 위해, 우리는 formal symbolic verification을 자연어 생성 과정과 동적으로 교차시키는 formal logic verification-guided framework를 도입하여, 오류가 발생하는 즉시 이를 탐지하고 수정할 수 있는 실시간 피드백을 제공한다. 수동적 사후 검증에 국한된 기존 neuro-symbolic method와 달리, 우리의 접근법은 reasoning chain의 중간 단계에서 발생하는 fallacy를 적극적으로 penalize한다. 우리는 이 framework를 formal logic verification-guided supervised fine-tuning과 policy optimization을 결합한 새로운 2단계 training pipeline으로 구현한다. 수학, 논리, 일반 추론을 아우르는 6개 benchmark에 대한 광범위한 평가는 우리의 7B 및 14B model이 각각 평균 10.4\\%와 14.2\\%의 margin으로 state-of-the-art baseline을 능가함을 보여준다. 이러한 결과는 formal verification이 advanced LLM reasoning의 성능 한계를 유의미하게 확장하는 확장 가능한 메커니즘으로 작동할 수 있음을 검증한다.

"},{"id":"60585","en":"ReMoE: Boosting Expert Reuse through Router Fine-Tuning in Memory-Constrained MoE LLM Inference","ko":"ReMoE: 메모리 제약 MoE LLM 추론에서 라우터 파인튜닝을 통한 전문가 재사용 향상","authors":"Xiongwei Zhu, Xiaojian Liao, Tianyang Jiang, Yusen Zhang, Liang Wang, Limin Xiao","pos":"#2009","link":"https://openreview.net/forum?id=ylAhgNb2ak","abs":"

Fine-grained Mixture-of-Experts (MoE) models sparsely activate a subset of parameters, significantly reducing computational costs while maintaining performance. However, in memory-constrained inference scenarios, only a small set of experts can be cached. Experts not in the cache must be fetched from slow external storage (e.g., UFS), leading to frequent evictions and substantial I/O overhead. We propose ReMoE, a router fine-tuning framework designed to boost token-wise expert reuse. By introducing a temporal inductive bias, ReMoE encourages the model to consistently select the same experts over time, which aligns the routing behavior with cache locality constraints, reducing the need to fetch experts from storage without adding any extra computation during inference. Experiments on DeepSeek and Qwen models show that ReMoE improves the expert reuse rate by 26\\%. Under a standard LRU caching policy simulation, ReMoE improves the cache hit rate by 15.7\\%, corresponds to a 7.8\\% reduction in median latency and an 8.5\\% increase in proxy throughput, while maintaining downstream task performance.

","absKo":"

Fine-grained Mixture-of-Experts (MoE) models는 파라미터의 일부만 희소하게 활성화하여 성능을 유지하면서도 계산 비용을 크게 줄입니다. 그러나 memory 제약이 있는 inference 시나리오에서는 소수의 expert만 cache할 수 있습니다. cache에 없는 expert는 느린 external storage(예: UFS)에서 가져와야 하므로, 빈번한 eviction과 상당한 I/O overhead가 발생합니다. 우리는 token-wise expert reuse를 높이기 위해 설계된 router fine-tuning framework인 ReMoE를 제안합니다. temporal inductive bias를 도입함으로써 ReMoE는 모델이 시간에 따라 동일한 expert를 일관되게 선택하도록 유도하며, 이는 routing behavior를 cache locality 제약과 정렬시켜 inference 중 추가 계산 없이 storage에서 expert를 가져와야 하는 필요를 줄입니다. DeepSeek 및 Qwen model에 대한 실험은 ReMoE가 expert reuse rate를 26\\% 향상시킴을 보여줍니다. 표준 LRU caching policy simulation 하에서 ReMoE는 cache hit rate를 15.7\\% 향상시키며, 이는 median latency 7.8\\% 감소와 proxy throughput 8.5\\% 증가에 해당하고, downstream task performance는 유지합니다.

"},{"id":"63073","en":"ReQAT: Achieving Full-Precision Reasoning Accuracy with 4-bit Floating-Point Quantization-Aware Training","ko":"ReQAT: 4비트 부동소수점 양자화 인지 학습으로 완전 정밀도 추론 정확도 달성","authors":"Janghwan Lee, Sihwa Lee, Jinseok Kim, Yongjik Kim, Jieun Lim, Jinwook Oh, Jungwook Choi","pos":"#2010","link":"https://openreview.net/forum?id=aKY52fnzgc","abs":"Large Reasoning Models (LRMs) achieve strong problem-solving through long chain-of-thought, but their deployment is constrained by the high cost of full-precision inference and growing KV cache footprints. Microscaled FP4 formats enable efficient FP4 deployment; however, fully quantizing weights, activations, and KV caches (W4A4KV4) causes severe reasoning degradation that existing PTQ and QAT fail to recover. We identify that FP4 failures concentrate on low-entropy tokens—precise symbolic commitments such as digits and operators—where quantization noise inflates sampling errors that cascade through reasoning traces. Based on this insight, we propose ReQAT, a reasoning-centric FP4 training framework with three components: (i) Trace-Aligned QAT (TAQ), which revisits identical reasoning traces to focus updates on critical low-entropy decisions; (ii) Selective Entropy Minimization (SEM), which reinforces confidence at low-entropy positions; and (iii) Q-FIT, a quantization-friendly initialization that jointly calibrates RoPE-consistent KV cache transformations to stabilize QAT. Under the same training budget, ReQAT not only recovers but surpasses BF16 fine-tuning accuracy—achieving while delivering up to $3.9\\times$ throughput speedup on NVIDIA DGX Spark and $3.1\\times$ on B200. This is the first demonstration that FP4 QAT can exceed full-precision accuracy for LRMs with over 3× speedup on production hardware.","absKo":"Large Reasoning Models (LRMs)는 긴 chain-of-thought를 통해 강력한 문제 해결 성능을 보이지만, full-precision inference의 높은 비용과 커져가는 KV cache footprint로 인해 배포가 제약된다. Microscaled FP4 format은 효율적인 FP4 배포를 가능하게 하지만, weight, activation, KV cache를 모두 완전히 quantize하는 W4A4KV4는 심각한 reasoning 성능 저하를 유발하며 기존 PTQ와 QAT는 이를 복구하지 못한다. 우리는 FP4 실패가 low-entropy token, 즉 digit와 operator 같은 정밀한 symbolic commitment에 집중된다는 점을 확인했다. 이 지점에서 quantization noise가 sampling error를 증폭시키고, 그 오류가 reasoning trace 전반으로 누적된다. 이 통찰을 바탕으로, 우리는 세 가지 구성요소를 갖는 reasoning-centric FP4 training framework인 ReQAT를 제안한다: (i) Trace-Aligned QAT (TAQ), 동일한 reasoning trace를 다시 검토하여 중요한 low-entropy decision에 업데이트를 집중시킨다; (ii) Selective Entropy Minimization (SEM), low-entropy 위치에서 confidence를 강화한다; (iii) Q-FIT, RoPE-consistent KV cache transformation을 함께 보정하여 QAT를 안정화하는 quantization-friendly initialization이다. 동일한 training budget에서 ReQAT는 BF16 fine-tuning accuracy를 복구하는 데 그치지 않고 이를 능가하며, NVIDIA DGX Spark에서 최대 $3.9\\times$, B200에서 $3.1\\times$의 throughput speedup을 달성한다. 이는 production hardware에서 3× 이상의 speedup과 함께 FP4 QAT가 LRMs에서 full-precision accuracy를 능가할 수 있음을 처음으로 보인 결과다."},{"id":"61102","en":"LoKiFormer: Locality-aware Attention with Decoupled Knowledge Memory for Efficient Large Language Model Pretraining","ko":"LoKiFormer: 효율적인 Large Language Model 사전학습을 위한 Decoupled Knowledge Memory 기반 Locality-aware Attention","authors":"Zimo Liu, Qiuwu Chen, Yuchen Li, Ying Sun, Yifan Zhang, Zhijie Qiu, Zeng You, Ryan Dong, Simeng Ma, Yaofo Chen, Mingkui Tan","pos":"#203","link":"https://openreview.net/forum?id=tgmYeWfvUU","abs":"

Large language models (LLMs) have achieved remarkable breakthroughs across various applications. However, their architectures remain inefficient in pretraining due to two main limitations: (i) self-attention lacks an explicit inductive bias for locality, leading to redundant modeling of sequence-internal local information; (ii) mixture-of-experts (MoE) implicitly couples knowledge storage with computational pathways, hindering flexible access to sequence-external global knowledge. To overcome these limitations, we propose LoKiFormer, a novel LLM architecture that augments the standard decoder with two dedicated modules: 1) Local Fusion Attention (LFA), which incorporates a convolutional fusion to attention, explicitly capturing local patterns and allowing the attention to operate on more informative representations; 2) Knowledge Memory Module (KMM), which introduces a parametric key–value memory that explicitly stores global knowledge in addressable slots, decoupling storage from computation and enabling direct knowledge retrieval. Together, these modules enable LoKiFormer to achieve more efficient and effective integration of information at both levels. Experimental results show that LoKiFormer converges 1.33x faster in pre-training than baseline models, underscoring its superiority over existing LLM architectures.

","absKo":"

대규모 언어 모델(LLMs)은 다양한 응용에서 놀라운 성과를 달성했다. 그러나 이들의 아키텍처는 사전학습(pretraining)에서 두 가지 주요 한계로 인해 여전히 비효율적이다: (i) self-attention은 locality에 대한 명시적 inductive bias가 없어, sequence 내부의 local information을 중복적으로 모델링하게 된다; (ii) mixture-of-experts (MoE)는 knowledge storage와 computational pathways를 암묵적으로 결합하여, sequence 외부의 global knowledge에 대한 유연한 접근을 방해한다. 이러한 한계를 극복하기 위해, 우리는 표준 decoder에 두 개의 전용 모듈을 추가하는 새로운 LLM 아키텍처 LoKiFormer를 제안한다: 1) Local Fusion Attention (LFA)는 convolutional fusion을 attention에 통합하여 local pattern을 명시적으로 포착하고, attention이 더 정보량이 풍부한 representation 위에서 동작하도록 한다; 2) Knowledge Memory Module (KMM)은 addressable slot에 global knowledge를 명시적으로 저장하는 parametric key–value memory를 도입하여, storage와 computation을 분리하고 직접적인 knowledge retrieval을 가능하게 한다. 이 두 모듈을 통해 LoKiFormer는 두 수준 모두에서 정보를 더 효율적이고 효과적으로 통합할 수 있다. 실험 결과, LoKiFormer는 baseline 모델보다 pre-training에서 1.33배 더 빠르게 수렴하며, 기존 LLM 아키텍처 대비 우수성을 입증한다.

"},{"id":"66571","en":"Op-CAD: Benchmarking and Investigating Operation-oriented CAD Generation","ko":"Op-CAD: 연산 지향 CAD 생성의 벤치마킹 및 탐구","authors":"Yixue Bai, Yufei Gu, Zeke Xie","pos":"#2102","link":"https://openreview.net/forum?id=2VSY8G3plF","abs":"

Recent research has made growing efforts to leverage large language models (LLMs) for computer-aided design (CAD), a domain that demands advanced geometric and spatial reasoning across long operation sequence. However, existing studies remain limited in addressing complex modeling tasks that necessitate step-by-step reasoning, primarily due to the scarcity of high-quality CAD datasets and the absence of fine-grained evaluation frameworks. In response to these challenges, we introduce Op-CAD, the first large-scale, multi-modal dataset for operation-oriented CAD generation, encompassing four operation types and five modalities. Furthermore, we introduce a novel CAD parsing module together with a geometry-guided hierarchical annotation pipeline, which decomposes modeling sequences into discrete operations and substantially improves the annotation accuracy of Vision-Language Models (VLMs). Based on our dataset, we redefine the CAD modeling task by decoupling geometric and spatial perspectives and introduce a novel metric, Chamfer/Fillet Intersection over Union (CF-IoU), to fill the void in assessing chamfer and fillet operations. By comprehensively evaluating eight LLMs on Op-CAD, we establish a benchmark for current models on operation-oriented tasks. Finally, we investigate performance enhancement strategies through fine-tuning on Op-CAD and propose Chain-of-Operation (COOP), a novel prompting strategy that emulates human-engineer reasoning.

","absKo":"

최근 연구들은 computer-aided design (CAD)을 위해 large language models (LLMs)를 활용하려는 노력을 점점 더 확대해 왔으며, CAD는 긴 operation sequence 전반에 걸쳐 고도화된 geometric reasoning과 spatial reasoning을 요구하는 도메인이다. 그러나 기존 연구들은 고품질 CAD dataset의 희소성과 세밀한 evaluation framework의 부재로 인해, 단계별 reasoning을 필요로 하는 복잡한 modeling task를 다루는 데 여전히 한계를 보인다. 이러한 문제에 대응하여, 우리는 operation-oriented CAD generation을 위한 최초의 대규모 multi-modal dataset인 Op-CAD를 소개하며, 이는 네 가지 operation type과 다섯 가지 modality를 포괄한다. 더 나아가, 우리는 새로운 CAD parsing module과 geometry-guided hierarchical annotation pipeline을 도입하여 modeling sequence를 개별 operation으로 분해하고, Vision-Language Models (VLMs)의 annotation 정확도를 크게 향상시킨다. 우리의 dataset을 바탕으로, 우리는 geometric 관점과 spatial 관점을 분리하여 CAD modeling task를 재정의하고, chamfer와 fillet operation의 평가 공백을 메우기 위해 Chamfer/Fillet Intersection over Union (CF-IoU)라는 새로운 metric을 제안한다. Op-CAD 위에서 여덟 개의 LLM을 포괄적으로 평가함으로써, 우리는 operation-oriented task에 대한 현재 모델들의 benchmark를 확립한다. 마지막으로, 우리는 Op-CAD에 대한 fine-tuning을 통해 성능 향상 전략을 검토하고, 인간 엔지니어의 reasoning을 모사하는 새로운 prompting strategy인 Chain-of-Operation (COOP)을 제안한다.

"},{"id":"66725","en":"One-shot Entropy Minimization for Language Model Reasoning","ko":"언어 모델 추론을 위한 One-shot Entropy Minimization","authors":"Zitian Gao, Yilong Chen, Haoming Luo, Joey Zhou, Bryan Dai","pos":"#2103","link":"https://openreview.net/forum?id=0yUoCFSUaY","abs":"

In this work, we propose One-shot Entropy Minimization (EM), a simple and fully unsupervised post-training approach that significantly improves reasoning and generation performance using only a single unlabeled data and approximately ten gradient steps. To avoid data contamination, we pretrain a 7-billion-parameter language model from scratch with strictly decontaminated data. Despite its extreme simplicity, one-shot EM yields substantial performance gains and improves reasoning abilities across a broad range of domains, including mathematical reasoning, logical reasoning, and coding. We further show that entropy minimization induces a characteristic right-skewed logit shift, amplifying high-probability tokens while suppressing low-probability tails, in contrast to reinforcement learning. Our findings suggest that entropy minimization primarily acts as a distribution shaping mechanism rather than a conventional learning process, offering an efficient and practical algorithm for post-training large language models.

","absKo":"

본 연구에서는 단 하나의 unlabeled data와 약 10회의 gradient step만으로 reasoning과 generation 성능을 크게 향상시키는, 단순하고 완전히 unsupervised인 post-training 접근법인 One-shot Entropy Minimization(EM)을 제안합니다. data contamination을 피하기 위해, 우리는 엄격하게 decontaminated된 data로 70억 파라미터 언어 모델을 scratch부터 pretrain합니다. 극도로 단순함에도 불구하고, one-shot EM은 substantial performance gain을 제공하며 수학적 추론, 논리적 추론, 코딩을 포함한 광범위한 영역에서 reasoning 능력을 향상시킵니다. 또한 우리는 entropy minimization이 reinforcement learning과는 대조적으로 특징적인 right-skewed logit shift를 유도하여, 저확률 tail을 억제하는 대신 고확률 token을 증폭시킨다는 점을 보입니다. 우리의 결과는 entropy minimization이 전통적인 learning process라기보다 주로 distribution shaping mechanism으로 작동함을 시사하며, large language model의 post-training을 위한 효율적이고 실용적인 알고리즘을 제공합니다.

"},{"id":"61069","en":"Native Parallel Reasoner: Reasoning in Parallelism via Self-Distilled Reinforcement Learning","ko":"Native Parallel Reasoner: Self-Distilled Reinforcement Learning을 통한 병렬 추론","authors":"Tong Wu, Michael Liu, Jun Bai, Zixia Jia, Shuyi Zhang, Ziyong Lin, Yanting Wang, Song-Chun Zhu, Zilong Zheng","pos":"#2113","link":"https://openreview.net/forum?id=tyqW6SYxWB","abs":"We introduce **Native Parallel Reasoner (NPR)**, a teacher-free framework that enables Large Language Models (LLMs) to self-evolve genuine parallel reasoning capabilities. NPR transforms the model from sequential emulation to native parallel cognition through three key innovations: 1) a **self-distilled** progressive training paradigm that transitions from ``cold-start'' format discovery to strict topological constraints without external supervision; 2) a novel **Parallel-Aware Policy Optimization (PAPO)** algorithm that optimizes branching policies directly within the execution graph, allowing the model to learn adaptive decomposition via trial and error; and 3) a robust **NPR Engine** that refactors memory management and flow control of SGLang to enable stable, large-scale parallel RL training. Across eight reasoning benchmarks, NPR trained on Qwen3-4B achieves performance gains of up to 24.5\\% and inference speedups up to 4.6$\\times$. Unlike prior baselines that often fall back to autoregressive decoding, NPR demonstrates 100\\% genuine parallel execution, establishing a new standard for self-evolving, efficient, and scalable agentic reasoning.","absKo":"우리는 **Native Parallel Reasoner (NPR)**를 소개한다. 이는 teacher 없이 Large Language Model(LLM)이 진정한 parallel reasoning capability를 스스로 진화시키도록 하는 framework이다. NPR은 세 가지 핵심 혁신을 통해 model을 sequential emulation에서 native parallel cognition으로 전환한다: 1) external supervision 없이 ``cold-start'' format discovery에서 엄격한 topological constraint로 넘어가는 **self-distilled** progressive training paradigm; 2) execution graph 내부에서 branching policy를 직접 최적화하여, trial and error를 통해 적응적 decomposition을 학습하게 하는 새로운 **Parallel-Aware Policy Optimization (PAPO)** 알고리즘; 3) SGLang의 memory management와 flow control을 재구성하여 안정적인 대규모 parallel RL training을 가능하게 하는 견고한 **NPR Engine**. 여덟 개의 reasoning benchmark 전반에서, Qwen3-4B로 학습된 NPR은 최대 24.5\\%의 성능 향상과 최대 4.6$\\\\times$의 inference speedup을 달성한다. 종종 autoregressive decoding으로 되돌아가는 기존 baseline과 달리, NPR은 100\\% 진정한 parallel execution을 보여주며, self-evolving, efficient, scalable agentic reasoning의 새로운 기준을 세운다."},{"id":"61268","en":"Many-Shot CoT-ICL: Making In-Context Learning Truly Learn","ko":"Many-Shot CoT-ICL: In-Context Learning이 진정으로 학습하게 하기","authors":"Tsz Ting Chung, Lemao Liu, Mo Yu, Dit-Yan Yeung","pos":"#2211","link":"https://openreview.net/forum?id=rzBWwdKik8","abs":"

In-context learning (ICL) adapts large language models (LLMs) to new tasks by conditioning on demonstrations in the prompt without parameter updates. With long-context models, many-shot ICL can use dozens to hundreds of examples and achieve performance comparable to fine-tuning, yet current understanding of its scaling behavior is largely derived from non-reasoning tasks. We study many-shot chain-of-thought in-context learning (CoT-ICL) for reasoning and show that standard many-shot rules do not transfer. Across non-reasoning and reasoning-oriented LLMs and across non-reasoning and reasoning tasks, we find: (i) a setting-dependent scaling effect, where increasing the number of CoT demonstrations is unstable for non-reasoning LLMs and benefits mainly reasoning-oriented LLMs; (ii) similarity-based retrieval helps on non-reasoning tasks but fails on reasoning, since semantic similarity poorly predicts procedural (i.e., CoT) compatibility; and (iii) an order-scaling effect, where performance variance grows with more CoT demonstrations. We interpret these behaviors by viewing many-shot CoT-ICL as in-context test-time learning rather than scaled pattern matching, and suggests two principles: (i) demonstrations should be easy for the target model to understand, and (ii) they should be ordered to support a smooth conceptual progression. Guided by the principle, we propose Curvilinear Demonstration Selection (CDS), a simple ordering method that yields an average 3.81\\% gain across math and narrative reasoning tasks. Overall, our results reframe the long context window from a retrieval buffer into a structured curriculum for in-context test-time learning.

","absKo":"

In-context learning (ICL)은 prompt 안의 demonstrations를 조건으로 하여 parameter update 없이 large language models (LLMs)을 새로운 task에 적응시킨다. long-context model에서는 many-shot ICL이 수십 개에서 수백 개의 예시를 사용할 수 있고 fine-tuning에 필적하는 성능을 달성할 수 있지만, 그 scaling behavior에 대한 현재의 이해는 대부분 non-reasoning task에서 도출되었다. 우리는 reasoning을 위한 many-shot chain-of-thought in-context learning (CoT-ICL)을 연구하며, 표준 many-shot 규칙이 그대로 적용되지 않음을 보인다. non-reasoning 및 reasoning-oriented LLM 전반과 non-reasoning 및 reasoning task 전반에서 우리는 다음을 발견한다: (i) setting-dependent scaling effect, 즉 CoT demonstration 수를 늘릴 때 non-reasoning LLM에서는 불안정하고 주로 reasoning-oriented LLM에만 이점이 있다; (ii) similarity-based retrieval은 non-reasoning task에서는 도움이 되지만 reasoning에서는 실패하는데, 이는 semantic similarity가 procedural(i.e., CoT) compatibility를 잘 예측하지 못하기 때문이다; (iii) order-scaling effect, 즉 CoT demonstration이 많아질수록 성능 분산이 커진다. 우리는 many-shot CoT-ICL을 scaled pattern matching이 아니라 in-context test-time learning으로 보아 이러한 행동을 해석하며, 두 가지 원칙을 제안한다: (i) demonstration은 target model이 이해하기 쉬워야 하고, (ii) demonstration은 매끄러운 개념적 진행을 지원하도록 순서가 배치되어야 한다. 이 원칙에 따라 우리는 Curvilinear Demonstration Selection (CDS)을 제안하는데, 이는 math 및 narrative reasoning task에서 평균 3.81\\%의 성능 향상을 보이는 간단한 ordering 방법이다. 전반적으로, 우리의 결과는 긴 context window를 retrieval buffer가 아니라 in-context test-time learning을 위한 구조화된 curriculum으로 재정의한다.

"},{"id":"65185","en":"EAGer: Entropy-Aware GEneRation for Adaptive Inference-Time Scaling","ko":"EAGer: 적응형 Inference-Time 스케일링을 위한 엔트로피 인지 생성","authors":"Daniel Scalena, Leonidas Zotos, Elisabetta Fersini, Malvina Nissim, Ahmet Üstün","pos":"#2404","link":"https://openreview.net/forum?id=G0Mw5MaPcR","abs":"

With the rise of reasoning language models and test-time scaling methods as a paradigm for improving model performance, substantial computation is often required to generate multiple candidate sequences from the same prompt. This enables exploration of different reasoning paths toward the correct solution, however, allocates the same compute budget for each prompt. Grounded on the assumption that different prompts carry different degrees of complexity, and thus different computation needs, we propose EAGer, a training-free generation method that leverages model uncertainty through token-wise entropy distribution to reduce redundant computation and concurrently improve overall performance. EAGer allows branching to multiple reasoning paths only in the presence of high-entropy tokens, and then reallocates the saved compute budget to the instances where exploration of alternative paths is most needed. We find that across multiple open-source models on complex reasoning benchmarks such as AIME 2025, EAGer can reallocate the budget without accessing target labels, achieving the best efficiency-performance trade-off in terms of reasoning length and Pass@k. When target labels are accessible, EAGer generates up to 65% fewer tokens (hence saving compute) and achieves up to 37% improvement in Pass@k compared to the Full Parallel Sampling. Our results show that EAGer consistently maximizes the efficiency-performance trade-off by enabling dynamic control over computation expenditure.

","absKo":"

reasoning language model과 test-time scaling method가 model performance 향상을 위한 패러다임으로 부상하면서, 동일한 prompt로부터 여러 candidate sequence를 생성하기 위해 상당한 computation이 종종 필요하다. 이는 정답으로 가는 다양한 reasoning path를 탐색할 수 있게 해주지만, 각 prompt에 동일한 compute budget을 할당하게 만든다. 서로 다른 prompt가 서로 다른 수준의 복잡도를 지니고, 따라서 서로 다른 computation 필요량을 가진다는 가정에 기반해, 우리는 token-wise entropy distribution을 통해 model uncertainty를 활용하여 중복 computation을 줄이면서 전체 성능을 동시에 향상시키는 training-free generation method인 EAGer를 제안한다. EAGer는 high-entropy token이 존재할 때만 여러 reasoning path로 branching을 허용하고, 이렇게 절약한 compute budget을 alternative path 탐색이 가장 필요한 instance에 재할당한다. 복잡한 reasoning benchmark인 AIME 2025를 포함한 여러 open-source model에서, EAGer는 target label에 접근하지 않고도 budget을 재분배할 수 있으며, reasoning length와 Pass@k 측면에서 최고의 efficiency-performance trade-off를 달성한다. target label에 접근할 수 있는 경우, EAGer는 Full Parallel Sampling과 비교해 최대 65% 적은 token을 생성해 compute를 절약하고, Pass@k를 최대 37% 향상시킨다. 우리의 결과는 EAGer가 computation expenditure를 동적으로 제어함으로써 efficiency-performance trade-off를 일관되게 극대화함을 보여준다.

"},{"id":"64994","en":"Critique-Guided Distillation for Robust Reasoning via Refinement","ko":"Critique-Guided Distillation: 정제를 통한 강건한 추론","authors":"Berkcan Kapusuzoglu, Supriyo Chakraborty, Zain Sarwar, Michael Lee, Sambit Sahu","pos":"#2507","link":"https://openreview.net/forum?id=HxyRSeT8z2","abs":"Supervised fine-tuning with expert demonstrations often produces models that imitate outputs without internalizing the reasoning processes needed for robust generalization. While critique-based approaches show promise, training models to generate critiques directly, such as Critique Fine-Tuning (CFT), can lead to output-format drift and degradation of general capabilities. We propose $\\textbf{C}$ritique-$\\textbf{G}$uided $\\textbf{D}$istillation (CGD), a training framework that decouples critique consumption from critique generation. During fine-tuning, the student is trained to refine flawed responses conditioned on teacher critiques. CGD treats critiques as a $\\textit{training-time-only}$ supervision signal, encouraging internalization of error-aware reasoning: critiques guide learning but are absent at inference. Across five model families, CGD consistently outperforms CFT and standard distillation on mathematical reasoning benchmarks, yielding 7\\% average improvements and gains of up to +15.0\\% on AMC23 and +12.2\\% on MATH-500. On challenging competition problems such as AIME24 and AIME25, CGD achieves substantially higher Pass@1 and stronger performance at low Pass@k, indicating improved reasoning quality per sample. Importantly, CGD preserves general instruction-following capabilities where CFT degrades significantly ($-$21.3\\% on IFEval). These results position CGD as a practical and compute-efficient intermediate training paradigm for reasoning-centric tasks without introducing inference-time overhead.","absKo":"Expert demonstration을 통한 supervised fine-tuning은 종종 출력만 모방하고 강건한 generalization에 필요한 reasoning process는 내재화하지 못하는 모델을 만든다. critique-based approach는 가능성을 보이지만, Critique Fine-Tuning(CFT)처럼 모델이 critique를 직접 생성하도록 학습시키는 것은 output-format drift와 일반 능력 저하를 초래할 수 있다. 우리는 critique 소비와 critique 생성을 분리하는 training framework인 $\\textbf{C}$ritique-$\\textbf{G}$uided $\\textbf{D}$istillation(CGD)을 제안한다. fine-tuning 동안 student는 teacher critique를 조건으로 결함이 있는 response를 개선하도록 학습된다. CGD는 critique를 $\\textit{training-time-only}$ supervision signal로 취급하여 error-aware reasoning의 내재화를 촉진한다. critique는 학습을 이끌지만 inference에서는 존재하지 않는다. 다섯 개 model family 전반에서 CGD는 수학 reasoning benchmark에서 CFT와 standard distillation을 일관되게 능가하며, 평균 7\\% 향상과 AMC23에서 최대 +15.0\\%, MATH-500에서 +12.2\\%의 향상을 보인다. AIME24와 AIME25 같은 어려운 competition problem에서는 CGD가 훨씬 높은 Pass@1과 낮은 Pass@k에서도 더 강한 성능을 달성하여, 샘플당 reasoning quality가 향상되었음을 시사한다. 중요한 점은 CGD가 CFT에서 크게 저하되는 일반 instruction-following 능력을 보존한다는 것이다(IFEval에서 $-$21.3\\%). 이러한 결과는 CGD를 inference-time overhead를 추가하지 않으면서 reasoning-centric task를 위한 실용적이고 compute-efficient한 intermediate training paradigm으로 자리매김하게 한다."},{"id":"62453","en":"CrispEdit: Low-Curvature Projections for Scalable Non-Destructive LLM Editing","ko":"CrispEdit: 확장 가능한 비파괴적 LLM 편집을 위한 Low-Curvature Projection","authors":"Zarif Ikram, Arad Firouzkouhi, Stephen Tu, Mahdi Soltanolkotabi, Paria Rashidinejad","pos":"#2508","link":"https://openreview.net/forum?id=gTHlB3WK0l","abs":"

A central challenge in large language model (LLM) editing is capability preservation: methods that successfully change targeted behavior can quietly game the editing proxy and corrupt general capabilities, producing degenerate behaviors reminiscent of proxy/reward hacking. We present CrispEdit, a scalable and principled second-order editing algorithm that treats capability preservation as an explicit constraint, unifying and generalizing several existing editing approaches. CrispEdit formulates editing as constrained optimization and enforces the constraint by projecting edit updates onto the low-curvature subspace of the capability-loss landscape. At the crux of CrispEdit is expressing capability constraint via Bregman divergence, whose quadratic form yields the Gauss–Newton Hessian exactly and even when the base model is not trained to convergence. We make this second-order procedure efficient at the LLM scale using Kronecker-factored approximate curvature (K-FAC) and a novel matrix-free projector that exploits Kronecker structure to avoid constructing massive projection matrices. Across standard model-editing benchmarks, CrispEdit achieves high edit success while keeping capability degradation below 1% on average across datasets, significantly improving over prior editors.

","absKo":"

대규모 언어 모델(LLM) editing의 핵심 과제는 capability preservation이다. 즉, target behavior를 성공적으로 변경하는 방법이 editing proxy를 교묘하게 속이고 일반적인 능력을 손상시켜, proxy/reward hacking을 연상시키는 퇴행적 behavior를 만들어낼 수 있다는 점이다. 우리는 capability preservation을 명시적 제약으로 취급하는 확장 가능하고 원칙적인 second-order editing algorithm인 CrispEdit을 제시하며, 여러 기존 editing 접근법을 통합하고 일반화한다. CrispEdit은 editing을 constrained optimization으로 정식화하고, edit update를 capability-loss landscape의 low-curvature subspace에 projection함으로써 제약을 강제한다. CrispEdit의 핵심은 Bregman divergence를 통해 capability constraint를 표현하는 데 있으며, 그 quadratic form은 base model이 수렴하도록 학습되지 않았더라도 Gauss–Newton Hessian을 정확히 도출한다. 우리는 Kronecker-factored approximate curvature(K-FAC)와 Kronecker structure를 활용해 거대한 projection matrix를 만들지 않는 새로운 matrix-free projector를 사용하여 이 second-order procedure를 LLM 규모에서 효율적으로 만든다. 표준 model-editing benchmark 전반에서 CrispEdit은 높은 edit success를 달성하면서 dataset 전반의 평균 capability degradation을 1% 미만으로 유지해, 기존 editor들보다 크게 향상된 성능을 보인다.

"},{"id":"63738","en":"Context Tuning for In-Context Optimization","ko":"In-Context Optimization을 위한 Context Tuning","authors":"Jack Lu, Ryan Teehan, Zhenbang Yang, Mengye Ren","pos":"#2510","link":"https://openreview.net/forum?id=UAJehyOPTS","abs":"

We introduce Context Tuning, a simple and effective method to significantly enhance few-shot adaptation of language models (LLMs) without fine-tuning model parameters. While prompt-based adaptation techniques have demonstrated the effectiveness of lightweight adaptation methods for LLMs, they typically initialize a trainable prompt or prefix with irrelevant tokens for the task at hand. In contrast, Context Tuning initializes the trainable prompt or prefix with task-specific demonstration examples, leveraging the model’s inherent In-Context Learning (ICL) ability to extract relevant information for improved few-shot learning performance. Extensive evaluations on benchmarks such as CrossFit, UnifiedQA, MMLU, BIG-Bench Hard, and ARC demonstrate that Context Tuning outperforms traditional prompt-based adaptation methods and achieves competitive accuracy to Test-Time Training with significantly higher training efficiency.

","absKo":"우리는 model parameter를 fine-tuning하지 않고도 language model(LLM)의 few-shot adaptation을 크게 향상시키는 간단하고 효과적인 방법인 Context Tuning을 제안한다. prompt-based adaptation 기법은 LLM에 대한 경량 adaptation 방법의 효과를 보여주었지만, 일반적으로 trainable prompt 또는 prefix를 현재 task와 무관한 token으로 초기화한다. 이에 반해 Context Tuning은 trainable prompt 또는 prefix를 task-specific demonstration example로 초기화하여, 모델의 고유한 In-Context Learning(ICL) 능력을 활용해 관련 정보를 추출함으로써 few-shot learning 성능을 향상시킨다. CrossFit, UnifiedQA, MMLU, BIG-Bench Hard, ARC와 같은 benchmark에 대한 광범위한 평가에서 Context Tuning은 기존 prompt-based adaptation 방법보다 우수하고, 훨씬 높은 training efficiency로 Test-Time Training에 준하는 경쟁력 있는 정확도를 달성함을 보여준다.

"},{"id":"62884","en":"BLISS: A Lightweight Bilevel Influence Scoring Method for Data Selection in Language Model Pretraining","ko":"BLISS: Language Model 사전학습 데이터 선택을 위한 가벼운 Bilevel Influence Scoring 방법","authors":"Jie Hao, Rui Yu, Wei Zhang, Huixia Judy Wang, Jie Xu, Mingrui Liu","pos":"#2607","link":"https://openreview.net/forum?id=cFWgTYm1hW","abs":"Effective data selection is essential for pretraining large language models (LLMs), enhancing efficiency and improving generalization to downstream tasks. However, existing approaches often require leveraging external pretrained models, making it difficult to disentangle the effects of data selection from those of the external pretrained models. In addition, they often overlook the long-term impact of selected data if the model is trained for a long period of time, primarily due to the prohibitive cost of full-scale LLM pretraining. In this paper, we introduce BLISS (**B**ileve**L** **I**nfluence **S**coring method for data **S**election): a lightweight data selection method that operates entirely \\emph{from scratch}, without relying on any external pretrained oracle models, while explicitly accounting for the long-term impact of selected data. BLISS leverages a small proxy model as a surrogate for the LLM and employs a score model to estimate the long-term influence of training samples if the proxy model is trained to convergence. We formulate data selection as a bilevel optimization problem, where the upper-level objective optimizes the score model to assign importance weights to training samples, ensuring that minimizing the lower-level objective (i.e., training the proxy model over the weighted training loss until convergence) leads to best validation performance. Once optimized, the trained score model predicts influence scores for the dataset, enabling efficient selection of high-quality samples for LLM pretraining. We validate BLISS by pretraining 410M/1B/2.8B Pythia and LLaMA-0.5B models on selected subsets of the C4 dataset. Notably, under the 1B model setting, BLISS achieves $1.7\\times$ speedup in reaching the same performance as the state-of-the-art method, demonstrating superior performance across multiple downstream tasks.","absKo":"대규모 language model (LLM)의 pretraining에서 효과적인 data selection은 효율성을 높이고 downstream task로의 일반화를 개선하는 데 필수적이다. 그러나 기존 접근법은 외부의 pretrained model을 활용해야 하는 경우가 많아, data selection의 효과와 외부 pretrained model의 효과를 분리해서 보기 어렵다. 더불어, 모델이 장기간 학습될 경우 선택된 데이터의 장기적 영향을 종종 간과하는데, 이는 전면적인 LLM pretraining의 비용이 지나치게 크기 때문이다. 본 논문에서는 BLISS(**B**ileve**L** **I**nfluence **S**coring method for data **S**election)를 소개한다. BLISS는 어떠한 외부 pretrained oracle model에도 의존하지 않고 완전히 \\emph{from scratch}로 동작하는 경량 data selection 방법이며, 선택된 데이터의 장기적 영향까지 명시적으로 고려한다. BLISS는 작은 proxy model을 LLM의 surrogate로 활용하고, proxy model이 수렴까지 학습되었을 때 training sample의 장기적 influence를 추정하기 위해 score model을 사용한다. 우리는 data selection을 bilevel optimization problem으로 정식화한다. 여기서 상위 수준 objective는 score model을 최적화하여 training sample에 중요도 가중치를 할당하고, 하위 수준 objective(즉, 가중된 training loss로 proxy model을 수렴까지 학습하는 것)를 최소화하면 best validation performance에 도달하도록 한다. 최적화가 완료되면, 학습된 score model이 dataset에 대한 influence score를 예측하여 LLM pretraining을 위한 고품질 sample을 효율적으로 선택할 수 있게 한다. 우리는 C4 dataset의 선택된 subset에서 410M/1B/2.8B Pythia와 LLaMA-0.5B model을 pretraining하여 BLISS를 검증한다. 특히 1B model 설정에서 BLISS는 state-of-the-art method와 동일한 성능에 도달하는 데 필요한 시간을 1.7배 단축했으며, 여러 downstream task 전반에서 우수한 성능을 보였다."},{"id":"66569","en":"Beyond Logits: Metastable Latent Dynamics for Sample-Efficient Best-of-N Selection in LLMs","ko":"Logit을 넘어서: LLM에서 표본 효율적 Best-of-N 선택을 위한 준안정 잠재 동역학","authors":"Xinrong Li, Zidong Zhou, Keyu Shen, Wenhao Zhou, Shangqi Guo","pos":"#2609","link":"https://openreview.net/forum?id=2W4ecSEJJx","abs":"

Best-of-N selection improves reasoning in large language models (LLMs) by allocating additional test-time compute to sample multiple candidate trajectories, but it fundamentally relies on reliable verification. However, widely used proxies based on logit confidence or sample agreement can suffer from calibration collapse, where confidence becomes misaligned with correctness. Instead, we move beyond output-level signals and analyze the model's latent dynamics during inference. Drawing from cognitive neuroscience, we hypothesize that effective reasoning exhibits \\textit{metastability}—a balance between stability and flexibility manifested as structured ``dwell-and-jump'' dynamics. We introduce Latent Velocity Entropy (LVE), a training-free metric that quantifies these dynamics via the entropy of internal representation updates. Extensive experiments on four reasoning benchmarks (AIME, GPQA, MATH, Brumo) demonstrate that the metric mitigates calibration collapse and consistently outperforms leading logit-based baselines. It surpasses the state-of-the-art baseline (UID) by 1.6\\% and majority voting by 4.0\\% in average accuracy. Remarkably, our method matches the performance of 10-sample majority voting using only 3 samples—a 70\\% reduction in inference cost.

","absKo":"

Best-of-N selection은 여러 candidate trajectory를 샘플링하기 위해 test-time compute를 추가로 할당함으로써 large language model (LLM)의 reasoning을 향상시키지만, 그 핵심은 신뢰할 수 있는 verification에 의존한다. 그러나 logit confidence나 sample agreement에 기반한 널리 쓰이는 proxy는 calibration collapse를 겪을 수 있으며, 이 경우 confidence와 정답 여부가 서로 맞지 않게 된다. 대신 우리는 output-level signal을 넘어 inference 동안 모델의 latent dynamics를 분석한다. cognitive neuroscience에서 착안하여, 우리는 효과적인 reasoning이 \\textit{metastability}, 즉 구조화된 ``dwell-and-jump'' dynamics로 나타나는 안정성과 유연성의 균형을 보인다고 가정한다. 우리는 내부 representation update의 entropy를 통해 이러한 dynamics를 정량화하는 training-free metric인 Latent Velocity Entropy (LVE)를 제안한다. 네 가지 reasoning benchmark (AIME, GPQA, MATH, Brumo)에 대한 광범위한 실험은 이 metric이 calibration collapse를 완화하고 선도적인 logit-based baseline을 일관되게 능가함을 보여준다. 평균 정확도에서 state-of-the-art baseline (UID)보다 1.6\\%, majority voting보다 4.0\\% 향상되었다. 주목할 만하게도, 우리의 방법은 3개의 sample만으로 10-sample majority voting과 동일한 성능을 달성하여 inference cost를 70\\% 줄인다.

"},{"id":"61881","en":"Bridging On-Device and Cloud LLMs for Collaborative Reasoning: A Unified Methodology for Local Routing and Post-Training","ko":"협업 추론을 위한 On-Device와 Cloud LLM 연결: Local Routing과 Post-Training의 통합 방법론","authors":"Wenzhi Fang, Dong-Jun Han, Liangqi Yuan, Evan Chen, Christopher G. Brinton","pos":"#2613","link":"https://openreview.net/forum?id=m0HB67qATB","abs":"

Device-cloud collaboration holds promise for deploying large language models (LLMs), leveraging lightweight on-device models for efficiency while relying on powerful cloud models for superior reasoning. A central challenge in this setting is determining, for each incoming query, whether it should be processed locally or offloaded to the cloud. Existing approaches typically rely on external routers, which often struggle to determine difficulty from the prompt itself, especially for tasks involving complex reasoning. Motivated by this limitation, we propose enabling on-device LLMs to decide internally whether to invoke cloud assistance at inference time, with this capability instilled through reinforcement learning based post-training. Casting on-device LLM post-training as a reward maximization problem, we design hierarchical rewards to encourage local problem solving and judicious cloud offloading. To solve the resulting problem, we develop an algorithm featuring a group-level policy gradient that stabilizes optimization, together with adaptive prompt filtering that provides complementary learning signals to mitigate policy collapse (i.e., exclusive local execution or exclusive cloud offloading). Extensive experiments on on-device-scale LLaMA and Qwen models across multiple reasoning benchmarks show that our method consistently outperforms baselines and significantly narrows the gap to full cloud LLMs.

","absKo":"

device-cloud collaboration은 경량 on-device model을 효율성 측면에서 활용하면서 강력한 cloud model을 통해 우수한 reasoning을 얻을 수 있다는 점에서 large language models(LLMs)의 배치에 유망하다. 이 설정의 핵심 과제는 들어오는 각 query에 대해 local에서 처리할지, 아니면 cloud로 offload할지를 결정하는 것이다. 기존 접근은 대개 external router에 의존하는데, 특히 복잡한 reasoning을 수반하는 task에서는 prompt 자체로부터 난이도를 판별하는 데 어려움을 겪는 경우가 많다. 이러한 한계에 착안하여, 우리는 inference 시점에 on-device LLM이 cloud assistance를 호출할지 여부를 내부적으로 결정하도록 하고, 이 능력을 reinforcement learning 기반 post-training을 통해 주입하는 방법을 제안한다. on-device LLM post-training을 reward maximization problem으로 정식화하고, local problem solving과 신중한 cloud offloading을 유도하는 hierarchical reward를 설계한다. 이 문제를 풀기 위해, 우리는 optimization을 안정화하는 group-level policy gradient와 policy collapse(즉, local execution만 수행하거나 cloud offloading만 수행하는 현상)를 완화하기 위한 보완적 learning signal을 제공하는 adaptive prompt filtering을 결합한 algorithm을 개발한다. 여러 reasoning benchmark에서 on-device-scale LLaMA와 Qwen model을 대상으로 한 광범위한 실험은, 우리의 방법이 baseline을 일관되게 능가하고 full cloud LLM과의 격차를 크게 좁힘을 보여준다.

"},{"id":"65448","en":"$\\phi$-Balancing for Mixture-of-Experts Training","ko":"Mixture-of-Experts 학습을 위한 φ-균형화","authors":"Lizhang Chen, Jonathan Li, Qi Wang, Runlong Liao, Shuozhe Li, Chen Liang, Ni Lao, qiang liu","pos":"#2705","link":"https://openreview.net/forum?id=DZbzIOguz4","abs":"Mixture-of-Experts (MoE) models rely on balanced expert utilization to fully realize their scalability. However, existing load-balancing methods are largely heuristic and operate on mini-batch assignment statistics, introducing bias relative to population-level objectives. We propose $\\phi$-balancing, a principled framework that directly targets population-level expert balance by minimizing a Schur-convex potential of the expected routing distribution. Using convex duality, we derive an equivalent min-max formulation and obtain a simple online algorithm via mirror descent, yielding an efficient EMA-based routing adjustment with negligible overhead. Across large-scale pretraining and downstream fine-tuning, $\\phi$-balancing consistently outperforms prior Switch-style and loss-free baselines, demonstrating more stable and effective expert utilization.","absKo":"Mixture-of-Experts(MoE) 모델은 확장성을 온전히 실현하기 위해 expert 활용의 균형에 의존한다. 그러나 기존 load-balancing 방법은 대체로 휴리스틱에 기반하며 mini-batch 할당 통계에 작동하므로, population-level objective에 비해 편향을 유발한다. 우리는 population-level expert balance를 직접 겨냥하여 기대 routing 분포의 Schur-convex potential을 최소화하는, 원리적인 프레임워크인 $\\phi$-balancing을 제안한다. convex duality를 사용해 동등한 min-max formulation을 유도하고, mirror descent를 통한 간단한 online algorithm을 얻어, 오버헤드가 거의 없는 효율적인 EMA 기반 routing 조정을 가능하게 한다. 대규모 pretraining과 downstream fine-tuning 전반에서 $\\phi$-balancing은 기존 Switch-style 및 loss-free baseline을 일관되게 능가하며, 더 안정적이고 효과적인 expert 활용을 보여준다."},{"id":"61444","en":"DV-World: Benchmarking Data Visualization Agents in Real-World Scenarios","ko":"DV-World: 실제 시나리오에서 Data Visualization Agents 벤치마킹","authors":"Jinxiang Meng, Shaoping Huang, Fangyu Lei, Jingyu Guo, Haoxiang Liu, Jiahao Su, Sihan Wang, Yao Wang, Enrui Wang, Ye Yang, Hongze Chai, Jinming Lyu, Anbang Yu, Huangjing Zhang, Yitong Zhang, Yiming Huang, Zeyao Ma, Shizhu He, Jun Zhao, Kang Liu","pos":"#301","link":"https://openreview.net/forum?id=q2kihJvzqC","abs":"

Real-world data visualization (DV) requires native environmental grounding, cross-platform evolution, and proactive intent alignment. Yet, existing benchmarks often suffer from code-sandbox confinement, single-language creation-only tasks, and assumption of perfect intent. To bridge these gaps, we introduce DV-World, a benchmark of 260 tasks designed to evaluate DV agents across real-world professional lifecycles. DV-World spans three domains: DV-Sheet for native spreadsheet manipulation including chart and dashboard creation as well as diagnostic repair; DV-Evolution for adapting and restructuring reference visual artifacts to fit new data across diverse programming paradigms and DV-Interact for proactive intent alignment with a user simulator that mimics real-world ambiguous requirements. Our hybrid evaluation framework integrates Table-value Alignment for numerical precision and MLLM-as-a-Judge with rubrics for semantic-visual assessment. Experiments reveal that state-of-the-art models achieve less than 50\\% overall performance, exposing critical deficits in handling the complex challenges of real-world data visualization. DV-World provides a realistic testbed to steer development toward the versatile expertise required in enterprise workflows. Data and code are available at \\url{https://anonymous.4open.science/r/DV-World-50D2}.

","absKo":"

실세계 데이터 시각화(DV)는 네이티브 환경 기반화, cross-platform 진화, 그리고 능동적인 intent alignment를 필요로 한다. 그러나 기존 benchmark는 코드 샌드박스에 갇힌 제약, 단일 언어의 생성 전용 과제, 그리고 완벽한 intent를 가정하는 문제를 자주 가진다. 이러한 격차를 메우기 위해, 우리는 실세계의 전문적 생애주기 전반에서 DV agent를 평가하도록 설계된 260개 task로 구성된 benchmark인 DV-World를 소개한다. DV-World는 세 가지 도메인으로 구성된다. DV-Sheet는 차트와 dashboard 생성뿐 아니라 진단적 수정을 포함한 네이티브 spreadsheet 조작을 다루고, DV-Evolution은 서로 다른 프로그래밍 패러다임 전반에서 참조 visual artifact를 새로운 데이터에 맞게 적응 및 재구성하는 과제를 다루며, DV-Interact는 실제의 모호한 요구사항을 모사하는 user simulator와의 proactive intent alignment를 다룬다. 우리의 하이브리드 평가 프레임워크는 수치 정밀도를 위한 Table-value Alignment와, semantic-visual 평가를 위한 rubric 기반 MLLM-as-a-Judge를 통합한다. 실험 결과, state-of-the-art 모델들도 전체 성능이 50\\%에 못 미쳐, 실제 데이터 시각화의 복잡한 도전 과제를 다루는 데 중대한 결함이 있음을 드러낸다. DV-World는 enterprise workflow에 필요한 다재다능한 전문성을 향하도록 개발 방향을 이끄는 현실적인 testbed를 제공한다. 데이터와 코드는 \\url{https://anonymous.4open.science/r/DV-World-50D2}에서 उपलब्ध하다.

"},{"id":"64291","en":"Convex Optimization for Alignment and Preference Learning on a Single GPU","ko":"단일 GPU에서 정렬 및 선호 학습을 위한 볼록 최적화","authors":"Miria Feng, Mert Pilanci","pos":"#3701","link":"https://openreview.net/forum?id=P4eXtzKPrl","abs":"Fine-tuning large language models (LLMs) to align with human preferences has driven the success of systems such as Gemini and ChatGPT. However, approaches like Reinforcement Learning from Human Feedback (RLHF) remain computationally expensive and complex. Direct Preference Optimization (DPO) offers a simpler alternative but has limitations such as inconsistent ranking accuracy, high dependence on expensive GPU resources, and sensitive hyperparameter tuning. We propose the Convex Optimization for Alignment and Preference Learning Algorithm (COALA): a novel lightweight strategy with strong theoretical guarantees. By leveraging the convex optimization reformulation of neural networks, COALA eliminates the need for a reference model and obtains significant reduction in both training time and VRAM consumption, thus enabling efficient training on a single GPU. Experiments across three datasets—including a 23,228-sample synthetic Educational Feedback dataset—and five models (including LLaMA-8B) demonstrate COALA's competitive performance and efficiency in utilizing as little as ${\\approx}17.6$% of DPO's total TFLOPS. COALA exhibits stable, monotonically increasing rewards and reaches peak margins in significantly less time than traditional methods such as DPO and ORPO. To the best of our knowledge, this is the first time convex optimization has been effectively applied to preference fine-tuning of LLMs.","absKo":"대형 언어 모델 (LLMs)을 human preference에 맞추기 위한 fine-tuning은 Gemini와 ChatGPT 같은 시스템의 성공을 이끌어 왔습니다. 그러나 Reinforcement Learning from Human Feedback (RLHF)과 같은 접근법은 계산 비용이 크고 복잡합니다. Direct Preference Optimization (DPO)은 더 단순한 대안이지만, 일관되지 않은 ranking accuracy, 고가의 GPU 자원에 대한 높은 의존성, 민감한 hyperparameter tuning 등의 한계를 가집니다. 우리는 강한 이론적 보장을 갖는 새로운 경량 전략인 Convex Optimization for Alignment and Preference Learning Algorithm (COALA)를 제안합니다. 신경망의 convex optimization 재정식을 활용함으로써 COALA는 reference model이 필요 없으며 training time과 VRAM consumption을 크게 줄여 단일 GPU에서의 효율적인 학습을 가능하게 합니다. 23,228개 샘플로 구성된 synthetic Educational Feedback dataset을 포함한 세 개의 dataset과 LLaMA-8B를 포함한 다섯 개의 model에 대한 실험은 COALA가 경쟁력 있는 성능과 함께 DPO의 총 TFLOPS의 약 ${\\approx}17.6$%만을 사용해도 효율적으로 동작함을 보여줍니다. COALA는 안정적이고 단조 증가하는 reward를 보이며, DPO와 ORPO 같은 전통적 방법보다 훨씬 짧은 시간에 peak margin에 도달합니다. 우리가 아는 한, 이는 convex optimization이 LLM의 preference fine-tuning에 효과적으로 적용된 첫 사례입니다."},{"id":"65772","en":"Deep Pre-Alignment for VLMs","ko":"VLM을 위한 심층 사전 정렬","authors":"Tianyu Yu, Kechen Fang, Zihao Wan, Kaidong Zhang, Yicheng Zhang, Jun Song, Bo Zheng, Yuan Yao","pos":"#3808","link":"https://openreview.net/forum?id=AGBuhcQbsO","abs":"

Most Vision Language Models (VLMs) directly map outputs from ViT encoders to the LLM via a lightweight projector. While effective, recent analysis suggests this architecture suffers from an alignment challenge: visual features remain distant from the text space in the initial layers of the LLM, forcing the model to waste critical depth on superficial modality alignment rather than deep understanding and complex reasoning. In this work, we propose Deep Pre-Alignment (DPA), a novel architecture that replaces the standard ViT encoder with a small VLM as perceiver, ensuring visual features are deeply aligned with the text space of the target large language model. Comprehensive experiments demonstrate the effectiveness of DPA. On the 4B parameter scale, DPA outperforms baselines by 1.9 points across 8 multimodal benchmarks, with gains widening to 3.0 points at the 32B scale. Moreover, by offloading alignment to the perceiver, DPA achieves a 32.9\\% reduction in language capability forgetting over 3 text benchmarks. We further demonstrate that these gains are consistent across different LLM families including Qwen3 and LLaMA 3.2, highlighting the generality of our approach. Beyond performance, DPA also offers a seamless upgrade path for current VLM development, requiring only a modular replacement for the visual encoder with marginal computation overhead.

","absKo":"

대부분의 Vision Language Model(VLM)은 ViT encoder의 출력을 경량 projector를 통해 LLM에 직접 매핑한다. 이 방식은 효과적이지만, 최근 분석에 따르면 이러한 아키텍처는 alignment 문제를 안고 있다. 즉, visual feature가 LLM의 초기 layer에서 text space와 여전히 멀리 떨어져 있어, 모델이 깊은 이해와 복잡한 추론이 아니라 피상적인 modality alignment에 중요한 depth를 낭비하게 된다. 본 연구에서는 Deep Pre-Alignment(DPA)를 제안한다. DPA는 표준 ViT encoder를 perceiver 역할의 작은 VLM으로 대체하여, visual feature가 목표 large language model의 text space와 깊게 aligned되도록 보장하는 새로운 아키텍처이다. 포괄적인 실험은 DPA의 효과를 입증한다. 4B parameter 규모에서 DPA는 8개 multimodal benchmark 전반에서 baseline을 1.9점 앞서며, 32B 규모에서는 그 향상이 3.0점으로 확대된다. 또한 alignment를 perceiver로 오프로딩함으로써, DPA는 3개 text benchmark에서 language capability forgetting을 32.9\\% 줄인다. 더 나아가 이러한 향상은 Qwen3와 LLaMA 3.2를 포함한 다양한 LLM family 전반에서 일관되게 나타나며, 우리의 접근법의 generality를 강조한다. 성능을 넘어, DPA는 현재 VLM 개발을 위한 seamless upgrade path도 제공하며, 시각 encoder의 modular replacement만으로도 거의 증가하지 않는 computation overhead를 요구한다.

"},{"id":"63325","en":"Video-OPD: Efficient Post-Training of Multimodal Large Language Models for Temporal Video Grounding via On-Policy Distillation","ko":"Video-OPD: On-Policy 증류를 통한 시간적 비디오 그라운딩을 위한 멀티모달 대규모 언어 모델의 효율적 사후 학습","authors":"Jiaze Li, Hao Yin, Haoran Xu, Boshen Xu, Wenhui Tan, Zewen He, Jianzhong Ju, Zhenbo Luo, Jian Luan","pos":"#409","link":"https://openreview.net/forum?id=Y4jQjDnVSd","abs":"

Reinforcement learning has emerged as a principled post-training paradigm for Temporal Video Grounding (TVG) due to its on-policy optimization, yet existing GRPO-based methods remain fundamentally constrained by sparse reward signals and substantial computational overhead. We propose Video-OPD, an efficient post-training framework for TVG inspired by recent advances in on-policy distillation. Video-OPD optimizes trajectories sampled directly from the current policy, thereby preserving alignment between training and inference distributions, while a frontier teacher supplies dense, token-level supervision via a reverse KL divergence objective. This formulation preserves the on-policy property critical for mitigating distributional shift, while converting sparse, episode-level feedback into fine-grained, step-wise learning signals. Building on Video-OPD, we introduce Teacher-Validated Disagreement Focusing (TVDF), a lightweight training curriculum that iteratively prioritizes trajectories that are both teacher-reliable and maximally informative for the student, thereby improving training efficiency. Empirical results demonstrate that Video-OPD consistently outperforms GRPO while achieving substantially faster convergence and lower computational cost, establishing on-policy distillation as an effective alternative to conventional reinforcement learning for TVG.

","absKo":"

Reinforcement learning은 on-policy optimization 덕분에 Temporal Video Grounding (TVG)을 위한 정당한 post-training paradigm으로 떠올랐지만, 기존 GRPO 기반 방법은 sparse reward signal과 상당한 계산 오버헤드라는 근본적 제약을 여전히 안고 있다. 우리는 최근 on-policy distillation의 발전에서 영감을 받은 TVG용 효율적 post-training 프레임워크 Video-OPD를 제안한다. Video-OPD는 현재 policy에서 직접 샘플링한 trajectory를 최적화함으로써 training과 inference 분포 간의 정렬을 유지하는 한편, frontier teacher가 reverse KL divergence objective를 통해 dense한 token-level supervision을 제공한다. 이 formulation은 distributional shift를 완화하는 데 핵심적인 on-policy 특성을 보존하면서, sparse한 episode-level feedback을 세밀한 step-wise learning signal로 전환한다. Video-OPD를 기반으로, 우리는 Teacher-Validated Disagreement Focusing (TVDF)을 도입한다. 이는 teacher가 신뢰할 수 있고 student에게 가장 정보량이 큰 trajectory를 반복적으로 우선순위화하는 경량 training curriculum으로, 결과적으로 training 효율을 높인다. 실증 결과는 Video-OPD가 GRPO를 일관되게 능가하면서도 훨씬 빠른 수렴과 더 낮은 계산 비용을 달성함을 보여 주며, on-policy distillation이 TVG를 위한 기존 reinforcement learning의 효과적인 대안임을 입증한다.

"},{"id":"65927","en":"Strategy Executability in Mathematical Reasoning: Leveraging Human–Model Differences for Effective Guidance","ko":"수학 추론에서의 전략 실행 가능성: 효과적 가이던스를 위한 인간-모델 차이 활용","authors":"Weida Liang, Yiyou Sun, Shuyuan Nan, Chuang Li, Dawn Song, Kenji Kawaguchi","pos":"#1801","link":"https://openreview.net/forum?id=8obMsWKFRJ","abs":"Example-based guidance is widely used to improve mathematical reasoning at inference time, yet its effectiveness is highly unstable across problems and models—even when the guidance is correct and problem-relevant. We show that this instability arises from a previously underexplored gap between *strategy usage*—whether a reasoning strategy appears in successful solutions—and *strategy executability*—whether the strategy remains effective when instantiated as guidance for a target model. Through a controlled analysis of paired human-written and model-generated solutions, we identify a systematic dissociation between usage and executability: human- and model-derived strategies differ in structured, domain-dependent ways, leading to complementary strengths and consistent source-dependent reversals under guidance. Building on this diagnosis, we propose *Selective Strategy Retrieval* (SSR), a test-time framework that explicitly models executability by selectively retrieving and combining strategies using empirical, multi-route, source-aware signals. Across multiple mathematical reasoning benchmarks, SSR yields reliable and consistent improvements over direct solving, in-context learning, and single-source guidance, improving accuracy by up to $+13$ points on AIME25 and $+5$ points on Apex for compact reasoning models.","absKo":"예시 기반 guidance는 inference time에 mathematical reasoning을 개선하는 데 널리 사용되지만, guidance가 정확하고 문제와 관련되어 있더라도 문제와 model에 따라 그 효과는 매우 불안정하다. 우리는 이러한 불안정성이 이전에는 충분히 탐구되지 않았던 *strategy usage*와 *strategy executability* 사이의 간극에서 비롯됨을 보인다. *strategy usage*는 reasoning strategy가 성공적인 해답에 나타나는지를, *strategy executability*는 그 strategy가 target model을 위한 guidance로 구현되었을 때도 여전히 효과적인지를 의미한다. 인간이 작성한 해답과 model이 생성한 해답의 짝을 맞춘 통제된 분석을 통해, 우리는 usage와 executability 사이에 체계적인 분리가 있음을 식별한다. 인간과 model에서 유도된 strategy는 구조화되고 domain-dependent한 방식으로 다르며, guidance 하에서 상보적인 강점과 source-dependent reversal을 일관되게 드러낸다. 이러한 진단을 바탕으로, 우리는 경험적 multi-route, source-aware signal을 사용하여 strategy를 선택적으로 검색하고 결합함으로써 executability를 명시적으로 모델링하는 test-time framework인 *Selective Strategy Retrieval*(SSR)을 제안한다. 여러 mathematical reasoning benchmark 전반에서 SSR은 직접 풀이, in-context learning, 단일 source guidance보다 안정적이고 일관된 향상을 제공하며, compact reasoning model에서 AIME25에서 최대 +13점, Apex에서 +5점의 정확도 향상을 달성한다."},{"id":"63630","en":"SuCo: Sufficiency-guided Continuous Adaptive Reasoning","ko":"SuCo: 충분성 유도 연속 적응 추론","authors":"Jiahao Wang, Bingyu Liang, Chenhao Hu, Longhui Zhang, Xuebo Liu, Min zhang, Jing Li, Xuelong Li","pos":"#1802","link":"https://openreview.net/forum?id=VHYljDMosS","abs":"

Despite remarkable performance on complex tasks, Large Reasoning Models (LRMs) often generate excessively long Chain-of-Thoughts (CoT), inflating computational costs even for simple queries. Existing efforts to mitigate this inefficiency typically rely on discrete reasoning modes or fixed budget tiers, lacking a principled criterion of when reasoning is sufficient. In this work, we introduce Minimal Sufficient CoT (MSC), defined as the shortest prefix of a CoT trajectory which is adequate for producing the correct answer. We empirically show that MSC not only reduces reasoning tokens, but also improves accuracy across difficulty levels. Building on MSC, we propose Sufficiency-guided Continuous Adaptive Reasoning (SuCo), a two-stage training framework for autonomous reasoning control along a continuous spectrum. In stage 1, MSC-Aligned Fine-Tuning (MFT) constructs MSC data using problem-adaptive sufficiency thresholds that naturally scale with question difficulty, then fine-tunes the model to internalize concise yet sufficient reasoning patterns. In stage 2, Sufficiency-Aware Policy Optimization (SAPO) further optimizes the model through reinforcement learning with dynamic complexity tracking and sufficiency-aware rewards that penalize both over- and under-thinking. Extensive experiments across mathematics, code, and science benchmarks show that SuCo consistently achieves improvements in both accuracy and reasoning efficiency.

","absKo":"

복잡한 과제에서 뛰어난 성능을 보이지만, Large Reasoning Models (LRMs)는 종종 지나치게 긴 Chain-of-Thoughts (CoT)를 생성하여, 단순한 질의에서도 계산 비용을 부풀린다. 이러한 비효율을 완화하려는 기존 연구는 대개 이산적인 reasoning mode나 고정된 budget tier에 의존해 왔으며, 언제 reasoning이 충분한지에 대한 원칙적인 기준은 부족하다. 이 연구에서는 정확한 답을 생성하기에 충분한 CoT trajectory의 가장 짧은 prefix로 정의되는 Minimal Sufficient CoT (MSC)를 제안한다. 실증적으로 MSC는 reasoning token을 줄일 뿐 아니라, 난이도 전반에 걸쳐 정확도도 향상시킴을 보인다. MSC를 기반으로, 연속적인 스펙트럼 전반에서 자율적인 reasoning 제어를 위한 2단계 학습 프레임워크인 Sufficiency-guided Continuous Adaptive Reasoning (SuCo)를 제안한다. 1단계에서는 MSC-Aligned Fine-Tuning (MFT)이 문제 적응적 sufficiency threshold를 사용해 MSC 데이터를 구성하는데, 이 threshold는 질문 난이도에 따라 자연스럽게 스케일되며, 이어 모델이 간결하면서도 충분한 reasoning pattern을 내재화하도록 fine-tuning한다. 2단계에서는 Sufficiency-Aware Policy Optimization (SAPO)이 동적 복잡도 추적과, 과도한 thinking과 부족한 thinking을 모두 페널티하는 sufficiency-aware reward를 사용한 reinforcement learning으로 모델을 추가 최적화한다. 수학, 코드, 과학 benchmark 전반에 걸친 광범위한 실험은 SuCo가 정확도와 reasoning 효율성 모두에서 일관된 향상을 달성함을 보여준다.

"},{"id":"64470","en":"TRACE: Toulmin-based Reasoning Assessment through Constructive Elements for LLM CoT Evaluation","ko":"TRACE: LLM CoT 평가를 위한 구성 요소 기반 Toulmin 추론 평가","authors":"Yundong Kim, Heyoung Yang","pos":"#1804","link":"https://openreview.net/forum?id=NJ9MZkCLAG","abs":"

Evaluating open-ended outputs from large language models (LLMs) remains challenging due to the absence of ground truth. We introduce TRACE (Toulmin-based Reasoning Assessment through Constructive Elements), a metric that analyzes Chain-of-Thought (CoT) reasoning processes. TRACE integrates Toulmin's argumentation theory with Flavell's metacognitive framework to assess reasoning structure. Experiments on 26.3K QA samples across 7 reasoning models show strong correlation with benchmark accuracy (r = 0.74). Furthermore, TRACE is effective as a reinforcement learning reward signal, outperforming accuracy-only baselines. These results suggest that TRACE serves as a complementary metric for evaluating open-ended outputs.

","absKo":"

ground truth의 부재로 인해 large language model(LLM)의 open-ended output을 평가하는 일은 여전히 어렵다. 우리는 Chain-of-Thought(CoT) reasoning process를 분석하는 지표인 TRACE(Toulmin-based Reasoning Assessment through Constructive Elements)를 소개한다. TRACE는 Toulmin의 argumentation theory와 Flavell의 metacognitive framework를 통합하여 reasoning structure를 평가한다. 7개의 reasoning model에서 26.3K QA sample을 대상으로 한 실험은 benchmark accuracy와 강한 상관관계(r = 0.74)를 보였다. 더 나아가 TRACE는 reinforcement learning reward signal로도 효과적이며, accuracy-only baseline을 능가한다. 이러한 결과는 TRACE가 open-ended output 평가를 위한 보완적 지표로 기능함을 시사한다.

"},{"id":"63181","en":"The Quality-Utility Paradox: Why High-Reward Data Impairs Small Model Reasoning","ko":"품질-유용성 역설: 고보상 데이터가 소형 모델 추론을 저해하는 이유","authors":"Haolong Qian, Xianliang Yang, Ma yinuo, Lirong Che, Feng Lu, Ye Guo, Lei Song, Jiang Bian, Chun Yuan","pos":"#1809","link":"https://openreview.net/forum?id=ZL2XH58JTi","abs":"

Knowledge distillation from powerful reasoning models underpins the development of Small Language Models (SLMs). A prevailing assumption in this paradigm is that training data with higher perceived quality, often defined by rigorous logic and superior reward scores, monotonically enhances downstream performance. In this paper, we identify a counter-intuitive \\textbf{Quality-Utility Paradox} across diverse model families(Qwen2.5, LLaMA-3, DeepSeek): data refined by a superior Synthesis Oracle consistently underperforms the SLM's self-generated Rejection Sampling (RFT) data, despite achieving higher reward scores. We argue that Oracle models introduce an intrinsic representation bias that shifts training data into a distribution incompatible with the target SLM, where the SLM allocates limited computational capacity to stylistic imitation rather than logical reasoning. We utilize a \\textbf{Style-Aligned Refinement} strategy to correct logical errors and strictly preserve the SLM's native syntax. Our experiments demonstrate that maintaining native syntax effectively mitigates syntactic adaptation costs, enabling distilled models to match or even surpass self-generated baselines. These findings underscore the necessity of syntactic alignment and advocate for model-aware reward designs that prioritize distributional compatibility alongside logical rigor. Our datasets and code will be publicly available.

","absKo":"

강력한 reasoning model로부터의 knowledge distillation은 Small Language Models (SLMs) 개발의 토대가 된다. 이 패러다임에서 널리 받아들여지는 가정은, 엄격한 logic과 더 높은 reward score로 정의되는 더 높은 품질의 training data가 downstream performance를 단조롭게 향상시킨다는 것이다. 본 논문에서는 다양한 model family(Qwen2.5, LLaMA-3, DeepSeek) 전반에서 반직관적인 \\textbf{Quality-Utility Paradox}를 식별한다. 더 우수한 Synthesis Oracle로 정제된 data가 더 높은 reward score를 달성함에도 불구하고, SLM이 스스로 생성한 Rejection Sampling (RFT) data보다 일관되게 낮은 성능을 보인다는 점이다. 우리는 Oracle model이 내재적 representation bias를 도입하여 training data를 target SLM과 호환되지 않는 distribution으로 이동시키며, 그 결과 SLM은 제한된 computational capacity를 logical reasoning이 아니라 stylistic imitation에 할당하게 된다고 주장한다. 우리는 logic error를 교정하면서 SLM 고유의 syntax를 엄격히 보존하는 \\textbf{Style-Aligned Refinement} 전략을 사용한다. 실험 결과, native syntax를 유지하는 것이 syntactic adaptation cost를 효과적으로 완화하여 distilled model이 self-generated baseline과 대등하거나 심지어 이를 능가할 수 있음을 보여준다. 이러한 결과는 syntactic alignment의 필요성을 강조하며, logical rigor와 함께 distributional compatibility를 우선시하는 model-aware reward design을 지지한다. 우리의 dataset과 코드는 공개될 예정이다.

"},{"id":"66264","en":"Thinking in Latent Space: Progressive Multimodal Simplification for Visual Reasoning","ko":"잠재 공간에서의 사고: 시각 추론을 위한 점진적 멀티모달 단순화","authors":"Yuesen Tang, Yiming Yang, Tengfei Bao, Yu Tong","pos":"#1811","link":"https://openreview.net/forum?id=5IA4cx8Gzz","abs":"

Recent Multimodal Large Language Models (MLLMs) have advanced cross-modal reasoning by extending Chain-of-Thought (CoT) prompting to visual tasks. However, existing methods still rely heavily on explicit textual reasoning steps, leading to information loss, unstable perception–reasoning interaction, and high computational cost. Inspired by human cognition, we argue that effective visual reasoning emerges from a dynamic interplay between perception and latent thought, rather than a purely linear verbalization process. Motivated by this insight, we propose Latent-Driven Progressive Visual Reasoning (LDPVR), a framework that formulates multimodal reasoning as a Markov Chain of Recursive State Simplification, where explicit textual states are progressively refined under the guidance of latent transitions. Central to LDPVR is Interleaved Latent Grounding, which leverages latent semantic intent to actively retrieve fine-grained visual evidence and drive robust state evolution, enabling the model to iteratively reduce uncertainty before committing to simplified textual states. To optimize this process, we introduce a three-stage curriculum combining supervised fine-tuning, latent-text distillation, and reinforcement learning via Group Relative Policy Optimization (GRPO). Experiments on six multimodal reasoning benchmarks demonstrate that LDPVR improves reasoning accuracy while maintaining low inference latency. Code will be made public.

","absKo":"

최근 Multimodal Large Language Models (MLLMs)은 Chain-of-Thought (CoT) prompting을 visual task로 확장함으로써 cross-modal reasoning을 발전시켜 왔다. 그러나 기존 방법은 여전히 명시적인 textual reasoning step에 크게 의존하여 정보 손실, 불안정한 perception–reasoning 상호작용, 그리고 높은 computational cost를 초래한다. 인간의 인지에서 영감을 받아, 우리는 효과적인 visual reasoning이 단순히 선형적인 verbalization 과정이 아니라, perception과 latent thought 사이의 동적 상호작용에서 나타난다고 주장한다. 이러한 통찰에 동기를 부여받아, 우리는 multimodal reasoning을 Recursive State Simplification의 Markov Chain으로 정식화하는 framework인 Latent-Driven Progressive Visual Reasoning (LDPVR)을 제안한다. 여기서 explicit textual state는 latent transition의 guidance 아래 점진적으로 정제된다. LDPVR의 핵심은 Interleaved Latent Grounding으로, latent semantic intent를 활용하여 세밀한 visual evidence를 능동적으로 검색하고 견고한 state evolution을 유도함으로써, 모델이 단순화된 textual state로 확정하기 전에 불확실성을 반복적으로 줄일 수 있게 한다. 이 과정을 최적화하기 위해, 우리는 supervised fine-tuning, latent-text distillation, 그리고 Group Relative Policy Optimization (GRPO)을 통한 reinforcement learning을 결합한 3단계 curriculum을 도입한다. 6개의 multimodal reasoning benchmark에서의 실험은 LDPVR이 낮은 inference latency를 유지하면서 reasoning accuracy를 향상시킴을 보여준다. 코드는 공개될 예정이다.

"},{"id":"60937","en":"Small Generalizable Prompt Predictive Models Can Steer Efficient RL Post-Training of Large Reasoning Models","ko":"작고 일반화 가능한 Prompt Predictive Model이 Large Reasoning Model의 효율적인 RL Post-Training을 조정할 수 있다","authors":"Yun Qu, Qi Wang, Yixiu Mao, Heming Zou, Yuhang Jiang, Weijie Liu, Clive Bai, Kai Yang, Yangkun Chen, Saiyong Yang, Xiangyang Ji","pos":"#1901","link":"https://openreview.net/forum?id=vCHOBek0dv","abs":"

Reinforcement learning enhances the reasoning capabilities of large language models but often involves high computational costs due to rollout-intensive optimization. Online prompt selection presents a plausible solution by prioritizing informative prompts to improve training efficiency. However, current methods either depend on costly, exact evaluations or construct prompt-specific predictive models lacking generalization across prompts. This study introduces Generalizable Predictive Prompt Selection (GPS), which performs Bayesian inference towards prompt difficulty using a lightweight generative model trained on the shared optimization history. Intermediate-difficulty prioritization and history-anchored diversity are incorporated into the batch acquisition principle to select informative prompt batches. The small predictive model also generalizes at test-time for efficient computational allocation. Experiments across varied reasoning benchmarks indicate GPS's substantial improvements in training efficiency, final performance, and test-time efficiency over superior baseline methods. The code is available at https://anonymous.4open.science/r/GPS-ICML.

","absKo":"

Reinforcement learning은 large language model의 reasoning 능력을 향상시키지만, rollout-intensive optimization 때문에 종종 높은 계산 비용을 수반한다. Online prompt selection은 정보성이 높은 prompt를 우선시함으로써 training efficiency를 개선하는 유망한 해법이다. 그러나 기존 방법은 비용이 큰 정확한 평가에 의존하거나, prompt별 predictive model을 구성하지만 prompt 간 generalization이 부족하다. 본 연구는 공유된 optimization history 위에서 학습된 가벼운 generative model을 사용해 prompt difficulty에 대한 Bayesian inference를 수행하는 Generalizable Predictive Prompt Selection(GPS)을 제안한다. 중간 난이도 우선순위와 history-anchored diversity는 batch acquisition principle에 통합되어 정보성 있는 prompt batch를 선택하는 데 사용된다. 이 작은 predictive model은 test-time에서도 generalize되어 효율적인 computational allocation을 가능하게 한다. 다양한 reasoning benchmark에 걸친 실험은, GPS가 우수한 baseline method들에 비해 training efficiency, final performance, test-time efficiency를 크게 향상시킴을 보여준다. Code는 https://anonymous.4open.science/r/GPS-ICML 에서 제공된다.

"},{"id":"65846","en":"Search for Truth from Reasoning: A Dynamic Representation Editing Framework for Steering LLM Trajectories","ko":"추론으로부터 진실 탐색: LLM 궤적 조종을 위한 동적 표현 편집 프레임워크","authors":"Tianlong Wang, Yuhang Wang, Weibin Liao, Xin Gao, Xinyu Ma, Yang Lin, Yasha Wang, Liantao Ma","pos":"#1904","link":"https://openreview.net/forum?id=9WIqPb44br","abs":"

Current approaches to enhance Large Language Model (LLM) reasoning, such as Chain-of-Thought and \"Wait\" prompts, primarily encourage models to think more, yet often fail to guide them toward Truth. While Representation Editing (RepE) offers a intrinsic control, its application to dynamic reasoning trajectories remains underexplored. In this work, we bridge this gap by investigating the geometry of truth within unfolding reasoning chains. We uncover three critical insights: (1) Truth is encoded at the sentence level and is entangled with latent reasoning patterns; (2) Effective intervention follows an Uncertainty Principle and a Decay Effect, requiring localization to early, high-entropy forks; (3) Naive steering vectors suffer from noise, risking collateral damage to correct trajectories. Based on these findings, we propose DynaSteer, a dynamic RepE framework. DynaSteer employs pattern clustering to disentangle reasoning manifolds and utilizes Fisher-LDA to project purified truth. By dynamically monitoring lookahead entropy, it selectively steers and rolls back trajectories only when necessary. Comprehensive experimental results on several MATH benchmark verify the effectiveness of DynaSteer, and experiments on out-of-domain coding tasks further confirm its generalization ability. Our code is publicly available at https://anonymous.4open.science/r/DynaSteer.

","absKo":"

Chain-of-Thought와 \"Wait\" prompt와 같은 Large Language Model (LLM) reasoning 향상 기법은 주로 모델이 더 오래 생각하도록 유도하지만, 종종 Truth로 이끄는 데는 실패한다. Representation Editing (RepE)은 intrinsic control을 제공하지만, dynamic reasoning trajectory에 대한 적용은 아직 충분히 탐구되지 않았다. 본 연구에서는 unfolding reasoning chain에서 truth의 geometry를 조사함으로써 이 공백을 메운다. 우리는 세 가지 핵심 통찰을 발견했다. (1) Truth는 sentence level로 인코딩되며 latent reasoning pattern과 얽혀 있다. (2) 효과적인 intervention은 Uncertainty Principle과 Decay Effect를 따른다. 즉, 초기의 high-entropy fork에 국소화되어야 한다. (3) Naive steering vector는 noise로 인해 오염되기 쉬우며, 올바른 trajectory에 부수적 피해를 줄 위험이 있다. 이러한 발견을 바탕으로, 우리는 dynamic RepE framework인 DynaSteer를 제안한다. DynaSteer는 pattern clustering을 사용해 reasoning manifold를 분리하고 Fisher-LDA를 활용해 정제된 truth를 투영한다. 또한 lookahead entropy를 동적으로 모니터링하여, 필요할 때에만 trajectory를 선택적으로 steering하고 rollback한다. 여러 MATH benchmark에서의 포괄적인 실험은 DynaSteer의 효과를 검증하며, out-of-domain coding task에서의 실험은 그 generalization ability를 추가로 확인한다. 코드는 https://anonymous.4open.science/r/DynaSteer 에 공개되어 있다.

"},{"id":"64164","en":"Scaling Prompt Synthesis for Large Language Model Reasoning","ko":"대형 언어 모델 추론을 위한 프롬프트 합성 확장","authors":"Xueliang Zhao, Wei Wu, Jian Guan, Zhuocheng Gong, Lingpeng Kong","pos":"#1906","link":"https://openreview.net/forum?id=QKFzXp5OLv","abs":"

Large language models (LLMs) are evolving from conversational systems into strong reasoners for tasks such as Olympiad mathematics and competitive programming. While scaling parameters and test-time computation has driven progress, a key bottleneck is the lack of high-quality training problems: human-curated datasets are costly and limited, while existing synthetic corpora are often too easy or narrow. PromptCoT showed that injecting rationales into prompt synthesis increases problem difficulty. Building on this, we present PromptScale, a scalable framework that replaces hand-crafted heuristics with an expectation-maximization (EM) loop, where rationales are iteratively refined to guide prompt construction. This produces problems that are both harder and more diverse than prior corpora. The synthetic prompts support two post-training regimes: (1) \\emph{Self-Play}, where strong models improve autonomously via verifiable feedback without stronger teachers; and (2) \\emph{Supervised Fine-Tuning (SFT)}, where weaker models learn from teacher-distilled traces. Extensive experiments demonstrate the effectiveness of this approach. In self-play, applying PromptScale to \\texttt{Qwen3-30B-A3B-Thinking-2507} sets new state-of-the-art results \\emph{at the 30B scale}, with +4.4, +4.8, and +5.3 on AIME 24/25 and HMMT 25, +6.1 and +5.0 on LiveCodeBench v5/v6, and +35 Elo on Codeforces. In SFT, training \\texttt{Qwen2.5-7B-Instruct} solely on synthetic prompts boosts accuracy to 73.1 (AIME 24), 65.6 (AIME 25), and 53.4 (LiveCodeBench v5), surpassing models trained on human or hybrid data. Analyses further confirm that PromptScale yields fundamentally harder and distributionally distinct problems. These results establish prompt synthesis as a new axis for scaling reasoning and position PromptScale as a scalable foundation for future open-source models.

","absKo":"

대규모 언어 모델(LLM)은 대화형 시스템에서 올림피아드 수학, 경쟁 프로그래밍과 같은 과제의 강력한 추론기로 진화하고 있다. 파라미터 규모와 테스트 시점 연산을 늘리는 방식이 진전을 이끌어 왔지만, 핵심 병목은 고품질 학습 문제의 부족이다. 인간이 선별한 데이터셋은 비용이 많이 들고 제한적이며, 기존 합성 코퍼스는 종종 너무 쉽거나 범위가 지나치게 좁다. PromptCoT는 추론을 프롬프트 합성에 주입하면 문제 난도가 높아진다는 점을 보였다. 이를 바탕으로, 우리는 수작업 휴리스틱을 expectation-maximization (EM) 루프로 대체하는 확장 가능한 프레임워크 PromptScale을 제안한다. 이 루프에서는 추론이 반복적으로 정교화되어 프롬프트 구성을 이끈다. 그 결과 이전 코퍼스보다 더 어렵고 더 다양한 문제를 생성할 수 있다. 이 합성 프롬프트는 두 가지 post-training regime을 지원한다. (1) \\emph{Self-Play}: 더 강한 teacher 없이 검증 가능한 피드백만으로 강력한 모델이 자율적으로 향상된다. (2) \\emph{Supervised Fine-Tuning (SFT)}: 약한 모델이 teacher-distilled trace로부터 학습한다. 광범위한 실험은 이 접근의 효과를 입증한다. self-play에서는 \\texttt{Qwen3-30B-A3B-Thinking-2507}에 PromptScale을 적용하여 \\emph{30B 규모에서} 새로운 state-of-the-art 결과를 달성했으며, AIME 24/25와 HMMT 25에서 각각 +4.4, +4.8, +5.3, LiveCodeBench v5/v6에서 +6.1과 +5.0, Codeforces에서 +35 Elo를 기록했다. SFT에서는 \\texttt{Qwen2.5-7B-Instruct}를 합성 프롬프트만으로 학습했을 때 정확도가 73.1 (AIME 24), 65.6 (AIME 25), 53.4 (LiveCodeBench v5)까지 향상되어, 인간 데이터 또는 하이브리드 데이터로 학습한 모델을 능가했다. 추가 분석은 PromptScale이 본질적으로 더 어렵고 분포적으로도 구별되는 문제를 생성함을 확인한다. 이 결과는 프롬프트 합성을 추론 스케일링의 새로운 축으로 자리매김하게 하며, PromptScale을 향후 open-source 모델을 위한 확장 가능한 기반으로 제시한다.

"},{"id":"65071","en":"Perceptual Flow Network for Visually Grounded Reasoning","ko":"시각적으로 기반한 추론을 위한 Perceptual Flow Network","authors":"Yangfu Li, Yuning Gong, Hongjian Zhan, Teng Li, Yuanhuiyi Lyu, Tianyi Chen, Qi Liu, Ziyuan Huang, Zhihang Zhong, DanDan Zheng, Yue Lu","pos":"#2001","link":"https://openreview.net/forum?id=H1ZEo8l2sH","abs":"

Despite the success of LVLMs, general optimization objectives (e.g., standard MLE) fail to constrain visual trajectories, leading to language bias and hallucination. To mitigate this, current methods introduce geometric priors from visual experts as additional supervision. However, we observe that such supervision is typically suboptimal: it is biased toward geometric precision and offers limited reasoning utility. To bridge this gap, we propose Perceptual Flow Network (PFlowNet), which eschews rigid alignment with the expert priors and achieves interpretable yet more effective visual reasoning. Specifically, PFlowNet decouples perception from reasoning to establish a self-conditioned generation process. Based on this, it integrates multi-dimensional rewards with vicinal geometric shaping via variational reinforcement learning, thereby facilitating reasoning-oriented perceptual behaviors while preserving visual reliability. PFlowNet delivers a provable performance guarantee and competitive empirical results, particularly setting new SOTA records on V* Bench (90.6%) and MME-RealWorld-lite (67.0%).

","absKo":"

LVLMs의 성공에도 불구하고, 일반적인 최적화 목적함수(예: standard MLE)는 visual trajectory를 제약하지 못해 language bias와 hallucination을 초래한다. 이를 완화하기 위해, 현재의 방법들은 visual expert로부터 얻은 geometric prior를 추가 supervision으로 도입한다. 그러나 우리는 이러한 supervision이 대체로 최적이 아님을 관찰했다: 이는 geometric precision 쪽으로 편향되어 있으며 reasoning utility는 제한적이다. 이 간극을 메우기 위해, 우리는 Perceptual Flow Network(PFlowNet)를 제안한다. PFlowNet은 expert prior와의 경직된 정렬을 지양하며, 해석 가능하면서도 더 효과적인 visual reasoning을 달성한다. 구체적으로, PFlowNet은 perception과 reasoning을 분리하여 self-conditioned generation process를 구축한다. 이를 바탕으로, variational reinforcement learning을 통해 multi-dimensional rewardsvicinal geometric shaping을 통합함으로써 visual reliability는 유지하면서 reasoning-oriented perceptual behavior를 촉진한다. PFlowNet은 증명 가능한 성능 보장을 제공하며 경쟁력 있는 실험 결과를 보이는데, 특히 V* Bench(90.6%)와 MME-RealWorld-lite(67.0%)에서 새로운 SOTA 기록을 수립한다.

"},{"id":"64047","en":"Reasoning Compartmentalization: Bridging the Concretization Gap via Abstraction-based Routing","ko":"추론 구획화: 추상화 기반 라우팅을 통한 구체화 격차 해소","authors":"Ling-I Wu, Jian Tong, Yu Sun, Xuan Gao, Xu Guo, Guoqiang Li, Qipeng Guo, Kai Chen","pos":"#2012","link":"https://openreview.net/forum?id=ROjScDyMtq","abs":"

While previous research has documented the sensitivity of Large Language Models (LLMs) to surface-level performance degradation, the underlying impact on internal representations and learning dynamics remains under-explored. In this work, we study this question using a controlled setup with paired reasoning tasks that are logically identical but expressed either in an abstract formal language (FL) or in natural language (NL). We find that converting FL problems into NL consistently degrades reasoning accuracy. More importantly, we show that FL and NL inputs activate largely separate internal representations and exhibit weak learning transfer between them. We refer to this phenomenon as reasoning compartmentalization. To test whether this compartmentalization can be mitigated, we introduce abstraction-based alignment, where models are trained to translate NL inputs into their corresponding FL forms. While this significantly improves reasoning performance, FL and NL representations remain largely distinct, and learning transfer across formulations remains limited. Through activation-level interventions, we further show that performance improvements arise not from representational fusion, but from improved routing. This suggests that abstraction alleviates formulation sensitivity by strengthening connections between formulation-specific reasoning pathways, rather than by aligning their representations.

","absKo":"이전 연구들은 Large Language Models (LLMs)가 표면적인 성능 저하에 민감하다는 점을 문서화해 왔지만, 내부 representation과 learning dynamics에 미치는 근본적 영향은 여전히 충분히 탐구되지 않았다. 본 연구에서는 논리적으로 동일하지만, 추상적인 formal language (FL) 또는 natural language (NL)로 표현된 paired reasoning task를 사용하는 통제된 설정에서 이 질문을 분석한다. 우리는 FL 문제를 NL로 변환하면 reasoning accuracy가 일관되게 저하됨을 발견한다. 더 중요한 점은, FL과 NL input이 대체로 분리된 내부 representation을 활성화하며, 둘 사이의 learning transfer가 약하다는 것을 보인다는 것이다. 우리는 이러한 현상을 reasoning compartmentalization이라고 부른다. 이 compartmentalization을 완화할 수 있는지 검증하기 위해, 우리는 모델이 NL input을 해당 FL 형식으로 번역하도록 학습하는 abstraction-based alignment를 도입한다. 이는 reasoning performance를 유의미하게 향상시키지만, FL과 NL representation은 여전히 대체로 구별되며, 서로 다른 formulation 간 learning transfer도 제한적이다. activation 수준의 개입을 통해, 우리는 성능 향상이 representation fusion에서 비롯된 것이 아니라 improved routing에서 비롯됨을 추가로 보인다. 이는 abstraction이 representation을 정렬해서가 아니라, formulation-specific reasoning pathway 간 연결을 강화함으로써 formulation sensitivity를 완화함을 시사한다.

"},{"id":"61954","en":"PLANTAIN: Plan-Answer Interleaved Reasoning","ko":"PLANTAIN: Plan-Answer 교차 추론","authors":"Anthony Liang, Jonathan Berant, Adam Fisch, Abhimanyu Goyal, Kalpesh Krishna, Jacob Eisenstein","pos":"#2100","link":"https://openreview.net/forum?id=lK2o9OjoXf","abs":"Reasoning models often spend a lot of time thinking before they generate a visible response. This creates a frustrating, but unfortunately common, experience: the user's time is wasted while the model reasons from a false premise that could have easily been corrected. In contrast, human speakers perform lightweight, incremental check-ins to ensure that conversational participants stay on common ground. With this motivation, we propose \\textit{interleaved reasoning} (IR), in which the model alternates between thinking and surfacing intermediate responses, as an alternative to the standard ``think-then-answer'' approach. By providing useful information to the user earlier, IR reduces perceived latency, the time a user waits for an initial output, without compromising the quality of the final response. We focus on a specialization of interleaved reasoning, \\method (\\textul{Plan}-\\textul{T}hought-\\textul{A}nswer \\textul{In}terleaving), where the first intermediate response is an explicit, step-by-step \\textit{plan} for executing the task. This plan-first strategy allows for user intervention and early feedback for subsequent reasoning steps. \\method\\ yields an $\\sim$6\\% improvement in pass@1 across several challenging math reasoning and coding benchmarks, while reducing time-to-first-response by over 60\\% relative to think-then-answer baselines.","absKo":"Reasoning model은 종종 눈에 보이는 응답을 생성하기 전에 오랜 시간 생각을 한다. 이는 답답하지만 안타깝게도 흔한 경험을 만든다. 즉, 사용자의 시간이 낭비되는 동안 모델은 쉽게 바로잡을 수 있었던 잘못된 전제 위에서 추론한다. 반면 사람 화자는 대화 참여자들이 common ground를 유지하도록 가벼운 incremental check-in을 수행한다. 이러한 동기를 바탕으로 우리는 표준적인 ``think-then-answer'' 접근의 대안으로, 모델이 thinking과 중간 응답 노출을 번갈아 수행하는 \\textit{interleaved reasoning} (IR)을 제안한다. 유용한 정보를 사용자에게 더 일찍 제공함으로써 IR은 perceived latency, 즉 사용자가 초기 출력을 기다리는 시간을 줄이면서도 최종 응답의 품질을 저하시키지 않는다. 우리는 interleaved reasoning의 한 특수화인 \\method (\\textul{Plan}-\\textul{T}hought-\\textul{A}nswer \\textul{In}terleaving)에 초점을 맞추며, 여기서 첫 번째 중간 응답은 task를 수행하기 위한 명시적인 step-by-step \\textit{plan}이다. 이러한 plan-first 전략은 이후 reasoning 단계들에 대해 사용자 개입과 조기 피드백을 가능하게 한다. \\method\\은 여러 어려운 수학 추론 및 코딩 benchmark에서 pass@1을 약 6\\% 향상시키는 동시에, think-then-answer baseline에 비해 time-to-first-response를 60\\% 이상 줄인다."},{"id":"63850","en":"On the Interplay of Pre-Training, Mid-Training, and RL on Reasoning Language Models","ko":"추론 언어 모델에서 pre-training, mid-training, RL의 상호작용","authors":"Charlie Zhang, Graham Neubig, Xiang Yue","pos":"#2105","link":"https://openreview.net/forum?id=TBaUfO9znF","abs":"

Recent reinforcement learning (RL) techniques have yielded impressive reasoning improvements in language models, yet it remains unclear whether post-training truly extends a model’s reasoning ability beyond what it acquires during pre-training. A central challenge is the lack of control in modern training pipelines: large-scale pre-training corpora are opaque, mid-training is often underexamined, and RL objectives interact with unknown prior knowledge in complex ways. To resolve this ambiguity, we develop a fully controlled experimental framework that isolates the causal contributions of pre-training, mid-training, and RL-based post-training. Our approach employs synthetic reasoning tasks with explicit atomic operations, parseable step-by-step reasoning traces, and systematic manipulation of training distributions. We evaluate models along two axes: extrapolative generalization to more complex compositions and contextual generalization across surface contexts. Using this framework, we reconcile competing views on RL’s effectiveness. We show that: 1) RL produces true capability gains (pass@128) only when pre-training leaves sufficient headroom and when RL data target the model’s edge of competence, tasks at the boundary that are difficult but not yet out of reach. 2) Contextual generalization requires minimal yet sufficient pre-training exposure, after which RL can reliably transfer. 3) Mid-training significantly enhances performance under fixed compute compared with RL only, demonstrating its central but underexplored role in training pipelines. 4) Process-level rewards reduce reward hacking and improve reasoning fidelity. Together, these results clarify the interplay between pre-training, mid-training, and RL, offering a foundation for understanding and improving reasoning LM training strategies.

","absKo":"

최근 reinforcement learning (RL) 기법은 language model의 추론 능력을 인상적으로 향상시켰지만, post-training이 실제로 pre-training 동안 획득한 것을 넘어 추론 능력을 확장하는지 여부는 여전히 불분명하다. 핵심 과제는 현대 training pipeline의 통제 부족이다. 대규모 pre-training corpus는 불투명하고, mid-training은 종종 충분히 검토되지 않으며, RL objective는 알려지지 않은 사전 지식과 복잡하게 상호작용한다. 이 모호성을 해소하기 위해, 우리는 pre-training, mid-training, RL 기반 post-training의 인과적 기여를 분리하는 완전히 통제된 실험 framework를 개발한다. 우리의 접근법은 명시적 atomic operation, parse 가능한 step-by-step reasoning trace, 그리고 training distribution의 체계적 조작을 갖춘 synthetic reasoning task를 사용한다. 우리는 모델을 두 축에서 평가한다. 더 복잡한 composition으로의 extrapolative generalization과, surface context 전반의 contextual generalization이다. 이 framework를 사용하여 우리는 RL의 효과성에 대한 상충된 관점을 조화시킨다. 우리는 다음을 보인다. 1) RL은 pre-training이 충분한 여지를 남기고 RL data가 모델 능력의 경계, 즉 어렵지만 아직 도달 불가능하지 않은 task를 목표로 할 때에만 실제 capability gain (pass@128)을 만든다. 2) contextual generalization은 최소하지만 충분한 pre-training 노출을 필요로 하며, 그 이후 RL은 안정적으로 transfer할 수 있다. 3) mid-training은 고정 compute에서 RL only보다 성능을 크게 향상시키며, training pipeline에서 중심적이지만 충분히 탐구되지 않은 역할을 수행함을 보여준다. 4) process-level reward는 reward hacking을 줄이고 reasoning fidelity를 향상시킨다. 종합하면, 이러한 결과는 pre-training, mid-training, RL의 상호작용을 명확히 하며, reasoning LM training strategy를 이해하고 개선하기 위한 기반을 제공한다.

"},{"id":"64487","en":"Modeling Hierarchical Thinking in Large Reasoning Models","ko":"대형 추론 모델에서 계층적 사고 모델링","authors":"G M Shahariar, Erfan Shayegani, Ali Nazari, Nael Abu-Ghazaleh","pos":"#2115","link":"https://openreview.net/forum?id=N44P3zrrgO","abs":"Large Reasoning Models (LRMs) solve complex tasks by generating long Chain-of-Thought (CoT) sequences; however, the emergent dynamics governing reasoning trajectories are not well understood and can lead to inconsistencies and reasoning pathologies. In this work, we propose to approximate LRM's emerging hierarchical reasoning dynamics as a trajectory within a Finite State Machine (FSM) transitioning among six abstract cognitive states. We demonstrate that these states and transitions can be captured in the latent state of the model. We believe that this representation can have different applications in the interpretability and optimization of LRM models. For example, by analyzing the topology of these transitions, we identify statistical shifts in reasoning strategies that help identify effective reasoning chains from those that fail. To illustrate these potential advantages, we propose $Q$-Value guided steering, a training-free inference-time control method that treats reasoning as a planning problem. We estimate the long-horizon utility of state transitions and apply sparse, orthogonal activation steering at sentence boundaries to align the CoT generation with optimal reasoning policies. Experiments across four benchmarks (AIME25, MATH-500, GSM8k, and GPQA Diamond) using three state-of-the-art open reasoning models demonstrate that $Q$-Value steering policy achieves significant performance gains with \"surgical'' efficiency, often requiring $25\\times$ fewer interventions than greedy and weighted baselines, which suggests that reasoning can be effectively controlled by guiding high-level cognitive dynamics rather than micro-managing token generation.","absKo":"대규모 추론 모델(LRMs)은 긴 Chain-of-Thought (CoT) sequence를 생성함으로써 복잡한 task를 해결한다. 그러나 추론 trajectory를 지배하는 emergent dynamics는 충분히 이해되지 않았으며, 불일치와 reasoning pathology를 초래할 수 있다. 이 연구에서 우리는 LRM의 emergent hierarchical reasoning dynamics를 여섯 개의 추상적 cognitive state 사이를 전이하는 Finite State Machine (FSM) 내의 trajectory로 근사할 것을 제안한다. 우리는 이러한 state와 transition이 모델의 latent state에서 포착될 수 있음을 보인다. 우리는 이 표현이 LRM 모델의 interpretability와 optimization에서 다양한 응용을 가질 수 있다고 본다. 예를 들어, 이러한 transition의 topology를 분석함으로써, 우리는 실패하는 reasoning chain과 효과적인 reasoning chain을 구분하는 데 도움이 되는 reasoning strategy의 statistical shift를 식별한다. 이러한 잠재적 이점을 보여주기 위해, 우리는 추론을 planning problem으로 취급하는 training-free inference-time control method인 $Q$-Value guided steering을 제안한다. 우리는 state transition의 long-horizon utility를 추정하고, sentence boundary에서 sparse, orthogonal activation steering을 적용하여 CoT 생성을 최적의 reasoning policy에 맞춘다. 세 개의 최첨단 open reasoning model을 사용한 네 개의 benchmark(AIME25, MATH-500, GSM8k, GPQA Diamond) 전반의 실험은 $Q$-Value steering policy가 \"surgical''한 효율성으로 상당한 성능 향상을 달성하며, greedy 및 weighted baseline보다 종종 $25\\times$ 더 적은 개입만 필요로 함을 보여준다. 이는 token generation을 미시적으로 관리하기보다 고수준 cognitive dynamics를 유도함으로써 reasoning을 효과적으로 제어할 수 있음을 시사한다."},{"id":"64433","en":"EDCO: Dynamic Curriculum Orchestration for Domain-specific Large Language Model Fine-tuning","ko":"EDCO: 도메인 특화 대형 언어 모델 미세조정을 위한 동적 커리큘럼 오케스트레이션","authors":"Jing-Cheng Pang, Sun Liu, Chang Zhou, Xian Tang, Haichuan Ma, KUN JIANG, Jianlong Wang, Kai Zhang, Sijie Wu, Haoran Cai, Chenwei Wu, Xubin Li, Xin Chen","pos":"#214","link":"https://openreview.net/forum?id=NiHC3GlPQU","abs":"

Domain-specific large language models (LLMs), typically developed by fine-tuning a pre-trained general-purpose LLM on specialized datasets, represent a significant advancement in applied AI. A common strategy in LLM fine-tuning is curriculum learning, which pre-orders training samples based on metrics like difficulty to improve learning efficiency compared to a random sampling strategy. However, most existing methods for LLM fine-tuning rely on a static curriculum, designed prior to training, which lacks adaptability to the model's evolving needs during fine-tuning. To address this, we propose EDCO, a novel framework based on two key concepts: \\textit{inference entropy} and \\textit{dynamic curriculum orchestration}. Inspired by recent findings that maintaining high answer entropy benefits long-term reasoning gains, EDCO prioritizes samples with high inference entropy in a continuously adapted curriculum. EDCO integrates three core components: an efficient entropy estimator that uses prefix tokens to approximate full-sequence entropy, an entropy-based curriculum generator that selects data points with the highest inference entropy, and an LLM trainer that optimizes the model on the selected curriculum. Comprehensive experiments in communication, medicine and law domains, EDCO outperforms traditional curriculum strategies for fine-tuning Qwen3-4B and Llama3.2-3B models under supervised and reinforcement learning settings. Furthermore, the proposed efficient entropy estimation reduces computational time by 83.5\\% while maintaining high accuracy.

","absKo":"도메인 특화 large language model (LLM)은 일반적으로 사전 학습된 범용 LLM을 특화된 dataset으로 fine-tuning하여 개발되며, 응용 AI의 중요한 진전을 나타낸다. LLM fine-tuning에서 흔한 전략 중 하나는 curriculum learning으로, 학습 효율을 random sampling 전략보다 높이기 위해 difficulty 같은 metric을 기준으로 training sample을 사전에 순서를 매긴다. 그러나 기존의 대부분의 LLM fine-tuning 방법은 training 전에 설계된 정적 curriculum에 의존하며, fine-tuning 동안 변화하는 model의 요구에 적응하지 못한다. 이를 해결하기 위해 우리는 \\textit{inference entropy}와 \\textit{dynamic curriculum orchestration}이라는 두 핵심 개념에 기반한 새로운 framework인 EDCO를 제안한다. 최근의 발견들, 즉 높은 answer entropy를 유지하는 것이 장기적인 reasoning 향상에 유리하다는 점에서 영감을 받아, EDCO는 지속적으로 조정되는 curriculum에서 inference entropy가 높은 sample을 우선시한다. EDCO는 세 가지 핵심 구성 요소를 통합한다. prefix token을 사용해 full-sequence entropy를 근사하는 효율적인 entropy estimator, inference entropy가 가장 높은 data point를 선택하는 entropy-based curriculum generator, 그리고 선택된 curriculum에 대해 model을 최적화하는 LLM trainer이다. communication, medicine, law 분야에서 수행한 포괄적 실험에서 EDCO는 supervised learning과 reinforcement learning 설정 모두에서 Qwen3-4B 및 Llama3.2-3B model을 fine-tuning할 때 기존 curriculum 전략보다 우수한 성능을 보였다. 더 나아가, 제안한 효율적인 entropy estimation은 높은 정확도를 유지하면서 computation time을 83.5\\% 줄였다.

"},{"id":"65181","en":"HyPER: Bridging Exploration and Exploitation for Scalable LLM Reasoning with Hypothesis Path Expansion and Reduction","ko":"HyPER: 가설 경로 확장과 축소를 통한 확장 가능한 LLM 추론을 위한 탐험과 활용의 연결","authors":"Shengxuan Qiu, Haochen Huang, Shuzhang Zhong, Pengfei Zuo, Meng Li","pos":"#2303","link":"https://openreview.net/forum?id=G29kBVeIZt","abs":"

Scaling test-time compute with multi-path chain-of-thought can improve reasoning accuracy, but its gains hinge on an effective exploration–exploitation trade-off. Existing methods handle this trade-off in rigid ways: tree-structured search hard-codes exploration via brittle expansion rules that disrupt post-trained reasoning, while parallel reasoning over-explores redundant hypothesis paths and relies on a weak answer selection strategy. Driven by the insight that the optimal balance is phase-dependent and that correct vs. incorrect paths often diverge only at late stages, we reconceptualize test-time scaling as a dynamic expand–reduce control problem over a pool of hypothesis paths. We introduce HyPER, a training-free online control policy for MoE multi-path decoding that reallocates compute under a fixed budget using lightweight path statistics. HyPER features (i) an online controller that shifts from exploration to exploitation as the hypothesis pool evolves, (ii) an MoE-based token-level refinement primitive for efficient generation-time exploitation without full-path resampling, and (iii) a length- and confidence-aware aggregation rule to bridge the existence–selection gap for reliable answer-time exploitation. Extensive experimental results across four MoE models and diverse benchmarks demonstrate HyPER consistently achieves the accuracy–compute Pareto frontier, outperforming prior-art methods by 8-10% while reducing token consumption by 25-40%.

","absKo":"

multi-path chain-of-thought로 test-time compute를 확장하면 reasoning 정확도를 높일 수 있지만, 그 이득은 효과적인 exploration–exploitation trade-off에 달려 있다. 기존 방법들은 이 trade-off를 경직된 방식으로 다룬다. tree-structured search는 brittle한 expansion rule을 통해 exploration을 하드코딩하여 post-trained reasoning을 방해하고, parallel reasoning은 중복되는 hypothesis path를 지나치게 탐색하며 약한 answer selection strategy에 의존한다. 최적의 균형은 phase-dependent이며, 올바른 경로와 잘못된 경로는 종종 후반 단계에서야 갈라진다는 통찰에 기반해, 우리는 test-time scaling을 hypothesis path pool 위에서 동적으로 작동하는 expand–reduce control problem으로 재정의한다. 우리는 MoE multi-path decoding을 위한 training-free online control policyHyPER를 제안하며, 가벼운 path statistics를 사용해 고정된 budget 내에서 compute를 재할당한다. HyPER의 특징은 (i) hypothesis pool이 진화함에 따라 exploration에서 exploitation으로 전환하는 online controller, (ii) full-path resampling 없이 효율적인 generation-time exploitation을 가능하게 하는 MoE 기반 token-level refinement primitive, 그리고 (iii) 존재–선택 간 격차를 메워 신뢰할 수 있는 answer-time exploitation을 제공하는 length- 및 confidence-aware aggregation rule이다. 4개의 MoE model과 다양한 benchmark에서 수행한 광범위한 실험 결과, HyPER는 일관되게 accuracy–compute Pareto frontier를 달성했으며, 기존 방법보다 8-10% 높은 성능을 보이면서 token consumption은 25-40% 줄였다.

"},{"id":"61070","en":"How to Fine-Tune a Reasoning Model? A Teacher–Student Cooperation Framework to Synthesize Student-Consistent SFT Data","ko":"Reasoning Model을 어떻게 Fine-Tune할 것인가? Student-Consistent SFT 데이터 합성을 위한 Teacher-Student 협력 프레임워크","authors":"Zixian Huang, Kaichen Yang, Xu Huang, Feiyang Hao, Qiming Ge, Bowen Li, He Du, Kai Chen, Qipeng Guo","pos":"#2305","link":"https://openreview.net/forum?id=tylCinzjUF","abs":"

A widely adopted strategy for model enhancement is to use synthetic data generated by a stronger model for supervised fine-tuning (SFT). However, for emerging reasoning models like Qwen3-8B, this approach often fails to improve reasoning capabilities and can even lead to a substantial drop in performance. In this work, we identify substantial stylistic divergence between teacher generated data and the distribution of student as a major factor impacting SFT. To bridge this gap, we propose a Teacher–Student Cooperation Data Synthesis framework (TESSY), which interleaves teacher and student models to alternately generate style and non-style tokens. Consequently, TESSY produces synthetic sequences that inherit the advanced reasoning capabilities of the teacher while maintaining stylistic consistency with the distribution of the student. In experiments on code generation using GPT-OSS-120B as the teacher, fine-tuning Qwen3-8B on teacher-generated data leads to performance drops of 3.25% on LiveCodeBench-Pro and 10.02% on OJBench, whereas TESSY achieves improvements of 11.25% and 6.68%.

","absKo":"

모델 성능 향상을 위한 널리 채택된 전략은 더 강한 모델이 생성한 synthetic data를 사용해 supervised fine-tuning(SFT)을 수행하는 것이다. 그러나 Qwen3-8B와 같은 emerging reasoning model에서는 이 접근이 reasoning 능력을 개선하지 못하는 경우가 많고, 오히려 성능을 크게 저하시킬 수도 있다. 본 연구에서는 teacher-generated data와 student 분포 사이의 상당한 stylistic divergence가 SFT에 영향을 미치는 주요 요인임을 확인한다. 이 간극을 메우기 위해 우리는 Teacher–Student Cooperation Data Synthesis framework(TESSY)를 제안한다. TESSY는 teacher와 student model을 교차적으로 사용하여 style token과 non-style token을 번갈아 생성한다. 그 결과 TESSY는 teacher의 고급 reasoning capability를 계승하면서도 student 분포와의 stylistic consistency를 유지하는 synthetic sequence를 생성한다. GPT-OSS-120B를 teacher로 사용한 code generation 실험에서, teacher-generated data로 Qwen3-8B를 fine-tuning하면 LiveCodeBench-Pro에서 3.25\\%, OJBench에서 10.02\\%의 성능 하락이 발생하는 반면, TESSY는 각각 11.25\\%와 6.68\\%의 향상을 달성한다.

"},{"id":"62721","en":"HGMem: Hypergraph-based Working Memory to Improve Multi-step RAG for Long-Context Complex Relational Modeling","ko":"HGMem: 장문 맥락 복합 관계 추론을 위한 Multi-step RAG 개선 Hypergraph 기반 Working Memory","authors":"Chulun Zhou, Chunkang Zhang, Guoxin Yu, Fandong Meng, Jie Zhou, Wai Lam, Mo Yu","pos":"#2307","link":"https://openreview.net/forum?id=dvrq3GX7qL","abs":"

Multi-step retrieval-augmented generation (RAG) has become a widely adopted strategy for enhancing large language models (LLMs) on tasks that demand global comprehension and intensive reasoning. Although many RAG systems incorporate a working memory to consolidate information, existing designs primarily function as a passive storage for isolated facts. This static nature overlooks crucial high-order correlations among primitive facts, thereby limiting the capacity for multi-step reasoning and resulting in fragmented reasoning and weak global sense-making within extended contexts. We introduce HGMem, a hypergraph-based memory mechanism that extends the concept of memory beyond simple storage into a dynamic, expressive structure for complex reasoning and global understanding. In our approach, memory is represented as a hypergraph where hyperedges correspond to distinct memory units, enabling the progressive formation of higher-order interactions within memory. This mechanism connects facts and thoughts around the focal problem, evolving the memory into an integrated and situated knowledge structure that provides strong propositions for deeper reasoning. We evaluate HGMem on several challenging global sense-making benchmarks. Extensive experiments and in-depth analyses demonstrate that our method consistently improves multi-step RAG and substantially outperforms strong baseline systems across diverse datasets.

","absKo":"

Multi-step retrieval-augmented generation (RAG)는 global comprehension과 intensive reasoning을 요구하는 task에서 large language models (LLMs)를 향상시키기 위한 널리 채택된 전략이 되었다. 많은 RAG system이 정보를 통합하기 위해 working memory를 포함하지만, 기존 설계는 주로 분리된 사실들을 위한 수동적 저장소로 기능한다. 이러한 정적인 특성은 원시 사실들 사이의 중요한 고차 상관관계를 간과하게 하며, 그 결과 multi-step reasoning 능력이 제한되고 extended context 내에서 reasoning이 단절되며 global sense-making이 약해진다. 우리는 memory의 개념을 단순한 저장을 넘어 복잡한 reasoning과 global understanding을 위한 동적이고 표현력 있는 구조로 확장하는 hypergraph 기반 memory mechanism인 HGMem을 소개한다. 우리의 접근에서는 memory가 hypergraph로 표현되며, hyperedge가 서로 다른 memory unit에 대응하여 memory 내부에서 higher-order interaction이 점진적으로 형성되도록 한다. 이 mechanism은 focal problem 주변의 사실과 생각을 연결하여, memory를 더 깊은 reasoning을 위한 강한 proposition을 제공하는 통합적이고 situated knowledge structure로 진화시킨다. 우리는 여러 까다로운 global sense-making benchmark에서 HGMem을 평가한다. 광범위한 실험과 심층 분석은 우리의 방법이 multi-step RAG를 일관되게 개선하고 다양한 dataset 전반에서 강력한 baseline system들을 크게 능가함을 보여준다.

"},{"id":"61345","en":"From Seeing to Thinking: Decoupling Perception and Reasoning Improves Post-Training of Vision-Language Models","ko":"보기에서 생각하기로: 지각과 추론의 분리가 Vision-Language Model의 Post-Training을 개선한다","authors":"Juncheng Wu, Hardy Chen, Haoqin Tu, Xianfeng Tang, Freda Shi, Hui Liu, Hanqing Lu, Cihang Xie, Yuyin Zhou","pos":"#2310","link":"https://openreview.net/forum?id=r7uOjvZdzO","abs":"

Recent advances in vision-language models (VLMs) emphasize long chain-of-thought reasoning; yet, we find that their performance on visual tasks is primarily limited by a lack of visual perception as opposed to reasoning itself. In this work, we systematically study the interplay between perception and reasoning in VLM post-training by decomposing their capabilities into three separate training stages: visual perception, visual reasoning, and textual reasoning, incorporating specialized training data. We demonstrate that visual perception (a) requires targeted optimization with specialized data; (b) serves as a fundamental scaffold that should be solidified through staged training before refining visual reasoning; and (c) is more effectively learned via RL than caption-based SFT. Our experiments across multiple VLMs demonstrate that staged training consistently improves both visual perception and reasoning performance over merged training. Notably, models trained with our approach achieve 1.5\\% higher reasoning accuracy with 20.8\\% shorter reasoning traces, suggesting that superior perception reduces the need for excessive reasoning. Finally, our staged-training models achieve superior performance among open-weight VLMs, establishing advanced results on several visual math and perception (e.g., +5.2\\% on WeMath and +3.7\\% on RealWorldQA) tasks compared with the base counterpart.

","absKo":"

vision-language models (VLMs)의 최근 발전은 long chain-of-thought reasoning을 강조하지만, 우리는 시각 과제에서의 성능이 reasoning 자체보다는 주로 visual perception의 부족에 의해 제한됨을 발견한다. 이 연구에서 우리는 VLM post-training에서 perception과 reasoning의 상호작용을 체계적으로 연구하기 위해, 그들의 능력을 세 개의 분리된 training stage로 분해한다: visual perception, visual reasoning, textual reasoning이며, 여기에 특화된 training data를 통합한다. 우리는 visual perception이 (a) 특화된 data로 표적화된 최적화를 필요로 하고, (b) visual reasoning을 정교화하기 전에 staged training을 통해 먼저 견고하게 만들어야 하는 근본적인 scaffold 역할을 하며, (c) caption-based SFT보다 RL을 통해 더 효과적으로 학습된다는 점을 보인다. 여러 VLM에 대한 실험은 staged training이 merged training보다 visual perception과 reasoning 성능을 일관되게 향상시킴을 보여준다. 특히, 우리 방법으로 학습된 모델은 reasoning trace를 20.8% 더 짧게 사용하면서도 reasoning 정확도를 1.5% 높게 달성하여, 더 나은 perception이 과도한 reasoning의 필요성을 줄인다는 점을 시사한다. 마지막으로, staged-training 모델은 open-weight VLM 중에서도 우수한 성능을 달성하며, base counterpart와 비교해 여러 visual math 및 perception(예:, WeMath에서 +5.2%, RealWorldQA에서 +3.7%) task에서 진전된 결과를 확립한다.

"},{"id":"63260","en":"DocVAL: Validated Chain-of-Thought Distillation for Grounded Document VQA","ko":"DocVAL: 근거 기반 문서 VQA를 위한 검증된 Chain-of-Thought 증류","authors":"Pinaki Prasad Guha Neogi, Ahmad Mohammadshirazi, Ser-Nam Lim, Rajiv Ramnath","pos":"#2403","link":"https://openreview.net/forum?id=Ygg3KJPeT6","abs":"

Document visual question answering requires models not only to answer questions correctly, but also to precisely localize answers within complex document layouts. While large vision-language models (VLMs) achieve strong spatial grounding, their inference cost and latency limit real-world deployment; on the other hand, compact VLMs are efficient but suffer substantial localization degradation under standard fine-tuning or distillation. To address this gap, we propose DocVAL, a validated chain-of-thought (CoT) distillation framework that transfers explicit spatial reasoning from large teacher models to compact, deployable student VLMs. DocVAL combines (1) teacher-generated spatial CoT supervision, (2) a rule-based dual-mode validator that filters low-quality training signals and provides fine-grained, pixel-level corrective feedback, and (3) a validation-driven two-stage training procedure with iterative refinement. Text detection is used only as training-time scaffolding for supervision and validation, enabling the final student to operate as a pure VLM without OCR or detection at inference. Across multiple document understanding benchmarks, the proposed DocVAL yields consistent improvements of up to 6--7 ANLS points over comparable compact VLMs. We further introduce mean Average Precision (mAP) as a localization metric for document question answering and report strong spatial grounding performance under this new evaluation. We release 95K validator-verified CoT traces and show that high-quality, validated supervision is more effective than scaling unfiltered data, enabling efficient and trustworthy document grounding. Dataset and implementation: https://anonymous.4open.science/r/DocVAL-1C14

","absKo":"

Document visual question answering은 모델이 질문에 정확히 답하는 것뿐 아니라, 복잡한 document layout 안에서 답을 정확하게 localize할 것을 요구한다. large vision-language models (VLMs)는 강한 spatial grounding을 달성하지만, 추론 비용과 latency가 실제 배포를 제한한다. 반면 compact VLMs는 효율적이지만 standard fine-tuning이나 distillation 하에서는 localization 성능이 크게 저하된다. 이러한 격차를 해소하기 위해 우리는 \\textbf{DocVAL}을 제안한다. DocVAL은 큰 teacher model에서 compact하고 배포 가능한 student VLMs로 명시적 spatial reasoning을 전달하는 validated chain-of-thought (CoT) distillation framework이다. DocVAL은 \\textbf{(1)} teacher-generated spatial CoT supervision, \\textbf{(2)} 저품질 training signal을 필터링하고 세밀한 pixel-level corrective feedback을 제공하는 rule-based dual-mode validator, \\textbf{(3)} 반복적 refinement를 포함한 validation-driven two-stage training procedure를 결합한다. Text detection은 supervision과 validation을 위한 training-time scaffolding으로만 사용되며, 이를 통해 최종 student는 inference 시 OCR이나 detection 없이 pure VLM으로 동작할 수 있다. 여러 document understanding benchmark 전반에서 제안한 \\textbf{DocVAL}은 비교 가능한 compact VLMs 대비 최대 \\textbf{6--7 ANLS} 포인트의 일관된 개선을 보인다. 또한 document question answering을 위한 localization metric으로 mean Average Precision (mAP)을 새롭게 도입하고, 이 새로운 평가 아래에서 강한 spatial grounding 성능을 보고한다. 우리는 \\textbf{95K validator-verified CoT traces}를 공개하며, 고품질의 validated supervision이 unfiltered data를 단순히 확장하는 것보다 더 효과적임을 보임으로써, 효율적이고 신뢰할 수 있는 document grounding을 가능하게 한다. Dataset and implementation: https://anonymous.4open.science/r/DocVAL-1C14

"},{"id":"66314","en":"Exploring Relational Reasoning Capabilities in LLMs with REL","ko":"REL을 이용한 LLM의 관계적 추론 능력 탐구","authors":"Lukas Fesser, Yasha Ektefaie, Ada Fang, Sham Kakade, Marinka Zitnik","pos":"#2412","link":"https://openreview.net/forum?id=4oU82peL8f","abs":"

Relational reasoning is the ability to infer relations that jointly bind multiple entities, attributes, or variables. While this capability is essential for scientific reasoning, most existing evaluations of relational reasoning in large language models focus on structured inputs such as tables, graphs, or synthetic relational tasks, and do not isolate the sources of difficulty that arise from higher-arity relational binding. We study this problem through the lens of Relational Complexity (RC), defined as the minimum number of independent entities or operands that must be simultaneously bound to apply a relation. RC provides a principled way to vary reasoning difficulty independently of confounders such as input size, vocabulary, and representational choices. Building on RC, we introduce REL, a generative benchmark framework spanning algebra, chemistry, and biology that varies RC within each domain. Evaluating frontier LLMs, we observe a consistent and monotonic degradation in performance as RC increases, even when the total number of entities is held fixed. This failure mode persists under increased test-time compute and with in-context learning, suggesting a limitation tied to the arity of the required relational binding rather than insufficient inference steps or exposure to examples. Our results identify a well-defined regime of higher-arity reasoning in which current models struggle and motivate revisiting reasoning benchmarks through the lens of relational complexity.

","absKo":"

Relational reasoning은 여러 entity, attribute 또는 variable을 함께 결합하는 관계를 추론하는 능력입니다. 이 능력은 scientific reasoning에 필수적이지만, large language model에서 relational reasoning을 평가하는 대부분의 기존 연구는 table, graph, 또는 synthetic relational task와 같은 structured input에 초점을 맞추며, 더 높은 arity의 relational binding에서 발생하는 난이도의 원인을 분리하지는 않습니다. 우리는 이 문제를 Relational Complexity (RC)의 관점에서 연구합니다. RC는 어떤 relation을 적용하기 위해 동시에 결합되어야 하는 독립적인 entity 또는 operand의 최소 개수로 정의됩니다. RC는 input size, vocabulary, representation choice와 같은 혼란 변수와 독립적으로 reasoning 난이도를 원리 있게 조절할 수 있는 방법을 제공합니다. RC를 바탕으로, 우리는 각 domain 내에서 RC를 변화시키는 algebra, chemistry, biology를 아우르는 생성형 benchmark framework인 REL을 소개합니다. frontier LLM을 평가한 결과, 총 entity 수를 고정하더라도 RC가 증가할수록 성능이 일관되고 단조롭게 저하되는 것을 관찰했습니다. 이러한 실패 양상은 test-time compute를 늘리거나 in-context learning을 사용해도 지속되며, 이는 추론 단계 수나 예시 노출의 부족보다는 필요한 relational binding의 arity와 관련된 한계를 시사합니다. 우리의 결과는 현재 모델이 어려움을 겪는 higher-arity reasoning의 명확한 영역을 밝혀내며, relational complexity의 관점에서 reasoning benchmark를 다시 살펴볼 필요가 있음을 보여줍니다.

"},{"id":"64843","en":"Didactic to Constructive: Turning Expert Solutions into Learnable Reasoning","ko":"교훈적에서 구성적으로: 전문가 솔루션을 학습 가능한 추론으로 전환","authors":"Ethan Mendes, Jungsoo Park, Alan Ritter","pos":"#2502","link":"https://openreview.net/forum?id=JG6f02X29a","abs":"Improving the reasoning capabilities of large language models (LLMs) typically relies either on the model's ability to sample a correct solution to be reinforced or the existence of a stronger model able to solve the problem. However, many difficult problems remain intractable for even current frontier models, preventing the extraction of valid training signals. A promising alternative is to leverage high-quality expert human solutions, yet naive imitation of this data fails because it is fundamentally out-of-distribution: expert solutions are typically didactic, containing implicit reasoning gaps intended for human readers rather than computational models. Furthermore, high-quality expert solutions are expensive, necessitating generalizable sample-efficient training methods. We propose Distribution Aligned Imitation Learning (DAIL), a two-step method that bridges the distributional gap by first transforming expert solutions into detailed, in-distribution reasoning traces and then applying a contrastive objective to focus learning on expert insights and methodologies. We find that DAIL can leverage fewer than 1000 high-quality expert solutions to achieve 10–25\\% pass@$k$ gains on Qwen2.5-Instruct and Qwen3 models, improve reasoning efficiency by $2\\times$ to $4\\times$, and enable out-of-domain generalization.","absKo":"대규모 언어 모델(LLM)의 reasoning 능력을 향상시키는 방식은 보통 두 가지에 의존한다. 하나는 강화할 수 있는 올바른 solution을 샘플링할 수 있는 모델의 능력이고, 다른 하나는 문제를 풀 수 있는 더 강한 모델의 존재이다. 그러나 많은 어려운 문제는 현재의 frontier model조차 해결하기 어려워 유효한 training signal을 추출할 수 없다. 유망한 대안은 고품질 expert human solution을 활용하는 것이지만, 이 데이터를 naive하게 imitation하면 본질적으로 out-of-distribution이기 때문에 실패한다. expert solution은 대개 인간 독자를 위한 didactic한 형태로 작성되어 있어, 계산 모델에 필요한 암묵적 reasoning gap을 포함하기 때문이다. 게다가 고품질 expert solution은 비용이 많이 들어, 일반화 가능하고 sample-efficient한 training method가 필요하다. 우리는 Distribution Aligned Imitation Learning(DAIL)을 제안한다. 이는 먼저 expert solution을 상세하고 in-distribution인 reasoning trace로 변환한 뒤, contrastive objective를 적용해 학습이 expert의 통찰과 방법론에 집중하도록 하는 2단계 방법이다. 우리는 DAIL이 1000개 미만의 고품질 expert solution만으로도 Qwen2.5-Instruct와 Qwen3 모델에서 10–25\\% pass@$k$ 향상을 달성하고, reasoning 효율을 $2\\times$에서 $4\\times$까지 개선하며, out-of-domain generalization을 가능하게 함을 확인했다."},{"id":"63300","en":"Curriculum-Guided Layer Scaling for Language Model Pretraining","ko":"언어 모델 사전학습을 위한 커리큘럼 유도 레이어 스케일링","authors":"Karanpartap Singh, Neil Band, Ehsan Adeli","pos":"#2506","link":"https://openreview.net/forum?id=YIOVZ7Xmoo","abs":"

As the cost of pretraining large language models grows, there is continued interest in strategies to improve learning efficiency during this core training stage. Motivated by cognitive development, where humans gradually build knowledge as their brains mature, we propose Curriculum-Guided Layer Scaling (CGLS), a framework for compute-efficient pretraining that synchronizes increasing data difficulty with model growth through progressive layer stacking (i.e. gradually adding layers during training). At the 100M parameter scale, using a curriculum transitioning from synthetic short stories to general web data, CGLS outperforms baseline methods on the question-answering benchmarks PIQA and ARC. Pretraining at the 1.2B scale, we stratify the DataComp-LM corpus with a DistilBERT-based classifier and progress from general text to highly technical or specialized content. Our results show that progressively increasing model depth alongside sample difficulty leads to better generalization and zero-shot performance on various downstream benchmarks. Altogether, our findings demonstrate that CGLS unlocks the potential of progressive stacking, offering a simple yet effective strategy for improving generalization on knowledge-intensive and reasoning tasks.

","absKo":"

대규모 language model의 pretraining 비용이 증가함에 따라, 이 핵심 training stage에서 learning efficiency를 개선하는 전략에 대한 관심이 계속되고 있다. 인간이 뇌가 성숙함에 따라 점진적으로 지식을 쌓는 cognitive development에서 영감을 받아, 우리는 progressive layer stacking, 즉 training 중에 layer를 점진적으로 추가하면서 증가하는 data difficulty와 model growth를 동기화하는 compute-efficient pretraining 프레임워크인 Curriculum-Guided Layer Scaling (CGLS)을 제안한다. 100M parameter 규모에서 synthetic short story에서 general web data로 전환하는 curriculum을 사용하면, CGLS는 question-answering benchmark인 PIQA와 ARC에서 baseline method를 능가한다. 1.2B 규모의 pretraining에서는 DataComp-LM corpus를 DistilBERT 기반 classifier로 stratify하고, 일반 텍스트에서 매우 기술적이거나 전문화된 내용으로 점진적으로 이동한다. 우리의 결과는 sample difficulty와 함께 model depth를 점진적으로 증가시키는 것이 다양한 downstream benchmark에서 더 나은 generalization과 zero-shot 성능으로 이어짐을 보여준다. 종합하면, 우리의 결과는 CGLS가 progressive stacking의 잠재력을 열어 주며, knowledge-intensive 및 reasoning task에서 generalization을 향상시키는 단순하면서도 효과적인 전략을 제공함을 보여준다.

"},{"id":"61485","en":"ConPress: Learning Efficient Reasoning from Multi-Question Contextual Pressure","ko":"ConPress: 다중 질문 맥락 압력에서 효율적 추론 학습","authors":"Jie Deng, Shining Liang, Jun Li, Hongzhi Li, Yutao Xie","pos":"#2512","link":"https://openreview.net/forum?id=piXLPucLRA","abs":"

Large reasoning models (LRMs) typically solve reasoning-intensive tasks by generating long chain-of-thought (CoT) traces, leading to substantial inference overhead. We identify a reproducible inference-time phenomenon, termed \\textbf{\\emph{Self-Compression}}: when multiple independent and answerable questions are presented within a single prompt, the model spontaneously produces shorter reasoning traces for each question. This phenomenon arises from \\emph{multi-question contextual pressure} during generation and consistently manifests across models and benchmarks. Building on this observation, we propose ConPress (Learning from Contextual Pressure, a lightweight self-supervised fine-tuning approach. ConPress constructs multi-question prompts to induce self-compression, samples the resulting model outputs, and parses and filters per-question traces to obtain concise yet correct reasoning trajectories. These trajectories are directly used for supervised fine-tuning, internalizing compressed reasoning behavior in single-question settings without external teachers, manual pruning, or reinforcement learning. With only 8k fine-tuning examples, ConPress reduces reasoning token usage by 59\\% on MATH500 and 33\\% on AIME25, while maintaining competitive accuracy.

","absKo":"

Large reasoning models (LRMs)는 일반적으로 긴 chain-of-thought (CoT) trace를 생성하여 reasoning-intensive task를 해결하므로 inference overhead가 상당하다. 우리는 \\textbf{\\emph{Self-Compression}}이라 불리는 재현 가능한 inference-time 현상을 확인했다. 즉, 여러 개의 독립적이고 답이 가능한 질문을 하나의 prompt에 함께 제시하면, 모델이 각 질문에 대해 더 짧은 reasoning trace를 자발적으로 생성한다. 이 현상은 생성 과정에서의 \\emph{multi-question contextual pressure}에서 비롯되며, 모델과 benchmark 전반에 걸쳐 일관되게 나타난다. 이 관찰을 바탕으로, 우리는 Learning from Contextual Pressure를 뜻하는 경량 self-supervised fine-tuning 접근법인 ConPress를 제안한다. ConPress는 self-compression을 유도하기 위해 multi-question prompt를 구성하고, 그에 따른 model output을 샘플링한 뒤, question별 trace를 파싱하고 필터링하여 간결하면서도 정답인 reasoning trajectory를 얻는다. 이 trajectory는 그대로 supervised fine-tuning에 사용되어, external teacher나 수동 pruning, reinforcement learning 없이 단일 질문 설정에서도 압축된 reasoning behavior를 내재화한다. 단 8k개의 fine-tuning example만으로도 ConPress는 MATH500에서 reasoning token 사용량을 59\\%, AIME25에서 33\\% 줄이면서 경쟁력 있는 accuracy를 유지한다.

"},{"id":"66033","en":"Beyond Scalars: Evaluating and Understanding LLM Reasoning via Geometric Progress and Stability","ko":"스칼라를 넘어서: 기하학적 진행과 안정성을 통한 LLM 추론 평가 및 이해","authors":"Xinyan Jiang, Ninghao Liu, Di Wang, Lijie Hu","pos":"#3201","link":"https://openreview.net/forum?id=7i44Lx4KEY","abs":"

Evaluating LLM reliability via scalar probabilities often fails to capture the structural dynamics of reasoning. We introduce TRACED, a framework that assesses reasoning quality through theoretically grounded geometric kinematics. By decomposing reasoning traces into Progress (displacement) and Stability (curvature), we reveal a distinct topological divergence: correct reasoning manifests as high-progress, stable trajectories, whereas hallucinations are characterized by low-progress, unstable patterns (stalled displacement with high curvature fluctuations). Leveraging these signatures, our probabilistic framework achieves competitive performance and superior robustness across diverse benchmarks. Crucially, TRACED bridges geometry and cognition by mapping high curvature to \"Hesitation Loops'' and displacement to ''Certainty Accumulation'', offering a physical lens to decode the internal dynamics of machine thought.

","absKo":"scalar probability를 통해 LLM reliability를 평가하는 방식은 reasoning의 구조적 dynamics를 포착하지 못하는 경우가 많다. 우리는 이론적으로 정당화된 geometric kinematics를 통해 reasoning quality를 평가하는 framework인 TRACED를 소개한다. reasoning trace를 Progress (displacement)와 Stability (curvature)로 분해함으로써, 우리는 뚜렷한 topological divergence를 드러낸다. 정답 reasoning은 high-progress, stable trajectory로 나타나는 반면, hallucination은 low-progress, unstable pattern, 즉 high curvature fluctuation을 동반한 stalled displacement로 특징지어진다. 이러한 signature를 활용하면 우리의 probabilistic framework는 다양한 benchmark 전반에서 경쟁력 있는 성능과 우수한 robustness를 달성한다. 결정적으로, TRACED는 high curvature를 \"Hesitation Loops''에, displacement를 ''Certainty Accumulation''에 매핑함으로써 geometry와 cognition을 연결하고, machine thought의 내부 dynamics를 해독하는 물리적 관점을 제공한다.

"},{"id":"64198","en":"Short Chains, Deep Thoughts: Balancing Reasoning Efficiency and Intra-Segment Capability via Split-Merge Optimization","ko":"짧은 사슬, 깊은 사고: 분할-병합 최적화를 통한 추론 효율성과 세그먼트 내 능력의 균형","authors":"Runquan Gui, Jie Wang, Zhihai Wang, Chi Ma, Jianye Hao, Feng Wu","pos":"#3705","link":"https://openreview.net/forum?id=Q3Dukl2zWg","abs":"

While Large Reasoning Models (LRMs) have demonstrated impressive capabilities in solving complex tasks through the generation of long reasoning chains, this reliance on verbose generation results in significant latency and computational overhead. To address these challenges, we propose \\textbf{CoSMo} (\\textbf{Co}nsistency-Guided \\textbf{S}plit-\\textbf{M}erge \\textbf{O}ptimization), a framework designed to eliminate structural redundancy rather than indiscriminately restricting token volume. Specifically, CoSMo utilizes a split-merge algorithm that dynamically refines reasoning chains by merging redundant segments and splitting logical gaps to ensure coherence. We then employ structure-aligned reinforcement learning with a novel segment-level budget to supervise the model in maintaining efficient reasoning structures throughout training. Extensive experiments across multiple benchmarks and backbones demonstrate that CoSMo achieves superior performance, improving accuracy by \\textbf{3.3} points while reducing segment usage by \\textbf{28.7\\%} on average compared to reasoning efficiency baselines.

","absKo":"

Large Reasoning Model(LRM)은 긴 reasoning chain을 생성함으로써 복잡한 task를 해결하는 인상적인 능력을 보여주었지만, 이러한 verbose generation에 대한 의존은 상당한 latency와 computational overhead를 초래한다. 이러한 문제를 해결하기 위해, 우리는 토큰 양을 무차별적으로 제한하는 대신 구조적 중복을 제거하도록 설계된 framework인 \\textbf{CoSMo}(\\textbf{Co}nsistency-Guided \\textbf{S}plit-\\textbf{M}erge \\textbf{O}ptimization)를 제안한다. 구체적으로, CoSMo는 분할-병합(split-merge) 알고리즘을 사용해 redundant segment를 병합하고 논리적 공백을 분할함으로써 reasoning chain을 동적으로 정제하고, 일관성을 보장한다. 이후 우리는 새로운 segment-level budget을 갖는 structure-aligned reinforcement learning을 사용해, 학습 전반에 걸쳐 모델이 효율적인 reasoning structure를 유지하도록 감독한다. 여러 benchmark와 backbone에 걸친 광범위한 실험은 CoSMo가 reasoning efficiency baseline과 비교해 평균적으로 segment 사용량을 \\textbf{28.7\\%} 줄이면서 정확도를 \\textbf{3.3}포인트 향상시켜 우수한 성능을 달성함을 보여준다.

"},{"id":"65098","en":"Teaching Models to Teach Themselves: Reasoning at the Edge of Learnability","ko":"스스로 가르치도록 모델을 가르치기: 학습 가능성의 경계에서의 추론","authors":"Shobhita Sundaram, John Quan, Ariel Kwiatkowski, Kartik Ahuja, Yann Ollivier, Julia Kempe","pos":"#3713","link":"https://openreview.net/forum?id=GnqHK8Ww98","abs":"

RL methods for finetuning large reasoning models stall on datasets with low initial success rates, and thus little training signal. We investigate a fundamental question: Can a pretrained LLM leverage latent knowledge to generate an automated curriculum for problems it cannot solve? We explore this with SOAR: A self-improvement framework designed to surface these pedagogical signals through meta-RL. A teacher model proposes synthetic problems for a student model, and is rewarded with its improvement on a subset of hard problems, thus grounding the curriculum in real student progress rather than proxy rewards. Our study on the hardest subsets of math benchmarks (0/128 success) reveal three core findings. First, it is possible to realize bi-level meta-RL that unlocks learning under sparse, binary rewards by sharpening a latent capacity of pretrained models to generate useful problems. Second, grounded rewards outperform intrinsic rewards used in prior LLM self-play, reliably avoiding the typical instability and diversity collapse modes. Third, the structure and well-posedness of questions are more critical for learning progress than solution correctness. Our results suggest that the ability to generate useful stepping stones does not require the preexisting ability to solve the hard problems, paving a principled path to escape reasoning plateaus without additional curated data.

","absKo":"

큰 reasoning model을 finetuning하기 위한 RL 방법은 초기 성공률이 낮아 학습 신호가 거의 없는 dataset에서 정체되기 쉽다. 우리는 근본적인 질문을 던진다. pretrained LLM이 해결할 수 없는 문제에 대해 latent knowledge를 활용하여 자동화된 curriculum을 생성할 수 있는가? 우리는 메타 RL을 통해 이러한 pedagogical signal을 드러내도록 설계된 self-improvement framework인 SOAR로 이를 탐구한다. teacher model이 student model을 위한 synthetic problem을 제안하고, hard problem의 일부에서 student가 개선되면 보상을 받는다. 이로써 curriculum은 proxy reward가 아니라 실제 student progress에 기반하게 된다. 가장 어려운 math benchmark subset(0/128 success)에 대한 우리의 연구는 세 가지 핵심 결과를 보여 준다. 첫째, pretrained model의 latent capacity를 유용한 문제 생성으로 날카롭게 함으로써 sparse한 binary reward 아래에서도 학습을 가능하게 하는 bi-level meta-RL을 실현할 수 있다. 둘째, grounded reward는 prior LLM self-play에서 사용된 intrinsic reward보다 우수하며, 일반적인 instability와 diversity collapse 모드를 안정적으로 회피한다. 셋째, 질문의 구조와 well-posedness가 solution correctness보다 학습 진행에 더 중요하다. 우리의 결과는 유용한 stepping stone을 생성하는 능력이 hard problem 자체를 풀 수 있는 기존 능력을 필요로 하지 않음을 시사하며, 추가로 정제된 데이터 없이 reasoning plateau를 탈출할 수 있는 원리적인 경로를 제시한다.

"},{"id":"61860","en":"Deep networks learn to parse uniform-depth context-free languages from local statistics","ko":"Deep Network는 국소 통계로부터 균일 깊이 Context-Free Language 파싱을 학습한다","authors":"Jack T. Parley, Francesco Cagnetta, Matthieu Wyart","pos":"#4607","link":"https://openreview.net/forum?id=mJgkPAFdiK","abs":"

Understanding how the structure of language can be learned from sentences alone is a central question in both cognitive science and machine learning. Studies of the internal representations of Large Language Models (LLMs) support their ability to parse text when predicting the next word, while representing semantic notions independently of surface form. Yet, which data statistics make these feats possible, and how much data is required, remain largely unknown. Probabilistic context-free grammars (PCFGs) provide a tractable testbed for studying these questions. However, prior work has focused either on the post-hoc characterization of the parsing-like algorithms used by trained networks; or on the learnability of PCFGs with fixed syntax, where parsing is unnecessary. Here, we (i) introduce a tunable class of PCFGs in which both the degree of ambiguity and the correlation structure across scales can be controlled; (ii) provide a learning mechanism---an inference algorithm inspired by the structure of deep convolutional networks---that links learnability and sample complexity to specific language statistics; and (iii) validate our predictions empirically across deep convolutional and transformer-based architectures. Overall, we propose a unifying framework where correlations at different scales lift local ambiguities, enabling the emergence of hierarchical representations of the data.

","absKo":"

문장만으로부터 language의 structure가 어떻게 학습될 수 있는지 이해하는 것은 cognitive science와 machine learning 모두에서 핵심적인 질문이다. Large Language Models (LLMs)의 내부 representation에 대한 연구는 다음 단어를 예측할 때 이들이 텍스트를 parsing할 수 있는 능력과, surface form과 독립적으로 semantic notion을 표현하는 능력을 뒷받침한다. 그러나 이러한 성능을 가능하게 하는 데이터 통계가 무엇인지, 그리고 얼마나 많은 데이터가 필요한지는 여전히 대부분 알려져 있지 않다. Probabilistic context-free grammars (PCFGs)는 이러한 질문을 연구하기 위한 tractable한 testbed를 제공한다. 하지만 기존 연구는 학습된 network가 사용하는 parsing-like algorithm을 사후적으로 특징짓는 데 초점을 맞추었거나, parsing이 필요 없는 fixed syntax의 PCFG 학습가능성에만 집중해 왔다. 여기서 우리는 (i) ambiguity의 정도와 scale 전반의 correlation structure를 모두 제어할 수 있는 tunable PCFG class를 도입하고; (ii) deep convolutional network의 structure에서 영감을 받은 inference algorithm인 learning mechanism을 제시하여, learnability와 sample complexity를 특정 language statistic과 연결하며; (iii) deep convolutional 및 transformer 기반 architecture 전반에서 우리의 예측을 실증적으로 검증한다. 전체적으로, 우리는 서로 다른 scale의 correlation이 local ambiguity를 해소하여 데이터의 hierarchical representation이 출현하도록 하는 통합 프레임워크를 제안한다.

"},{"id":"64884","en":"Causal Matrix Completion under Multiple Treatments via Mixed Synthetic Nearest Neighbors","ko":"Mixed Synthetic Nearest Neighbor를 통한 다중 처치 하의 Causal Matrix Completion","authors":"Minrui Luo, Zhiheng Zhang","pos":"#3710","link":"https://openreview.net/forum?id=Ir6N7U5Kea","abs":"

Synthetic Nearest Neighbors (SNN) provides a principled solution to causal matrix completion under missing-not-at-random (MNAR) by exploiting local low-rank structure through fully observed anchor submatrices. However, its effectiveness critically relies on sufficient data availability within each treatment level, a condition that often fails in settings with multiple or complex treatments. In this work, we propose Mixed Synthetic Nearest Neighbors (MSNN), a new entry-wise causal identification estimator that integrates information across treatment levels. We show that MSNN retains the finite-sample error bounds and asymptotic normality guarantees of SNN, while enlarging the effective sample size available for estimation. Empirical results on synthetic and real-world datasets illustrate the efficacy of the proposed approach, especially under data-scarce treatment levels.

","absKo":"

Synthetic Nearest Neighbors(SNN)는 fully observed anchor submatrix를 통해 local low-rank structure를 활용함으로써 missing-not-at-random(MNAR) 상황에서 causal matrix completion에 대한 원리적인 해법을 제공한다. 그러나 그 효과성은 각 treatment level 내에서 충분한 데이터 가용성에 결정적으로 의존하며, 이는 여러 개 또는 복잡한 treatment가 있는 설정에서는 자주 성립하지 않는다. 본 연구에서는 treatment level 전반의 정보를 통합하는 새로운 entry-wise causal identification estimator인 Mixed Synthetic Nearest Neighbors(MSNN)를 제안한다. 우리는 MSNN이 SNN의 finite-sample error bound와 asymptotic normality 보장을 유지하면서도, 추정을 위해 사용 가능한 effective sample size를 확장함을 보인다. synthetic 및 실제 데이터셋에 대한 실험 결과는 제안한 접근법의 효용을 보여주며, 특히 데이터가 희소한 treatment level에서 두드러진다.

"},{"id":"68775","en":"Multi-Accurate CATE is Robust to Unknown Covariate Shifts","ko":"Multi-Accurate CATE는 알 수 없는 Covariate Shift에 강건하다","authors":"Christoph Kern, Michael Kim, Angela Zhou","pos":"#3905","abs":"

Estimating heterogeneous treatment effects is important to tailor treatments to those individuals who would most likely benefit. However, conditional average treatment effect predictors may often be trained on one population but possibly deployed on different, possibly unknown populations. We use methodology for learning multi-accurate predictors to post-process CATE T-learners (differenced regressions) to become robust to unknown covariate shifts at the time of deployment. The method works in general for pseudo-outcome regression, such as the DR-learner. We show how this approach can combine (large) confounded observational and (smaller) randomized datasets by learning a confounded predictor from the observational dataset, and auditing for multi-accuracy on the randomized controlled trial. We show improvements in bias and mean squared error in simulations with increasingly larger covariate shift, and on a semi-synthetic case study of a parallel large observational study and smaller randomized controlled experiment. Overall, we establish a connection between methods developed for multi-distribution learning and achieve appealing desiderata (e.g. external validity) in causal inference and machine learning.

","absKo":"

이질적인 treatment effect를 추정하는 것은 가장 큰 이익을 얻을 가능성이 높은 개인에게 치료를 맞춤화하는 데 중요하다. 그러나 conditional average treatment effect predictor는 종종 한 집단에서 학습되지만, 다른 집단, 심지어는 알려지지 않은 집단에 배포될 수 있다. 우리는 multi-accurate predictor를 학습하는 방법론을 사용하여 CATE T-learner(차분 회귀)를 후처리함으로써 배포 시점의 알 수 없는 covariate shift에 대해 강건하게 만든다. 이 방법은 DR-learner와 같은 pseudo-outcome regression 전반에 일반적으로 적용된다. 우리는 이 접근법이 관측 데이터셋으로부터 confounded predictor를 학습하고, randomized controlled trial에서 multi-accuracy를 감사함으로써 (큰) confounded observational dataset과 (더 작은) randomized dataset을 결합할 수 있음을 보인다. 우리는 점점 더 커지는 covariate shift에 대한 시뮬레이션과, 대규모 관측 연구 및 더 작은 randomized controlled experiment가 병행된 semi-synthetic case study에서 bias와 mean squared error의 개선을 보인다. 전체적으로 우리는 multi-distribution learning을 위해 개발된 방법과의 연결을 확립하고, causal inference와 machine learning에서 외적 타당성(external validity)과 같은 매력적인 desiderata를 달성한다.

"},{"id":"61982","en":"Use What You Know: Causal Foundation Models with Partial Graphs","ko":"알고 있는 것을 활용하라: partial graph를 이용한 causal foundation models","authors":"Arik Reuter, Anish Dhir, Cristiana Diaconu, Jake Robertson, Ole Ossen, Frank Hutter, Adrian Weller, Mark van der Wilk, Bernhard Schölkopf","pos":"#4109","link":"https://openreview.net/forum?id=l1PVCATRCx","abs":"

Estimating causal quantities traditionally relies on bespoke estimators tailored to specific assumptions. Recently proposed Causal Foundation Models (CFMs) promise a more unified approach by amortising causal discovery and inference in a single step. However, in their current state, they do not allow for the incorporation of any domain knowledge, which can lead to suboptimal predictions. We bridge this gap by introducing methods to condition CFMs on causal information, such as the causal graph or more readily available ancestral information. When access to complete causal graph information is too strict a requirement, our approach also effectively leverages partial causal information. We systematically evaluate conditioning strategies and find that injecting learnable biases into the attention mechanism is the most effective method to utilise full and partial causal information. Our experiments show that this conditioning allows a general-purpose CFM to match the performance of specialised models trained on specific causal structures. Overall, our approach addresses a central hurdle on the path towards all-in-one causal foundation models: the capability to answer causal queries in a data-driven manner while effectively leveraging any amount of domain expertise.

","absKo":"

causal quantity를 추정하는 전통적인 방법은 특정 가정에 맞춘 bespoke estimator에 의존한다. 최근 제안된 Causal Foundation Model (CFM)은 causal discovery와 inference를 한 번에 amortize하는 보다 통합된 접근을 약속한다. 그러나 현재 상태의 CFM은 어떠한 domain knowledge도 통합할 수 없어, 이는 최적 이하의 예측으로 이어질 수 있다. 우리는 causal graph나 더 쉽게 확보 가능한 ancestral information과 같은 causal information으로 CFM을 조건화하는 방법을 도입함으로써 이 간극을 메운다. 완전한 causal graph 정보에 접근하는 것이 너무 엄격한 요구일 때, 우리의 접근법은 부분적인 causal information도 효과적으로 활용한다. 우리는 conditioning strategy를 체계적으로 평가하여, attention mechanism에 learnable bias를 주입하는 것이 full 및 partial causal information을 활용하는 가장 효과적인 방법임을 발견한다. 우리의 실험은 이러한 conditioning이 general-purpose CFM이 특정 causal structure에서 학습된 specialized model의 성능과 맞먹도록 해준다는 것을 보여준다. 전반적으로, 우리의 접근법은 all-in-one causal foundation model로 가는 길에서 핵심적인 장애물을 해결한다. 즉, 데이터 기반 방식으로 causal query에 답하면서도 도메인 전문 지식을 가능한 한 많이 효과적으로 활용할 수 있는 능력이다.

"},{"id":"63252","en":"Unveiling Prior-data Fitted Networks on Causal Effect Estimation: Pre-training or Finetuning?","ko":"인과 효과 추정에서 Prior-data Fitted Network 규명: 사전학습인가 미세조정인가?","authors":"Haotian Wang, Xinpeng Lv, Hao Zou, Yanghao Xiao, Shanzhi Gu, Yang Shi, Yunxin Mao, Yuanxing Zhang, Mingyang Geng, Shaowu Yang, Haoxuan Li, Wenjing Yang, Peng Cui, Zhouchen Lin","pos":"#4110","link":"https://openreview.net/forum?id=YjPDsCH5zg","abs":"

Amortized causal inference via Prior-data Fitted Networks (PFNs) has emerged as a promising paradigm, enabling zero-shot estimation of causal effects without the need for dataset-specific model tuning. However, the principled effectiveness of unified pre-training across general interventional regimes remains an underexplored question. In this paper, we investigate interventions on subsets of variables within Structural Causal Models (SCMs) and identify a fundamental theoretical limitation of current pre-training approaches. Theoretically, we prove that a single observational SCM induces an exponentially large space of interventional distributions, resulting in a phenomenon we term prior uncoverage. Consequently, this uncoverage yields a mismatch between the learned meta-prior and the true grounding prior, leading to unavoidable posterior inconsistency and estimation bias. To address this, we posit that fine-tuning is a fundamental necessity and propose a target-specific strategy named Point-Wise Interventional Fine-tuning (PWF), enabling the local generalization property. We further scale this approach via Meta-Sampling Fine-tuning (MSF) from a budgeted active learning perspective, thereby achieving uniform generalization on any interventional distribution.

","absKo":"

Prior-data Fitted Networks(PFNs)를 통한 amortized causal inference는 dataset별 model tuning 없이 zero-shot으로 causal effect를 추정할 수 있게 해주는 유망한 패러다임으로 떠올랐다. 그러나 일반적인 intervention regime 전반에 걸친 통합 pre-training의 원리적 효과는 아직 충분히 탐구되지 않았다. 본 논문에서는 Structural Causal Models(SCMs)에서 변수 부분집합에 대한 intervention을 살펴보고, 현재의 pre-training 접근법이 지닌 근본적인 이론적 한계를 규명한다. 이론적으로 우리는 하나의 observational SCM이 exponentially large한 interventional distribution 공간을 유도함을 증명하며, 이를 prior uncoverage라는 현상으로 정의한다. 그 결과 이러한 uncoverage는 학습된 meta-prior와 실제 grounding prior 간의 불일치를 초래하고, 이는 피할 수 없는 posterior inconsistency와 estimation bias로 이어진다. 이를 해결하기 위해 우리는 fine-tuning이 근본적으로 필요하다는 입장을 취하고, local generalization property를 가능하게 하는 target-specific 전략인 Point-Wise Interventional Fine-tuning(PWF)을 제안한다. 또한 budgeted active learning 관점에서 Meta-Sampling Fine-tuning(MSF)으로 이 접근을 확장하여, 어떤 interventional distribution에서도 uniform generalization을 달성한다.

"},{"id":"66812","en":"Uncovering Bias Mechanisms in Observational Studies","ko":"관찰 연구에서 편향 메커니즘 밝히기","authors":"Ilker Demirel, Zeshan Hussain, Piersilvio De Bartolomeis, David Sontag","pos":"#4111","link":"https://openreview.net/forum?id=072o1IbwJi","abs":"

Observational studies are a key resource for causal inference but are often affected by systematic biases. Prior work has focused mainly on detecting these biases, via sensitivity analyses and comparisons with randomized controlled trials, or mitigating them through debiasing techniques. However, there remains a lack of methodology for uncovering the underlying mechanisms driving these biases, e.g., whether due to hidden confounding or selection of participants. In this work, we show that the relationship between bias magnitude and the predictive performance of nuisance function estimators (in the observational study) can help distinguish among common sources of bias. We validate our methodology through extensive synthetic experiments and a real-world case study, demonstrating its effectiveness in revealing the mechanisms behind observed biases. Our framework offers a new lens for understanding and characterizing bias in observational studies, with practical implications for improving causal inference.

","absKo":"

Observational study는 causal inference의 핵심 자원이지만, 체계적 bias의 영향을 받는 경우가 많다. 기존 연구는 주로 sensitivity analysis와 randomized controlled trial과의 비교를 통해 이러한 bias를 탐지하거나, debiasing technique을 통해 완화하는 데 초점을 맞춰 왔다. 그러나 hidden confounding이나 participant selection 때문인지 등, 이러한 bias를 유발하는 근본 메커니즘을 밝혀내는 방법론은 여전히 부족하다. 본 연구에서는 observational study에서 nuisance function estimator의 predictive performance와 bias magnitude 사이의 관계가 일반적인 bias source를 구분하는 데 도움이 될 수 있음을 보인다. 우리는 광범위한 synthetic experiment와 실제 사례 연구를 통해 이 방법론을 검증했으며, 관측된 bias의 배후 메커니즘을 밝히는 데 효과적임을 입증했다. 우리의 framework는 observational study에서 bias를 이해하고 특성화하는 새로운 관점을 제공하며, causal inference 개선에 실질적인 함의를 가진다.

"},{"id":"63541","en":"Treatment Responder Classification with Abstention","ko":"거부를 포함한 치료 반응자 분류","authors":"Haoxiang Wang, Haoxuan Li, Ziyan Wang, Zhiheng Zhang, Aoqi Zuo, Erdun Gao, Kun Zhang, Mingming Gong","pos":"#4112","link":"https://openreview.net/forum?id=WFdQSjmchK","abs":"

Treatment responder classification seeks to learn a rule to classify individuals who will benefit from the treatment. This paper studies a new scenario in treatment responder classification when abstention is allowed, i.e., practitioners can opt out of making uncertain classification on some individuals for further investigation. By revealing the implicit relation between causal misclassification risk with abstention and Conditional Value at Risk (CVaR), we develop a doubly robust method named TRECA to learn the classification rule under loose convergence conditions on nuisance parameters, and further extend it to deal with possible violation on key assumptions such as monotonicity and unconfoundedness. Rigorous theories and extensive experiments on two real-world datasets demonstrate the theoretical and experimental guarantee on our methods in learning treatment responders classification rules with low regret at the cost of limited abstention.

","absKo":"

Treatment responder classification은 치료로부터 이익을 얻을 개인을 분류하는 규칙을 학습하는 것을 목표로 한다. 본 논문은 abstention이 허용되는 treatment responder classification의 새로운 상황을 연구한다. 즉, 실무자는 추가 조사를 위해 일부 개인에 대해 불확실한 분류를 하지 않기로 선택할 수 있다. 우리는 abstention이 있는 causal misclassification risk와 Conditional Value at Risk (CVaR) 사이의 암묵적 관계를 밝힘으로써, nuisance parameter에 대한 느슨한 convergence condition 하에서 classification rule을 학습하는 doubly robust method인 TRECA를 개발하고, 이를 monotonicity와 unconfoundedness 같은 핵심 가정의 잠재적 위반을 다루도록 확장한다. 엄밀한 이론과 두 개의 real-world dataset에 대한 광범위한 실험은 제한된 abstention의 대가로 낮은 regret로 treatment responder classification rule을 학습하는 우리 방법의 이론적·실험적 보장을 입증한다.

"},{"id":"61852","en":"Modeling Covariate Transition for Efficient Estimation of Longitudinal Treatment Effects in Randomized Experiments","ko":"무작위 실험에서 종단적 처치 효과의 효율적 추정을 위한 Covariate Transition 모델링","authors":"Naoki Chihara, Tatsushi Oka, Yasuko Matsubara, Yasushi Sakurai, Shota Yasui","pos":"#4200","link":"https://openreview.net/forum?id=mLZlTzepbr","abs":"

We present a regression-adjustment framework designed to estimate longitudinal treatment effects in randomized experiments under static regimes. Although regression-adjustment methods are useful for variance reduction in randomized experiments through the use of pre-treatment covariates, they usually focus only on average effects, which cannot capture valuable knowledge about when effects appear and how long they continue. To address this limitation, we need to consider intermediate outcomes and evolving post-treatment covariates over time, and we represent these transitions using transition kernels. Furthermore, we establish the asymptotic normality and the semiparametric efficiency bound for our estimator, enabling more powerful statistical inference. Simulation studies and empirical analysis using A/B test data from a streaming platform in Japan demonstrate the practical advantages of our method.

","absKo":"

우리는 static regime 하의 randomized experiment에서 longitudinal treatment effect를 추정하기 위한 regression-adjustment framework를 제시한다. regression-adjustment 방법은 pre-treatment covariate를 사용해 randomized experiment의 variance를 줄이는 데 유용하지만, 보통 average effect에만 초점을 맞추며, 이는 효과가 언제 나타나고 얼마나 오래 지속되는지에 대한 중요한 정보를 포착하지 못한다. 이 한계를 해결하기 위해서는 시간에 따라 변화하는 intermediate outcome과 post-treatment covariate를 고려해야 하며, 우리는 이러한 전이를 transition kernel을 사용해 표현한다. 더 나아가, 우리는 우리의 estimator에 대한 asymptotic normality와 semiparametric efficiency bound를 확립하여 더 강력한 statistical inference를 가능하게 한다. simulation study와 일본의 streaming platform A/B test data를 사용한 empirical analysis는 본 방법의 실용적 이점을 보여준다.

"},{"id":"63521","en":"Relational Structural Causal Models","ko":"관계적 구조적 인과 모델","authors":"Adiba Ejaz, Elias Bareinboim","pos":"#4207","link":"https://openreview.net/forum?id=WTBaZHtIra","abs":"

An artificial intelligence must have a model of its environment that is causal, supporting reasoning about interventions and counterfactuals, and also combinatorial, supporting generalization to unseen combinations of objects. In this work, we formally study when and how such a model can be learned. We develop relational structural causal models, extending structural causal models (Pearl 2009) to settings where objects and their relations vary. First, we show how answers to not only causal but also observational queries about unseen combinations of objects can not be identified without further assumptions. To enable such identification---including in the presence of unobserved confounding---we define relational causal graphs and derive symbolic identification criteria. Finally, we propose relational neural causal models, a provably correct approach that outperforms non-relational baselines on simulated traffic scenes with varying cars, signals, and pedestrians.

","absKo":"

인공 지능은 개입과 반사실에 대한 추론을 지원하는 인과적 환경 모델을 가져야 하며, 동시에 보지 못한 객체 조합에 대한 일반화를 지원하는 조합적 모델도 가져야 한다. 이 연구에서는 그러한 모델을 언제, 그리고 어떻게 학습할 수 있는지를 형식적으로 다룬다. 우리는 구조적 인과 모델(Structural Causal Models, Pearl 2009)을 객체와 그 관계가 변하는 설정으로 확장한 관계적 구조적 인과 모델을 개발한다. 먼저, 추가 가정 없이는 보지 못한 객체 조합에 대한 인과 질의뿐 아니라 관측 질의에 대한 답도 식별할 수 없음을 보인다. 관측되지 않은 교란이 존재하는 경우까지 포함하여 이러한 식별을 가능하게 하기 위해, 우리는 관계적 인과 그래프를 정의하고 기호적 식별 기준을 도출한다. 마지막으로, 시뮬레이션된 교통 장면에서 서로 다른 차량, 신호, 보행자를 대상으로 비관계적 베이스라인보다 더 나은 성능을 보이는, 이론적으로 정당화된 접근법인 관계적 신경 인과 모델을 제안한다.

"},{"id":"66032","en":"Robust Causal Discovery in Real-World Time Series with Power-Laws","ko":"멱법칙을 갖는 실세계 시계열에서의 강건한 인과 발견","authors":"Matteo Tusoni, Giuseppe Masi, Andrea Coletta, Aldo Glielmo, Viviana Arrigoni, Novella Bartolini","pos":"#4208","link":"https://openreview.net/forum?id=7i8d203tky","abs":"

Exploring causal relationships in stochastic time series is a challenging yet crucial task with a vast range of applications, including finance, economics, neuroscience, and climate science. Many algorithms for Causal Discovery (CD) have been proposed; however, they often exhibit a high sensitivity to noise, resulting in spurious causal inferences on real data. In this paper, we observe that the frequency spectra of many real-world time series follow a power-law distribution, notably due to an inherent self-organizing behavior. Leveraging this insight, we build a robust CD method based on the extraction of power‑law spectral features that amplify genuine causal signals. Our method consistently outperforms state-of-the-art alternatives on both synthetic benchmarks and real-world datasets with known causal structures, demonstrating its robustness and practical relevance.

","absKo":"확률적 time series에서 causal relationship을 탐색하는 것은 finance, economics, neuroscience, climate science를 포함한 매우 넓은 응용 범위를 지닌 어렵지만 핵심적인 과제이다. Causal Discovery (CD)를 위한 많은 algorithm이 제안되었지만, 이들은 종종 noise에 매우 민감하여 실제 데이터에서 spurious causal inference를 낳는다. 본 논문에서는 많은 실제 time series의 frequency spectrum이 power-law distribution을 따른다는 점을 관찰하며, 이는 특히 내재된 self-organizing behavior 때문임을 보인다. 이러한 통찰을 활용해, 우리는 진정한 causal signal을 증폭시키는 power-law spectral feature 추출에 기반한 robust CD method를 구축한다. 우리의 method는 synthetic benchmark와 알려진 causal structure를 가진 real-world dataset 모두에서 state-of-the-art 대안들을 일관되게 능가하며, 그 robust성과 실용적 타당성을 입증한다.

"},{"id":"62870","en":"Spatial Deconfounder: Interference-Aware Deconfounding for Spatial Causal Inference","ko":"Spatial Deconfounder: 공간 인과 추론을 위한 간섭 인지 deconfounding","authors":"Ayush Khot, Miruna Oprescu, Maresa Schröder, Ai Kagawa, Xihaier Luo","pos":"#4212","link":"https://openreview.net/forum?id=cOkJvV16Fd","abs":"

Causal inference in spatial domains faces two intertwined challenges: (1) unmeasured spatial factors, such as weather, air pollution, or mobility, that confound treatment and outcome, and (2) interference from nearby treatments that violate standard no-interference assumptions. While existing methods typically address one by assuming away the other, we show they are deeply connected: interference reveals structure in the latent confounder. Leveraging this insight, we propose the Spatial Deconfounder, a two-stage method that reconstructs a substitute confounder from local treatment vectors using a conditional variational autoencoder (C-VAE) with a spatial prior, then estimates causal effects via a flexible outcome model. We show that this approach enables nonparametric identification of both direct and spillover effects under weak assumptions—without requiring multiple treatment types or a known model of the latent field. Empirically, we extend SpaCE, a benchmark suite for spatial confounding, to include treatment interference, and show that the Spatial Deconfounder consistently improves effect estimation across real-world datasets in environmental health and social science. By turning interference into a multi-cause signal, our framework bridges spatial and deconfounding literatures to advance robust causal inference in structured spatial data.

","absKo":"

공간 도메인에서의 causal inference는 두 가지 맞물린 문제에 직면한다: (1) treatment와 outcome을 혼란시키는 날씨, 대기 오염, 이동성 같은 관측되지 않은 spatial factor, 그리고 (2) 표준의 no-interference 가정을 위반하는 주변 treatment로부터의 interference이다. 기존 방법들은 대개 하나를 다루는 대신 다른 하나를 가정으로 없애 버리지만, 우리는 이 둘이 깊게 연결되어 있음을 보인다. 즉, interference는 latent confounder의 구조를 드러낸다. 이 통찰을 활용해 우리는 Spatial Deconfounder를 제안한다. 이는 spatial prior를 갖는 conditional variational autoencoder(C-VAE)를 사용해 local treatment vector로부터 substitute confounder를 복원한 뒤, 유연한 outcome model로 causal effect를 추정하는 2단계 방법이다. 우리는 이 접근이 여러 treatment type이나 latent field의 알려진 model 없이도, 약한 가정하에서 direct effect와 spillover effect를 모두 nonparametric하게 식별할 수 있음을 보인다. 실험적으로는 spatial confounding을 위한 benchmark suite인 SpaCE를 treatment interference를 포함하도록 확장하고, Spatial Deconfounder가 환경 보건과 사회과학의 실제 데이터셋 전반에서 effect estimation을 일관되게 개선함을 보인다. interference를 multi-cause signal로 바꾸는 이 framework는 공간 데이터에서의 robust causal inference를 발전시키기 위해 spatial literatures와 deconfounding literatures를 연결한다.

"},{"id":"62411","en":"Many Experiments, Few Repetitions, Unpaired Data, and Sparse Effects: Is Causal Inference Possible?","ko":"많은 실험, 적은 반복, 비대응 데이터, 희소 효과: 인과 추론이 가능한가?","authors":"Felix Schur, Niklas Pfister, Peng Ding, Sach Mukherjee, Jonas Peters","pos":"#4301","link":"https://openreview.net/forum?id=gqa99Ev4C4","abs":"In many applications, practical constraints prevent measuring covariates and outcomes on the same units, resulting in unpaired data. We study the problem of estimating causal effects under hidden confounding in the following unpaired data setting: we observe some covariates $X$ and an outcome $Y$ under different experimental conditions (environments) but do not observe them jointly -- we either observe $X$ or $Y$. Under appropriate regularity conditions, the problem can be cast as an instrumental variable (IV) regression with the environment acting as a (possibly high-dimensional) instrument. When there are many environments but only a few observations per environment, standard two-sample IV estimators fail to be consistent. We propose a GMM-type estimator based on cross-fold sample splitting of the instrument–covariate sample that also applies in standard IV settings. We prove that it is consistent as the number of environments grows but the sample size per environment remains constant. We further extend the method to sparse causal effects via $\\ell_1$-regularized estimation and post-selection refitting.","absKo":"많은 응용에서 실제 제약 때문에 공변량과 결과를 같은 단위에서 측정할 수 없어, 짝지어지지 않은(unpaired) 데이터가 발생한다. 우리는 숨은 교란(hidden confounding)이 있는 다음의 unpaired data 설정에서 인과 효과를 추정하는 문제를 연구한다. 즉, 서로 다른 실험 조건(environment) 아래에서 일부 공변량 $X$와 결과 $Y$는 관측되지만, 둘을 함께는 관측하지 못하며, $X$ 또는 $Y$ 중 하나만 관측한다. 적절한 정규성 조건 하에서 이 문제는 환경이 (잠재적으로 고차원인) instrument로 작용하는 instrumental variable (IV) regression으로 정식화할 수 있다. 환경 수는 많지만 각 환경당 관측 수가 적을 때, 표준 two-sample IV 추정기는 일관적이지 않다. 우리는 instrument–covariate sample에 대한 cross-fold sample splitting을 기반으로 한 GMM-type 추정기를 제안하며, 이는 표준 IV setting에도 적용 가능하다. 우리는 환경 수는 증가하지만 환경당 샘플 크기는 일정하게 유지될 때 이 추정기가 일관적임을 증명한다. 또한 $\\ell_1$-regularized estimation과 post-selection refitting을 통해 sparse causal effect로 방법을 확장한다."},{"id":"65599","en":"Interventional Processes For Causal Uncertainty Quantification","ko":"인과적 불확실성 정량화를 위한 개입 프로세스","authors":"Hugh Dance, Peter Orbanz, Arthur Gretton","pos":"#4302","link":"https://openreview.net/forum?id=BzG0xtGjjr","abs":"

Reliable uncertainty quantification for causal effects is crucial in various applications, but remains difficult in nonparametric models, particularly for continuous treatments. We introduce IMPspec, a Gaussian process (GP) framework for modeling uncertainty over interventional causal functions under continuous treatments, which can be represented using reproducing Kernel Hilbert Spaces (RKHSs). By using principled function class expansions and a spectral representation of RKHS features, IMPspec yields tractable training and inference, a spectral algorithm to calibrate posterior credible intervals, and avoids the underfitting and variance collapse pathologies of earlier GP-on-RKHS methods. Across synthetic benchmarks and an application in healthcare, IMPspec delivers state-of-the-art performance in causal uncertainty quantification and downstream causal Bayesian optimization tasks.

","absKo":"

Causal effect에 대한 신뢰할 수 있는 uncertainty quantification은 다양한 응용에서 매우 중요하지만, 비모수적 모델, 특히 continuous treatment에서는 여전히 어렵다. 우리는 continuous treatment 하의 intervention causal function에 대한 uncertainty를 모델링하기 위한 Gaussian process(GP) framework인 IMPspec을 제안하며, 이는 reproducing Kernel Hilbert Spaces(RKHSs)로 표현될 수 있다. 원리적인 function class expansion과 RKHS feature의 spectral representation을 사용함으로써, IMPspec은 tractable한 training과 inference, posterior credible interval을 보정하는 spectral algorithm, 그리고 기존의 GP-on-RKHS 방법에서 나타나던 underfitting과 variance collapse 병리를 회피한다. 합성 benchmark와 healthcare 응용 전반에서, IMPspec은 causal uncertainty quantification 및 downstream causal Bayesian optimization task에서 state-of-the-art 성능을 제공한다.

"},{"id":"62998","en":"Evaluating bivariate causal statements based on mutual compatibility","ko":"상호 호환성 기반 bivariate 인과 문장 평가","authors":"Erik Jahn, Dominik Janzing","pos":"#4306","link":"https://openreview.net/forum?id=b3EvCd8sYE","abs":"For many real-world systems, causal ground truth is difficult to obtain, making claims about causal effects hard to assess. We develop methods for evaluating collections of $\\binom{n}{2}$ bivariate causal statements over a set of $n$ variables. In the setting of acyclic linear statements, any such collection can be extended to a unique multivariate causal model, but we argue that this induced model is implausible if it imposes substantial additional confounding to explain observed correlations. We introduce a compatibility score that quantifies this notion of plausibility, notably without relying on the faithfulness assumption. Additionally, we define an incompatibility score for purely graphical bivariate causal statements, based on global consistency constraints that are derived from acyclicity and faithfulness assumptions. We give theoretical and empirical evidence that both scores can successfully distinguish correct from incorrect causal statements in generic settings. Moreover, we demonstrate the practical applicability of our methods by analyzing causal claims made by large language models. Our work aims to provide a foundation for assessing the reliability of causal information derived from human experts or artificial intelligence in settings where alternative forms of validation are unavailable.","absKo":"많은 실제 시스템에서 causal ground truth를 얻기 어렵기 때문에, causal effect에 대한 주장을 평가하기가 쉽지 않다. 우리는 n개의 변수 집합에 대해 \\binom{n}{2}개의 bivariate causal statement 모음을 평가하는 방법을 개발한다. acyclic linear statement 설정에서 이러한 모음은 어떤 경우에도 유일한 multivariate causal model로 확장될 수 있지만, 우리는 관측된 correlation을 설명하기 위해 상당한 추가 confounding을 부과한다면 그 유도된 model은 그럴듯하지 않다고 주장한다. 우리는 faithfulness assumption에 의존하지 않으면서 이러한 그럴듯함을 정량화하는 compatibility score를 도입한다. 또한 acyclicity와 faithfulness assumption에서 도출되는 global consistency constraint를 기반으로, 순수한 graphical bivariate causal statement에 대한 incompatibility score를 정의한다. 우리는 이 두 score가 일반적인 설정에서 올바른 causal statement와 잘못된 causal statement를 성공적으로 구분할 수 있다는 이론적 및 실증적 근거를 제시한다. 더 나아가 large language model이 제시한 causal claim을 분석함으로써 우리의 방법이 실제로 적용 가능함을 보인다. 우리의 연구는 다른 형태의 검증이 불가능한 상황에서 human expert나 artificial intelligence로부터 도출된 causal information의 신뢰성을 평가하기 위한 기초를 제공하는 것을 목표로 한다."},{"id":"61894","en":"Distributionally Robust Causal Abstractions","ko":"분포 강건 Causal 추상화","authors":"Yorgos Felekis, Theodoros Damoulas, Paris Giampouras","pos":"#4308","link":"https://openreview.net/forum?id=lu2d1gQhHS","abs":"

Causal Abstraction (CA) theory provides a principled framework for relating causal models that describe the same system at different levels of granularity while ensuring interventional consistency between them. Recent methods for learning CAs, however, assume fixed and well-specified exogenous distributions, leaving them vulnerable to environmental shifts and model misspecification. In this work, we address these limitations by introducing the first class of distributionally robust CAs and their associated learning algorithms. The latter cast robust causal abstraction learning as a constrained min-max optimization problem with Wasserstein ambiguity sets. We provide theoretical guarantees for both empirical and Gaussian environments, enabling principled selection of ambiguity-set radii and establish quantitative guarantees on worst-case abstraction error. Furthermore, we present empirical evidence across different problems and CA learning methods, demonstrating our framework’s robustness not only to environmental shifts but also to structural and intervention mapping misspecification.

","absKo":"

Causal Abstraction (CA) 이론은 서로 다른 granularity 수준에서 같은 시스템을 설명하는 causal model을 상호 연관시키면서, 그들 사이의 intervention consistency를 보장하는 원칙적인 프레임워크를 제공한다. 그러나 최근 CA 학습 방법은 고정되고 잘 명세된 exogenous distribution을 가정하므로, 환경 변화와 model misspecification에 취약하다. 이 연구에서 우리는 최초의 distributionally robust CA와 그에 상응하는 학습 알고리즘을 도입하여 이러한 한계를 해결한다. 후자는 robust causal abstraction learning을 Wasserstein ambiguity set을 갖는 constrained min-max optimization 문제로 정식화한다. 우리는 empirical 및 Gaussian environment 모두에 대해 이론적 보장을 제공하여 ambiguity-set radius를 원칙적으로 선택할 수 있게 하고, worst-case abstraction error에 대한 정량적 보장을 확립한다. 더 나아가, 다양한 문제와 CA 학습 방법에 걸친 실증적 증거를 제시하여, 우리의 프레임워크가 환경 변화뿐 아니라 structural 및 intervention mapping misspecification에 대해서도 견고함을 보임을 보여준다.

"},{"id":"63808","en":"DISCO: Mitigating Bias in Deep Learning with Conditional Distance Correlation","ko":"조건부 distance correlation을 이용한 딥러닝 편향 완화 DISCO","authors":"Emre Kavak, Tom Nuno Wolf, Christian Wachinger","pos":"#4309","link":"https://openreview.net/forum?id=TYjTWxDOOC","abs":"Dataset bias often leads deep learning models to exploit spurious correlations instead of task-relevant signals. We introduce the Standard Anti-Causal Model (SAM), a unifying causal framework that characterizes bias mechanisms and yields a conditional independence criterion for causal stability. Building on this theory, we propose DISCO$_m$ and sDISCO, efficient and scalable estimators of conditional distance correlation that enable independence regularization in black-box models. Across six diverse datasets, our methods consistently outperform or are competitive in existing bias mitigation approaches, while requiring fewer hyperparameters and scaling seamlessly to multi-bias scenarios. This work bridges causal theory and practical deep learning, providing both a principled foundation and effective tools for robust prediction.","absKo":"Dataset bias는 deep learning model이 task와 관련된 신호 대신 spurious correlation을 이용하도록 유도하는 경우가 많다. 우리는 bias mechanism을 특징짓고 causal stability를 위한 conditional independence criterion을 도출하는 통합 causal framework인 Standard Anti-Causal Model(SAM)을 제안한다. 이 이론에 기반하여, black-box model에서 independence regularization을 가능하게 하는 conditional distance correlation의 효율적이고 확장 가능한 추정량인 DISCO$_m$과 sDISCO를 제안한다. 서로 다른 여섯 개의 데이터셋 전반에서, 우리의 방법은 기존 bias mitigation 접근법보다 일관되게 우수하거나 최소한 경쟁력 있는 성능을 보이며, 더 적은 hyperparameter를 요구하고 multi-bias scenario로도 매끄럽게 확장된다. 본 연구는 causal theory와 실용적 deep learning을 연결하며, 원리적인 기반과 강건한 예측을 위한 효과적인 도구를 함께 제공한다."},{"id":"64614","en":"Conditional Distributional Treatment Effects: Doubly Robust Estimation and Testing","ko":"조건부 분포적 처치 효과: Doubly Robust 추정 및 검정","authors":"Saksham Jain, Alex Luedtke","pos":"#4310","link":"https://openreview.net/forum?id=LfUXGnwVzw","abs":"

Beyond conditional average treatment effects, treatments may impact the entire outcome distribution in covariate-dependent ways, for example, by altering the variance or tail risks for specific subpopulations. We propose a novel estimand to capture such conditional distributional treatment effects, and develop a doubly robust estimator that is minimax optimal in the local asymptotic sense. Using this, we develop a test for the global homogeneity of conditional potential outcome distributions that accommodates discrepancies beyond the maximum mean discrepancy (MMD), has provably valid type 1 error, and is consistent against fixed alternatives---the first test, to our knowledge, with such guarantees in this setting. Furthermore, we derive exact closed-form expressions for two natural discrepancies (including the MMD), and provide a computationally efficient, permutation-free algorithm for our test.

","absKo":"

conditional average treatment effect를 넘어서, treatment는 covariate에 따라 outcome distribution 전체에 영향을 미칠 수 있으며, 예를 들어 특정 subpopulation에서 분산이나 tail risk를 변화시킬 수 있다. 우리는 이러한 conditional distributional treatment effect를 포착하기 위한 새로운 estimand를 제안하고, local asymptotic sense에서 minimax optimal인 doubly robust estimator를 개발한다. 이를 바탕으로, 우리는 최대 평균 불일치(maximum mean discrepancy, MMD)를 넘어서는 discrepancy를 수용하고, type 1 error가 이론적으로 타당함이 보장되며, 고정된 alternative에 대해 consistent한 conditional potential outcome distribution의 global homogeneity를 검정하는 test를 개발한다. 이는 우리가 아는 한, 이러한 보장을 갖는 이 설정의 첫 번째 test이다. 또한, 우리는 두 가지 자연스러운 discrepancy(MMD를 포함)에 대한 정확한 closed-form expression을 유도하고, 우리의 test를 위한 계산적으로 효율적인 permutation-free algorithm을 제공한다.

"},{"id":"66736","en":"CausalProfiler: Generating Synthetic Benchmarks for Rigorous and Transparent Evaluation of Causal Machine Learning","ko":"CausalProfiler: Causal Machine Learning의 엄밀하고 투명한 평가를 위한 합성 벤치마크 생성","authors":"Panayiotis Panayiotou, Audrey Poinsot, Alessandro Leite, Nicolas CHESNEAU, Marc Schoenauer, Özgür Şimşek","pos":"#4311","link":"https://openreview.net/forum?id=0wCl7EifsY","abs":"

Causal machine learning (Causal ML) aims to answer \"what if\" questions using machine learning algorithms, making it a promising tool for high-stakes decision-making. Yet, empirical evaluation practices in Causal ML remain limited. Existing benchmarks often rely on a handful of hand-crafted or semi-synthetic datasets, leading to brittle, non-generalizable conclusions. To bridge this gap, we introduce CausalProfiler, a synthetic benchmark generator for Causal ML methods. Based on a set of explicit design choices about the class of causal models, queries, and data considered, the CausalProfiler randomly samples causal models, data, queries, and ground truths constituting the synthetic causal benchmarks. In this way, Causal ML methods can be rigorously and transparently evaluated under a variety of conditions. This work offers the first random generator of synthetic causal benchmarks with coverage guarantees and transparent assumptions operating on the three levels of causal reasoning: observation, intervention, and counterfactual. We demonstrate its utility by evaluating several state-of-the-art methods under diverse conditions and assumptions, both in and out of the identification regime, illustrating the types of analyses and insights the CausalProfiler enables.

","absKo":"

Causal machine learning (Causal ML)은 머신러닝 알고리즘을 사용해 \"what if\" 질문에 답하는 것을 목표로 하며, 고위험 의사결정에 유망한 도구다. 그러나 Causal ML에서의 실증적 평가 관행은 여전히 제한적이다. 기존 benchmark는 대개 손으로 설계한 또는 반합성(semi-synthetic) 데이터셋 몇 개에 의존하므로, 결론이 취약하고 일반화 가능성이 낮다. 이러한 간극을 메우기 위해, 우리는 Causal ML 방법을 위한 합성 benchmark 생성기인 CausalProfiler를 제안한다. 인과 모델, query, data의 종류에 대한 명시적 설계 선택을 바탕으로, CausalProfiler는 합성 causal benchmark를 구성하는 causal model, data, query, ground truth를 무작위로 샘플링한다. 이를 통해 Causal ML 방법을 다양한 조건에서 엄밀하고 투명하게 평가할 수 있다. 본 연구는 observation, intervention, counterfactual이라는 세 가지 수준의 causal reasoning에서 작동하면서 coverage guarantee와 투명한 가정을 갖춘, 합성 causal benchmark의 첫 무작위 생성기를 제시한다. 또한 identification regime 안팎의 다양한 조건과 가정에서 여러 최신 방법을 평가함으로써, CausalProfiler가 가능하게 하는 분석과 통찰의 유형을 보여준다.

"},{"id":"66592","en":"Causal discovery for time series with endogenous context variables","ko":"내생적 맥락 변수를 갖는 시계열의 인과 발견","authors":"Oana-Iuliana Popescu, Wiebke Günther, Martin Rabel, Jakob Runge","pos":"#4312","link":"https://openreview.net/forum?id=2Nw6JjbwfG","abs":"

Many real-world systems exhibit both context- and time-dependent causal dynamics, where the dynamical system state also influences its context. For instance, soil moisture is driven by precipitation, yet also provides the context for heat-flux realization. We capture such dynamics in Structural Causal Models (SCMs) by introducing endogenous and time-dependent discrete context variables, also allowing for possibly lagged dependencies with the system variables. While context variables are discrete, they may also be proxies of continuous variables. The enabling assumptions for causal discovery of our model are either persistence of the context or sparsity of the context–system dependencies. We design two new PCMCI-based algorithms for causal discovery with endogenous context variables for time series and prove their soundness. A systematic evaluation on synthetic benchmarks and an application to a real-world land-atmosphere feedback problem demonstrate their effectiveness and applicability.

","absKo":"많은 실제 시스템은 context-와 time-dependent causal dynamics를 동시에 보이며, 이때 dynamical system의 상태가 그 context에도 영향을 미친다. 예를 들어 토양 수분은 강수에 의해 구동되지만, 동시에 heat-flux realization의 context를 제공한다. 우리는 endogenous이고 time-dependent인 discrete context variable을 도입하여 이러한 동역학을 Structural Causal Model(SCM)로 포착하며, 시스템 변수와의 지연된 의존성도 허용한다. context variable은 discrete이지만 continuous variable의 proxy일 수도 있다. 우리의 모델에서 causal discovery를 가능하게 하는 가정은 context의 persistence 또는 context-system 의존성의 sparsity 중 하나이다. 우리는 time series를 위한 endogenous context variable을 갖는 causal discovery에 대해 PCMCI 기반의 두 가지 새로운 알고리즘을 설계하고, 그 soundness를 증명한다. synthetic benchmark에 대한 체계적 평가와 실제 land-atmosphere feedback 문제에의 적용은 이 방법들의 효과성과 적용 가능성을 보여준다.

"},{"id":"63618","en":"Anti-causal domain generalization: Leveraging unlabeled data","ko":"반인과적 도메인 일반화: 라벨 없는 데이터 활용","authors":"Sorawit Saengkyongam, Juan L. Gamella, Andrew Miller, Jonas Peters, Nicolai Meinshausen, Christina Heinze-Deml","pos":"#4314","link":"https://openreview.net/forum?id=VOrzQ8GBDh","abs":"

The problem of domain generalization concerns learning predictive models that are robust to distribution shifts when deployed in new, previously unseen environments. Existing methods typically require labeled data from multiple training environments, limiting their applicability when labeled data are scarce. In this work, we study domain generalization in an anti-causal setting, where the outcome causes the observed covariates. Under this structure, environment perturbations that affect the covariates do not propagate to the outcome, which motivates regularizing the model's sensitivity to these perturbations. Crucially, estimating these perturbation directions does not require labels, enabling us to leverage unlabeled data from multiple environments. We propose two methods that penalize the model's sensitivity to variations in the mean and covariance of the covariates across environments, respectively, and prove that these methods have worst-case optimality guarantees under certain classes of environments. Finally, we demonstrate the empirical performance of our approach on a controlled physical system and a physiological signal dataset.

","absKo":"

domain generalization 문제는 새로운, 이전에 보지 못한 환경에 배포될 때 distribution shift에 robust한 predictive model을 학습하는 데 관한 것이다. 기존 방법은 보통 여러 training environment에서 라벨이 있는 데이터를 필요로 하므로, 라벨 데이터가 부족할 때 적용 가능성이 제한된다. 본 연구에서는 outcome이 observed covariate의 원인이 되는 anti-causal setting에서 domain generalization을 연구한다. 이러한 구조에서는 covariate에 영향을 주는 environment perturbation이 outcome으로 전파되지 않으므로, 이러한 perturbation에 대한 모델의 sensitivity를 regularize하는 것이 동기가 된다. 중요한 점은, 이러한 perturbation direction을 추정하는 데 라벨이 필요하지 않으므로, 여러 environment의 unlabeled data를 활용할 수 있다는 것이다. 우리는 각각 covariate의 mean과 covariance 변동에 대한 모델의 sensitivity를 penalize하는 두 가지 방법을 제안하고, 특정 class의 environment 하에서 이 방법들이 worst-case optimality guarantee를 갖는다는 것을 증명한다. 마지막으로, 제안 방법의 실증적 성능을 통제된 물리 시스템과 생리 신호 데이터셋에서 보여준다.

"},{"id":"60473","en":"Addressing Instrument-Outcome Confounding in Mendelian Randomization through Representation Learning","ko":"Representation Learning을 통한 Mendelian Randomization의 Instrument-Outcome Confounding 해결","authors":"Shimeng Huang, Matthew Robinson, Francesco Locatello","pos":"#4415","link":"https://openreview.net/forum?id=zxJXgfCm63","abs":"

Mendelian Randomization (MR) is a prominent observational epidemiological research method, designed to address unobserved confounding when estimating causal effects. It is closely related to instrumental variable (IV) methods, where genetic variants serve as instruments to infer causal relationships from observational data. However, the core assumptions required for valid IV analysis---particularly the independence between instruments and unobserved confounders---are untestable and often violated in practice. In MR, such violations commonly arise when genetic variants are correlated with environmental factors (e.g., population stratification and assortive mating), leading to confounding between instruments and outcomes. At the same time, MR studies increasingly include data collected across multiple environments or populations, providing an opportunity to address these violations. Leveraging this setting, we propose a representation learning framework that exploits multi-environment data to recover latent exogenous components of genetic instruments suitable for causal inference. We provide theoretical insights into when and how the learned components can act as valid instruments, and we demonstrate the effectiveness of our approach through simulations and semi-synthetic experiments using genetic data from the All of Us Biobank.

","absKo":"

Mendelian Randomization (MR)은 인과 효과를 추정할 때 관측되지 않은 confounding을 다루기 위해 고안된 대표적인 observational epidemiological research method이다. 이는 instrumental variable (IV) methods와 밀접하게 관련되며, 여기서 genetic variant는 observational data로부터 causal relationship을 추론하기 위한 instrument로 사용된다. 그러나 유효한 IV 분석에 필요한 핵심 가정, 특히 instrument와 관측되지 않은 confounder 사이의 독립성은 검증할 수 없고 실제로 자주 위배된다. MR에서는 이러한 위배가 일반적으로 genetic variant가 환경 요인과 상관될 때 발생하며(예: population stratification 및 assortive mating), 그 결과 instrument와 outcome 사이에 confounding이 생긴다. 동시에 MR 연구는 점점 더 여러 environment 또는 population에서 수집된 데이터를 포함하고 있어, 이러한 위배를 다룰 기회를 제공한다. 우리는 이러한 설정을 활용하여, 인과 추론에 적합한 genetic instrument의 잠재적 exogenous component를 복원하기 위해 multi-environment data를 활용하는 representation learning framework를 제안한다. 우리는 학습된 component가 언제, 그리고 어떻게 valid instrument로 작동할 수 있는지에 대한 이론적 통찰을 제공하며, All of Us Biobank의 genetic data를 사용한 simulation과 semi-synthetic experiment를 통해 제안 방법의 효과를 입증한다.

"},{"id":"64474","en":"Causal Fine-Tuning under Latent Confounded Shift","ko":"잠재 교란 변화 하에서의 인과적 미세조정","authors":"Jialin Yu, Yuxiang Zhou, Haoxuan Li, Junchi Yu, Mengyue Yang, Yulan He, Nevin Zhang, Phil Torr, Ricardo Silva","pos":"#3015","link":"https://openreview.net/forum?id=NDPvfDxsz4","abs":"

Adapting to latent confounded shift remains a core challenge in modern AI. This setting is driven by hidden variables that induce spurious correlations between inputs and outputs during training, leading models to rely on non-causal shortcuts. For example, a model may learn to treat metadata (e.g., data source like \"Amazon\") as a proxy for positive sentiment, causing failure when the source becomes predominantly negative during deployment. To address this latent confounded shift, we introduce Causal Fine-Tuning (CFT). Using a structural causal model as an inductive bias, we derive sufficient conditions under which the causal effect of inputs is identifiable (despite latent confounding), and translate these insights into a fine-tuning objective that decomposes representations into high-level causal and low-level spurious components. Instantiating this framework in BERT, we show that learning such causal/spurious representations and adjusting them accordingly yield a more robust predictor. Experiments on spurious correlation injection attacks in text demonstrate that our method outperforms black-box domain generalization baselines, highlighting the benefits of explicitly modeling causal structure.

","absKo":"

latent confounded shift에 적응하는 것은 현대 AI의 핵심 과제이다. 이 설정은 학습 중 입력과 출력 사이에 spurious correlation을 유발하여 모델이 비인과적 shortcut에 의존하게 만드는 숨겨진 변수에 의해 주도된다. 예를 들어, 모델이 메타데이터(예: \"Amazon\"과 같은 데이터 출처)를 positive sentiment의 대리 변수로 학습하면, 배포 시점에 그 출처가 주로 negative가 될 때 실패가 발생할 수 있다. 이러한 latent confounded shift를 해결하기 위해 우리는 Causal Fine-Tuning (CFT)을 제안한다. structural causal model을 inductive bias로 사용하여, latent confounding이 존재하더라도 입력의 causal effect가 식별 가능해지는 충분 조건을 도출하고, 이러한 통찰을 high-level causal component와 low-level spurious component로 representation을 분해하는 fine-tuning objective로 변환한다. 이 framework를 BERT에 구현하여, 이러한 causal/spurious representation을 학습하고 그에 맞게 조정하면 더 robust한 predictor를 얻을 수 있음을 보인다. 텍스트에서 spurious correlation injection attack에 대한 실험은 우리의 방법이 black-box domain generalization baseline보다 우수함을 보여주며, causal structure를 명시적으로 모델링하는 이점을 강조한다.

"},{"id":"62154","en":"NBCG: Nash-Bargained Causal Game for Long-Tailed Multi-Label NLP","ko":"NBCG: Long-Tailed Multi-Label NLP를 위한 Nash-Bargained Causal Game","authors":"Jing Yang, Jusheng Zhang, Keze Wang","pos":"#308","link":"https://openreview.net/forum?id=jIHoooImnn","abs":"

Long-tailed multi-label text classification is often treated as a data scarcity problem, addressed by re-sampling or fixed re-weighting. We argue that a central failure mode is \\textit{dominant coalition capture}: frequent labels, amplified by spurious co-occurrences, form dominant coalitions that dominate shared representations and gradient allocation during optimization. As a result, rare labels are learned via superficial shortcuts, yielding brittle generalization under distribution shifts. We propose \\textbf{NBCG}, a Nash-Bargained Causal Game that reformulates multi-label learning as a cooperative bargaining process among label coalitions. NBCG first leverages Neural Structural Equation Models to learn a directed dependency structure, inducing causally coherent coalitions---rather than random partitions---and coalition-specific communication masks. We then optimize a Nash bargaining objective over coalition utilities relative to an adaptive disagreement point, which serves as a principled credit-allocation mechanism: it adaptively prioritizes under-served coalitions while maintaining a Pareto-efficient trade-off among all players.

","absKo":"

long-tailed multi-label text classification은 종종 데이터 부족 문제로 간주되며, re-sampling이나 고정된 re-weighting으로 다뤄진다. 우리는 핵심 실패 양상이 \\textit{dominant coalition capture}라고 주장한다. 즉, spurious co-occurrence에 의해 증폭된 frequent label들이 dominant coalition을 형성하여, optimization 동안 shared representation과 gradient allocation을 지배한다. 그 결과 rare label은 피상적인 shortcut을 통해 학습되고, distribution shift 하에서 취약한 generalization을 보이게 된다. 우리는 \\textbf{NBCG}, 즉 Nash-Bargained Causal Game을 제안하며, 이는 multi-label learning을 label coalition 간의 cooperative bargaining process로 재정식화한다. NBCG는 먼저 Neural Structural Equation Model을 활용해 directed dependency structure를 학습하고, 무작위 partition이 아닌 causally coherent한 coalition과 coalition-specific communication mask를 유도한다. 이후 adaptive disagreement point에 대한 coalition utility 위에서 Nash bargaining objective를 최적화하는데, 이는 원칙적인 credit-allocation mechanism으로 작동한다. 즉, 모든 player 간 Pareto-efficient trade-off를 유지하면서도 under-served coalition을 적응적으로 우선시한다.

"},{"id":"62126","en":"A Recursive Decomposition Framework for Causal Structure Learning in the Presence of Latent Variables","ko":"잠재 변수가 있는 상황에서 인과 구조 학습을 위한 재귀적 분해 프레임워크","authors":"Zheng Li, Feng Xie, Shenglan Nie, Xichen Guo, Ruxin Wang, Hao Zhang","pos":"#3511","link":"https://openreview.net/forum?id=jdbQzlnYll","abs":"

Constraint-based causal discovery is widely used for learning causal structures, but heavy reliance on conditional independence (CI) testing makes it computationally expensive in high-dimensional settings. To mitigate this limitation, many divide-and-conquer frameworks have been proposed, but most assume causal sufficiency, i.e., no latent variables. In this paper, we show that divide-and-conquer strategies can be theoretically generalized beyond causal sufficiency to settings with latent variables. Specifically, we propose a recursive decomposition framework, termed DiCoLa, that enables divide-and-conquer causal discovery in the presence of latent variables. It recursively decomposes the global learning task into smaller subproblems and integrates their solutions through a principled reconstruction step to recover the global structure. We theoretically establish the soundness and completeness of the proposed framework. Extensive experiments on synthetic data demonstrate that our approach significantly improves computational efficiency across a range of causal discovery algorithms, while experiments on a real-world dataset further illustrate its practical effectiveness.

","absKo":"

Constraint-based causal discovery는 causal structure 학습에 널리 사용되지만, conditional independence (CI) testing에 크게 의존하기 때문에 고차원 설정에서 계산 비용이 큽니다. 이러한 한계를 완화하기 위해 많은 divide-and-conquer framework가 제안되었지만, 대부분은 causal sufficiency, 즉 latent variable이 없다는 가정을 합니다. 본 논문에서는 divide-and-conquer 전략이 causal sufficiency를 넘어 latent variable이 있는 설정으로 이론적으로 일반화될 수 있음을 보입니다. 구체적으로, DiCoLa라고 부르는 recursive decomposition framework를 제안하여 latent variable 존재 하에서도 divide-and-conquer causal discovery를 가능하게 합니다. 이 방법은 전역 학습 과제를 더 작은 subproblem으로 재귀적으로 분해하고, 원리 있는 reconstruction step을 통해 각 해를 통합하여 전역 구조를 복원합니다. 우리는 제안한 framework의 soundness와 completeness를 이론적으로 확립합니다. synthetic data에 대한 광범위한 실험은 우리의 접근법이 다양한 causal discovery algorithm에서 계산 효율성을 크게 향상시킴을 보여주며, real-world dataset에 대한 실험은 그 실용적 효과를 추가로 입증합니다.

"},{"id":"66168","en":"Personalized Policy Learning through Discrete Experimentation","ko":"이산적 실험을 통한 개인화 정책 학습","authors":"Zhiqi Zhang, Zhiyu Zeng, Ruohan Zhan, Dennis Zhang","pos":"#3804","link":"https://openreview.net/forum?id=698yKwEXf0","abs":"While Randomized controlled trials (RCTs), or A/B tests, are the gold standard for optimizing online-platform policies, they are limited by discrete testing levels. This approach is suboptimal for continuous variables (e.g., prices and incentives), as it fails to extrapolate to untested values or account for user heterogeneity. We address this by developing Deep Learning for Policy Targeting (\\textsf{DLPT}) to learn personalized continuous policies from discrete RCTs using high-dimensional features. We prove our estimators are asymptotically unbiased and consistent, achieving a $\\sqrt{n}$-regret bound. In a collaboration with a leading social media platform to optimize creator incentives, we show that \\textsf{DLPT} substantially outperforms existing benchmarks. In a collaboration with a leading social media platform to optimize creator incentives, we show that \\textsf{DLPT} substantially outperforms existing benchmarks.","absKo":"Randomized controlled trial (RCT), 즉 A/B test는 온라인 플랫폼 policy를 최적화하는 데 있어 gold standard이지만, 이산적인 testing level에 의해 제한된다. 이 접근법은 가격이나 인센티브와 같은 연속 변수에는 비최적이며, 미시험 값으로의 extrapolation을 하지 못하고 사용자 이질성도 반영하지 못한다. 우리는 고차원 feature를 사용해 discrete RCT로부터 personalized continuous policy를 학습하는 Deep Learning for Policy Targeting(\\textsf{DLPT})을 개발하여 이를 해결한다. 우리는 우리의 estimator가 점근적으로 unbiased이며 consistent함을 증명하고, $\\sqrt{n}$-regret bound를 달성함을 보인다. 크리에이터 인센티브 최적화를 위해 선도적인 소셜 미디어 플랫폼과 협업한 결과, \\textsf{DLPT}가 기존 benchmark를 크게 능가함을 보였다. 크리에이터 인센티브 최적화를 위해 선도적인 소셜 미디어 플랫폼과 협업한 결과, \\textsf{DLPT}가 기존 benchmark를 크게 능가함을 보였다."},{"id":"61729","en":"Toward Scalable and Valid Conditional Independence Testing with Spectral Representations","ko":"Spectral Representation을 활용한 확장 가능하고 유효한 Conditional Independence Testing을 향해","authors":"Alek Fröhlich, Vladimir Kostic, Karim Lounici, Daniel Rodrigues Perazzo, Daniel Tiezzi, Massimiliano Pontil","pos":"#4113","link":"https://openreview.net/forum?id=nPzckCXmHE","abs":"

Conditional independence (CI) is central to causal inference, feature selection, and graphical modeling, yet it is untestable in many settings without additional assumptions. Existing CI tests often rely on restrictive structural conditions, limiting their validity. Kernel methods using partial covariance operators offer a more principled approach but suffer from limited adaptivity, slow convergence, and poor scalability. In this work, we explore whether representation learning can help address these limitations. Specifically, we focus on representations derived from the singular value decomposition of partial covariance operators and use them to construct a simple test statistic. We also introduce a bi-level contrastive algorithm to learn these representations. Our theory links representation learning error to test performance and establishes asymptotic validity and power guarantees. Experiments on real and synthetic data suggest that this approach offers a principled and statistically grounded path toward scalable CI testing, bridging kernel-based theory with modern representation learning.

","absKo":"

Conditional independence (CI)는 causal inference, feature selection, graphical modeling의 핵심이지만, 추가 가정 없이 많은 설정에서 검정할 수 없습니다. 기존 CI test는 종종 제한적인 구조적 조건에 의존하여 유효성이 제약됩니다. partial covariance operator를 사용하는 kernel method는 더 원리적인 접근법을 제공하지만, 적응성이 제한되고 수렴이 느리며 확장성이 떨어집니다. 본 연구에서는 representation learning이 이러한 한계를 해소하는 데 도움이 되는지 탐구합니다. 구체적으로는 partial covariance operator의 singular value decomposition에서 얻은 representation에 초점을 맞추고, 이를 사용해 단순한 test statistic을 구성합니다. 또한 이러한 representation을 학습하기 위한 bi-level contrastive algorithm도 제안합니다. 우리의 이론은 representation learning error와 test performance를 연결하고, auxiliary reasoning chain이 관찰되는 설정에서 asymptotic validity와 power 보장을 확립합니다. 실제 및 synthetic data 실험은 이 접근법이 scalable CI testing을 향한 원리적이고 통계적으로 근거 있는 경로를 제공하며, kernel-based theory와 현대 representation learning을 연결함을 시사합니다.

"},{"id":"60623","en":"Multi-View Causal Discovery without Non-Gaussianity: Identifiability and Algorithms","ko":"비가우시안성 없는 멀티뷰 인과 발견: 식별 가능성과 알고리즘","authors":"Ambroise Heurtebise, Lemir Omar Chehab, Pierre Ablin, Alexandre Gramfort, Aapo Hyvarinen","pos":"#4201","link":"https://openreview.net/forum?id=yMKdZPwPyL","abs":"

Causal discovery is a difficult problem that typically relies on strong assumptions on the data-generating model, such as non-Gaussianity. In practice, many modern applications provide multiple related views of the same system, which has rarely been considered for causal discovery. Here, we leverage this multi-view structure to achieve causal discovery with weak assumptions. We propose a multi-view linear Structural Equation Model (SEM) that extends the well-known framework of non-Gaussian disturbances by alternatively leveraging correlation over views. We prove the identifiability of the model for acyclic SEMs. Subsequently, we propose several multi-view causal discovery algorithms, inspired by single-view algorithms (DirectLiNGAM, PairwiseLiNGAM, and ICA-LiNGAM). The new methods are validated through simulations and applications on neuroimaging data, where they enable the estimation of causal graphs between brain regions.

","absKo":"

인과 발견은 보통 non-Gaussianity와 같은 데이터 생성 모델에 대한 강한 가정에 의존하는 어려운 문제이다. 실제로 많은 현대 응용에서는 동일한 시스템의 서로 관련된 여러 view를 제공하지만, 인과 발견에서 이러한 다중 view 구조는 거의 고려되지 않았다. 여기서는 약한 가정하에서 인과 발견을 달성하기 위해 이 multi-view 구조를 활용한다. 우리는 view 간 상관을 대안적으로 활용함으로써 잘 알려진 non-Gaussian disturbance 프레임워크를 확장하는 multi-view linear Structural Equation Model (SEM)을 제안한다. 또한 acyclic SEM에 대해 이 모델의 identifiability를 증명한다. 이어서 single-view 알고리즘(DirectLiNGAM, PairwiseLiNGAM, ICA-LiNGAM)에서 영감을 받은 여러 multi-view causal discovery 알고리즘을 제안한다. 새로운 방법들은 시뮬레이션과 neuroimaging 데이터 응용을 통해 검증되며, 이를 통해 뇌 영역 간 causal graph를 추정할 수 있음을 보인다.

"},{"id":"61193","en":"Nonlinear Covariate Balance in Experimental Design","ko":"실험 설계에서의 비선형 공변량 균형","authors":"Qing Chen, Peng Zhang","pos":"#4202","link":"https://openreview.net/forum?id=srhmyF8k2w","abs":"

We study experimental designs that balance nonlinear functions of covariates, extending classical methods that primarily target linear balance. Building on the Gram-Schmidt Walk (GSW) framework of Harshaw et al (2024) for linear covariate balancing, we introduce a design that directly controls imbalance in nonlinear structure, including polynomial and more general smooth function classes. Like GSW, the proposed design retains sufficient robustness against model misspecification. Our implementation operates directly on a Gram matrix, avoiding the expensive step of explicitly constructing the nonlinear covariate expansions. We further accelerate the nonlinear design via a low-rank approximation of the Gram matrix, achieving runtimes comparable to the GSW of Harshaw et al (2024) while preserving nonlinear covariate balance and robustness.

","absKo":"

우리는 covariate의 nonlinear function을 균형 있게 맞추는 experimental design을 연구하며, 주로 linear balance를 목표로 했던 고전적 방법을 확장한다. 선형 covariate balancing을 위한 Harshaw et al (2024)의 Gram-Schmidt Walk (GSW) framework를 기반으로, 우리는 polynomial 및 더 일반적인 smooth function class를 포함하는 nonlinear structure의 imbalance를 직접 제어하는 design을 제안한다. GSW와 마찬가지로, 제안하는 design은 model misspecification에 대한 충분한 robustness를 유지한다. 우리의 구현은 Gram matrix 위에서 직접 동작하므로, nonlinear covariate expansion을 명시적으로 구성하는 비용이 큰 단계를 피한다. 또한 Gram matrix의 low-rank approximation을 통해 nonlinear design을 더욱 가속하여, Harshaw et al (2024)의 GSW와 유사한 runtime을 달성하면서도 nonlinear covariate balance와 robustness를 보존한다.

"},{"id":"65686","en":"Off-Policy Evaluation Beyond Overlap under Network Interference","ko":"네트워크 간섭 하에서 중첩을 넘어선 오프폴리시 평가","authors":"Ziyan Wang, Yiran Liu, Zhiheng Zhang","pos":"#4203","link":"https://openreview.net/forum?id=BAxL3dqhLn","abs":"

Offline Policy Evaluation (OPE) aims to estimate the value of a target policy from historical logged data without interating with the environment, thereby assessing policy performance. In settings with network interference, individuals no longer satisfy the SUTVA assumption: an individual’s outcome is influenced not only by their own treatment but also by the treatments of their neighbors, which makes the definition and estimation of policy value more complex. To capture this interference mechanism, we allow all neighbors to affect individual outcomes through a unified exposure mapping, and we use a decaying higher-order neighborhood aggregation to characterize the influence of more distant neighbors. Moreover, in real-world applications, the target policy and the logging policy often do not fully overlap (non-overlap), so the policy value in non-overlap regions cannot be point-identified. To address this issue, we partially identify the policy value over non-overlap regions and, under a smoothness assumption, formulate the estimation of the lower and upper bounds as a linear program, yielding valid bounds on the offline policy value. Finally, we conduct systematic experiments on semi-synthetic network data to validate the effectiveness and robustness of the proposed method under network interference and limited overlap.

","absKo":"

Offline Policy Evaluation(OPE)은 환경과 상호작용하지 않고 과거에 기록된 logged data로부터 target policy의 가치를 추정함으로써 policy performance를 평가하는 것을 목표로 한다. network interference가 있는 설정에서는 개체들이 더 이상 SUTVA 가정을 만족하지 않는다. 즉, 한 개체의 결과는 자신의 treatment뿐 아니라 이웃의 treatment에도 영향을 받으며, 이로 인해 policy value의 정의와 추정이 더 복잡해진다. 이러한 interference mechanism을 포착하기 위해, 우리는 모든 이웃이 통합된 exposure mapping을 통해 개체 결과에 영향을 미치도록 허용하고, 더 먼 이웃의 영향을 특성화하기 위해 decaying higher-order neighborhood aggregation을 사용한다. 또한 실제 응용에서는 target policy와 logging policy가 종종 완전히 겹치지 않기 때문에(non-overlap), non-overlap 구간에서의 policy value는 point-identified 될 수 없다. 이 문제를 해결하기 위해, 우리는 non-overlap 구간에 대해 policy value를 partially identify하고, smoothness assumption 하에서 lower 및 upper bound의 추정을 linear program으로 정식화하여 offline policy value에 대한 valid bound를 도출한다. 마지막으로, network interference와 limited overlap 하에서 제안 방법의 효과와 견고성을 검증하기 위해 semi-synthetic network data에서 체계적인 실험을 수행한다.

"},{"id":"67225","en":"Position: Causality is Key for Interpretability Claims to Generalise","ko":"Position: 인과성은 해석 가능성 주장이 일반화되기 위한 핵심이다","authors":"Shruti Joshi, Aaron Mueller, David Klindt, Wieland Brendel, Dhanya Sridhar, Patrik Reizinger","pos":"#4205","link":"https://openreview.net/forum?id=6l7cKIR6nf","abs":"

Interpretability research on large language models (LLMs) has produced methods that align model components to high-level concepts, yet their use has been accompanied by recurring failures: findings that do not generalise, and causal language that outruns the evidence. Our position is that Pearl’s causal hierarchy formally defines what constitutes a good alignment, what data or assumptions it requires, and what inferences it supports. Specifically, observations of model behaviour support only associational claims; interventions enable cause-effect claims, but not necessarily predictions of model behaviour; counterfactuals, or predictions of behaviour on unseen examples, are often unverifiable in current studies. We show how interpretability research can benefit from causal representation learning (CRL), which provides tools for provably extracting semantic variables and their relationships from activations, and outline practical requirements for generalisable insights: robustness to distribution shifts, sensitivity to assumptions, and compositionality of interventions. Our diagnostic framework helps practitioners select appropriate methods and mitigate failures to ensure that claims match evidence and findings generalise.

","absKo":"

large language model (LLM)에 대한 interpretability 연구는 model component를 고수준 개념에 정렬시키는 방법들을 만들어 냈지만, 그 사용은 반복적인 실패와 함께해 왔다: 일반화되지 않는 발견, 그리고 증거를 앞서는 인과적 언어. 우리의 입장은 Pearl의 causal hierarchy가 무엇이 좋은 alignment를 구성하는지, 어떤 data나 assumption이 필요한지, 그리고 어떤 inference를 지지하는지를 형식적으로 정의한다는 것이다. 구체적으로, model behavior에 대한 observation은 associational claim만을 지지하며; intervention은 cause-effect claim을 가능하게 하지만, 반드시 model behavior 예측을 의미하지는 않는다; counterfactual, 즉 보지 못한 example에서의 behavior 예측은 현재의 연구에서 종종 검증 불가능하다. 우리는 interpretability 연구가 causal representation learning (CRL)로부터 혜택을 받을 수 있음을 보이고, CRL이 activation에서 semantic variable과 그 관계를 증명 가능하게 추출하는 도구를 제공함을 설명하며, 일반화 가능한 통찰을 위한 실천적 요구사항 - distribution shift에 대한 강건성, assumption에 대한 민감도, intervention의 compositionality - 을 제시한다. 우리의 진단 프레임워크는 실무자가 적절한 방법을 선택하고 실패를 완화하여, 주장과 증거가 일치하고 발견이 일반화되도록 돕는다.

"},{"id":"65359","en":"Root Cause Analysis of Failures in Microservices via Bayesian Root Cause Discovery","ko":"Bayesian 근본 원인 발견을 통한 마이크로서비스 장애의 근본 원인 분석","authors":"Kenneth Lee, Zihan Zhou, Murat Kocaoglu","pos":"#4209","link":"https://openreview.net/forum?id=EOSV5tlpqg","abs":"Modern cloud systems rely on architectures with many interconnected microservices, which enable scalability and flexibility but make troubleshooting failures difficult. Identifying the root cause requires navigating complex dependencies, often beyond the capacity of domain experts. Causal models offer a principled approach to root cause analysis (RCA), but prior methods are typically sample inefficient, as they assume access to the full causal graph or require large numbers of post-failure interventions. We introduce Bayesian Root Cause Discovery (BRCD), which leverages a partial causal structure (a CPDAG learned during the pre-failure period) and performs Bayesian inference without enumerating all DAGs from each interventional Markov equivalence class ($\\mathcal{I}$-MEC) for each root cause candidate. Using a recent uniform DAG sampling framework (Wienöbst et al., 2023), BRCD provides the first statistical consistency guarantees for nonparametric RCA, with both identifiability and finite-sample posterior bounds under $\\varepsilon$-vanishing approximation. Empirically, across synthetic benchmarks and three microservice systems (Online Boutique, Sockshop, Petshop), BRCD achieves state-of-the-art top-$l$ accuracy while remaining effective in low-failure-sample regimes and scaling to large graphs.","absKo":"현대 클라우드 시스템은 수많은 상호 연결된 microservice로 구성된 아키텍처에 의존하는데, 이는 확장성과 유연성을 제공하지만 장애 트러블슈팅을 어렵게 만든다. 근본 원인을 식별하려면 복잡한 의존성을 따라가야 하며, 이는 종종 도메인 전문가의 역량을 넘어선다. causal model은 root cause analysis(RCA)를 위한 원칙적인 접근을 제공하지만, 기존 방법은 대개 전체 causal graph에 대한 접근을 가정하거나 장애 이후의 개입을 대량으로 요구하기 때문에 sample inefficiency가 크다. 우리는 Bayesian Root Cause Discovery(BRCD)를 제안하는데, 이는 부분적인 causal structure(장애 이전 기간에 학습된 CPDAG)를 활용하고, 각 root cause candidate마다 각 interventional Markov equivalence class($\\mathcal{I}$-MEC)의 모든 DAG를 열거하지 않고 Bayesian inference를 수행한다. 최근의 uniform DAG sampling framework(Wienöbst et al., 2023)를 사용하여, BRCD는 비모수 RCA를 위한 최초의 statistical consistency 보장을 제공하며, $\\varepsilon$-vanishing approximation 하에서 identifiability와 유한 표본 posterior bound를 모두 갖는다. 경험적으로, 합성 benchmark와 세 개의 microservice 시스템(Online Boutique, Sockshop, Petshop) 전반에서 BRCD는 저장애 샘플 환경에서도 효과적으로 작동하고 대규모 graph로 확장되면서 state-of-the-art top-$l$ 정확도를 달성한다."},{"id":"60689","en":"SCOUT: Cyclic Causal Discovery Under Soft Interventions with Unknown Targets","ko":"SCOUT: Unknown Target을 갖는 Soft Intervention 하의 Cyclic Causal Discovery","authors":"Alpar Turkoglu, Muralikrishnna Guruswamy Sethuraman, Faramarz Fekri","pos":"#4210","link":"https://openreview.net/forum?id=xXyreH61EN","abs":"

Learning causal relationships between variables from data is a fundamental research area with many applications across disciplines. Most of the existing causal discovery algorithms rely on the assumptions that (i) the underlying system is acyclic, (ii) the exogenous noise variables are Gaussian, and (iii) that the intervention targets for the data generating experiments are known. While these assumptions simplify the analysis, they are violated in real-life systems. Most existing methods that address these issues either assume the underlying model is linear or are constrained to operate in limited interventional settings. To that end, we propose SCOUT, a novel causal discovery framework to learn nonlinear causal cyclic relationships from soft interventional data with unknown targets. Our main approach maximizes the data log-likelihood to recover the graph structure, using two normalizing-flow architectures—contractive residual flows and neural spline flows. By conducting experiments on synthetic and real-world data, we show that SCOUT outperforms state-of-the-art methods in both causal graph and unknown target recovery across various interventional and noise settings.

","absKo":"

변수 간 causal relationship을 데이터로부터 학습하는 것은 다양한 분야에 걸쳐 많은 응용을 갖는 기초 연구 분야이다. 기존 causal discovery algorithm의 대부분은 (i) underlying system이 acyclic하고, (ii) exogenous noise variable이 Gaussian이며, (iii) 데이터 생성 실험의 intervention target이 알려져 있다는 가정에 의존한다. 이러한 가정은 분석을 단순화하지만 실제 시스템에서는 위배된다. 이러한 문제를 다루는 기존 방법들은 대부분 underlying model이 linear라고 가정하거나 제한된 interventional setting에서만 동작하도록 제약된다. 이에 따라 우리는 unknown target을 가진 soft interventional data로부터 nonlinear causal cyclic relationship을 학습하는 새로운 causal discovery framework인 SCOUT를 제안한다. 우리의 핵심 접근은 두 개의 normalizing-flow architecture, 즉 contractive residual flow와 neural spline flow를 사용하여 data log-likelihood를 최대화함으로써 graph structure를 복원한다. synthetic 및 real-world data에 대한 실험을 통해, SCOUT가 다양한 interventional 및 noise setting에서 causal graph 복원과 unknown target 복원 모두에서 state-of-the-art method를 능가함을 보인다.

"},{"id":"62976","en":"Smooth Multi-Policy Causal Effect Estimation in Longitudinal Settings","ko":"종단 간 데이터에서 Smooth Multi-Policy Causal Effect Estimation","authors":"Wenxin Chen, Weishen Pan, Kyra Gan, Fei Wang","pos":"#4211","link":"https://openreview.net/forum?id=bIcz7bIZSo","abs":"

Comparative evaluation of multiple dynamic treatment policies is essential for healthcare and policy decisions, yet conventional longitudinal causal inference methods estimate each in isolation, preventing information sharing across counterfactuals. We demonstrate that this separate estimation paradigm induces a structurally uncontrolled second-order bias, inflating finite-sample variance even after standard debiasing with longitudinal targeted maximum likelihood estimation (LTMLE). To address this, we propose a policy-aware reparameterization of Iterative Conditional Expectation (ICE) Q-functions that enables joint estimation through shared representations. We implement this approach in the Policy-Encoded Q Network (PEQ-Net), an architecture centered on a shared policy encoder. The encoder is trained using kernel mean embeddings, ensuring that the learned representation space reflects population-level policy dissimilarities. After applying an LTMLE correction step, we prove this design imposes a structural constraint on the second-order remainder, thereby stabilizing finite-sample variance. Experiments on semi-synthetic datasets demonstrate that PEQ-Net consistently outperforms existing ICE-based methods, reducing root-mean-square error by 13–62\\%.

","absKo":"

multiple dynamic treatment policy의 비교 평가는 의료 및 정책 의사결정에 필수적이지만, 전통적인 longitudinal causal inference 방법은 각 정책을 고립적으로 추정하므로 counterfactual 간 정보 공유가 불가능합니다. 우리는 이러한 분리 추정 패러다임이 구조적으로 통제되지 않은 second-order bias를 유발하여, longitudinal targeted maximum likelihood estimation (LTMLE)로 표준적인 debiasing을 수행한 뒤에도 finite-sample variance를 증폭시킨다는 점을 보입니다. 이를 해결하기 위해 우리는 공유 표현을 통한 joint estimation을 가능하게 하는 Iterative Conditional Expectation (ICE) Q-function의 policy-aware reparameterization을 제안합니다. 우리는 이를 공유 policy encoder를 중심으로 한 Policy-Encoded Q Network (PEQ-Net)에 구현합니다. 이 encoder는 kernel mean embedding을 사용해 학습되며, 이를 통해 학습된 representation space가 population-level policy dissimilarity를 반영하도록 보장합니다. LTMLE correction step을 적용한 뒤, 우리는 이 설계가 second-order remainder에 구조적 제약을 부과함을 증명하여 finite-sample variance를 안정화합니다. semi-synthetic dataset에서의 실험은 PEQ-Net이 기존 ICE 기반 방법을 일관되게 능가하며 root-mean-square error를 13–62\\% 줄인다는 것을 보여줍니다.

"},{"id":"64597","en":"Spiked-CFR: Causal Representation Learning from LLMs via Wasserstein Projection Pursuit","ko":"Spiked-CFR: Wasserstein Projection Pursuit를 통한 LLM으로부터의 Causal Representation 학습","authors":"Fan Wang, Hengyu Yue, Yu Bowen, Weiming Liu, Zongxin Yang, Xuyun Zhang, Xiaolin Zheng, Chaochao Chen, Shuiguang Deng","pos":"#4213","link":"https://openreview.net/forum?id=LqXY6aCtkA","abs":"Estimating treatment effects from observational text is increasingly practical with Large Language Models (LLMs). However, applying causal representation learning directly to high-dimensional LLM embeddings faces a fundamental barrier: empirical Wasserstein matching suffers from the curse of dimensionality, rendering standard generalization guarantees effectively vacuous. We propose SPIKED-CFR, a framework bridging this gap by assuming a Spiked Confounding Structure, where treatment selection bias concentrates in a low-dimensional subspace of the semantic representation. We develop Wasserstein Projection Pursuit, a minimax objective that adversarially learns an orthogonal projection on the Stiefel manifold to identify and balance only this confounding subspace while preserving prognostic information. Under a spiked confounding structure, we show the projected discrepancy can be estimated at a rate governed by the intrinsic dimension $k \\ll D$, and we derive a tighter PEHE generalization bound that depends on $k$ rather than the ambient embedding dimension. Experiments on four semi-synthetic benchmarks and four real-world clinical benchmarks demonstrate improved accuracy and robustness over strong baselines. Code is available at \\url{https://anonymous.4open.science/r/SpikedCFR-7E13}.","absKo":"관찰된 text로부터 treatment effect를 추정하는 일은 Large Language Models (LLMs) 덕분에 점점 더 실용적이 되고 있다. 그러나 causal representation learning을 고차원 LLM embedding에 직접 적용하면 근본적인 장벽에 부딪힌다. 경험적 Wasserstein matching은 차원의 저주를 겪어, 표준 generalization guarantee가 사실상 공허해진다. 우리는 treatment selection bias가 semantic representation의 저차원 subspace에 집중된다는 Spiked Confounding Structure를 가정함으로써 이 간극을 메우는 프레임워크인 SPIKED-CFR를 제안한다. 우리는 Stiefel manifold 위에서 orthogonal projection을 적대적으로 학습하여 이 confounding subspace만 식별하고 균형을 맞추면서 prognostic information은 보존하는 minimax objective인 Wasserstein Projection Pursuit를 개발한다. spiked confounding structure 하에서, projected discrepancy는 intrinsic dimension $k \\ll D$에 의해 지배되는 rate로 추정될 수 있음을 보이고, ambient embedding dimension이 아니라 $k$에 의존하는 더 타이트한 PEHE generalization bound를 유도한다. 네 개의 semi-synthetic benchmark와 네 개의 실제 clinical benchmark에서 수행한 실험은 강력한 baseline들보다 향상된 정확도와 robustness를 보여준다. 코드는 \\url{https://anonymous.4open.science/r/SpikedCFR-7E13}에서 확인할 수 있다."},{"id":"64305","en":"Streaming Covariate Balancing via Discrepancy-Based Feature Coresets","ko":"불일치 기반 특징 코어셋을 통한 스트리밍 공변량 균형화","authors":"YiXin Ren, Chenghou Jin, Yewei Xia, Zichuan Lin, Deheng Ye, Hao Zhang, Jihong Guan, Shuigeng Zhou","pos":"#4214","link":"https://openreview.net/forum?id=Ovk43Z9KGC","abs":"

Real-time estimation of average treatment effects (ATE) in streaming observational data poses two key challenges: strict memory constraints that preclude storing the full data history, and distributional shifts in both treatment assignment and outcome-generating process. Existing methods either require offline access to the entire dataset for covariate balancing or rely on parametric online models that are vulnerable to model misspecification under such shifts. This paper proposes a novel model-agnostic method for ATE estimation in streaming data, which effectively addresses the above challenges. Based on discrepancy theory, we first compress streaming data into feature coresets that preserve covariate balancing objectives over a rich nonparametric function class, enabling linear-time updates with bounded memory. Then, by directly learning balancing weights and bypassing parametric propensity score estimation, we enhance the model's robustness against the shift in treatment assignment, while by balancing over an expressive function space we make the model more adaptive to the shift in the outcome-generating process. Theoretically, we establish convergence guarantees with explicit bounds on memory usage and computational complexity. Empirically, extensive experiments on both synthetic and real-world datasets show the effectiveness and robustness of the proposed method, consistently outperforming existing techniques.

","absKo":"

스트리밍 관측 데이터에서 average treatment effects (ATE)를 실시간으로 추정하는 문제는 두 가지 핵심 과제를 수반한다: 전체 데이터 이력을 저장할 수 없게 만드는 엄격한 메모리 제약과, treatment assignment 및 outcome-generating process 양쪽에서 발생하는 distributional shift이다. 기존 방법들은 covariate balancing을 위해 전체 데이터셋에 대한 offline 접근을 요구하거나, 이러한 shift 하에서 model misspecification에 취약한 parametric online model에 의존한다. 본 논문은 스트리밍 데이터에서 ATE 추정을 위한 새로운 model-agnostic 방법을 제안하며, 위의 과제들을 효과적으로 해결한다. discrepancy theory에 기반하여, 우리는 먼저 스트리밍 데이터를 풍부한 nonparametric function class에 걸친 covariate balancing objective를 보존하는 feature coreset으로 압축함으로써, bounded memory 하에서 linear-time update를 가능하게 한다. 이어서, balancing weight를 직접 학습하고 parametric propensity score estimation을 우회함으로써 treatment assignment의 shift에 대한 model의 robustness를 높이고, expressive function space에서 balancing을 수행함으로써 outcome-generating process의 shift에 더 잘 적응하도록 한다. 이론적으로는 메모리 사용량과 계산 복잡도에 대한 explicit bound를 포함한 convergence guarantee를 제시한다. 실험적으로는 synthetic 및 real-world dataset 전반에 걸친 광범위한 실험에서 제안 방법의 효과성과 robustness를 입증하며, 기존 기법들을 일관되게 상회함을 보인다.

"},{"id":"65152","en":"MiniMax Learning of Interpretable Factored Stochastic Policies from Conjoint Data, with Uncertainty Quantification","ko":"결합 데이터로부터 해석 가능한 인수분해 확률적 정책의 MiniMax 학습과 불확실성 정량화","authors":"Connor T Jerzak, Priyanshi Chandra, Rishi Hazra","pos":"#4300","link":"https://openreview.net/forum?id=GJblFvJcMb","abs":"We study offline learning of factored stochastic policies over extremely large, combinatorial action spaces and show how standard conjoint data can be used to estimate such policies with asymptotically valid uncertainty under conditions. Conjoint analyses typically report AMCEs by averaging over opponent attributes and thus ignore strategic interdependence. We instead learn \\emph{stochastic interventions}---product-of-Categorical policies over factor levels—that (i) optimize expected outcomes in an average-case setting and (ii) extend to a two-player \\emph{minimax} (adversarial) setting that realistically captures simultaneous strategic candidate selection. Methodologically, we derive a closed-form solution for the average-case optimizer under two-way interactions with $L_2$ variance regularization, and provide a general gradient-based procedure for richer model classes. Uncertainty from the outcome model propagates asymptotically to both the optimal policy and its value via a Delta method approximation. We further model institutional details (e.g., primaries) inside the minimax objective and introduce a data-driven measure of strategic divergence between parties. On synthetic data, we empirically characterize finite-sample error and coverage as dimensionality and $n$ vary. On a U.S. presidential conjoint, adversarially learned policies produce restricted-equilibrium vote shares that align with historical election ranges in our data, in stark contrast to non-adversarial (averaging) optimizers.","absKo":"우리는 매우 큰 조합적 action space에서 factorized stochastic policies의 offline learning을 연구하고, 표준 conjoint data를 사용해 조건부로 asymptotically valid uncertainty를 지닌 정책을 추정할 수 있음을 보인다. Conjoint 분석은 일반적으로 상대 후보의 속성에 대해 평균을 내어 AMCE를 보고하므로 전략적 상호의존성을 무시한다. 우리는 대신 stochastic interventions, 즉 factor level에 대한 product-of-Categorical policies를 학습하여 (i) average-case setting에서 기대 성과를 최적화하고, (ii) 동시에 전략적 후보 선택을 현실적으로 포착하는 two-player minimax(adversarial) setting으로 확장한다. 방법론적으로, 우리는 $L_2$ variance regularization이 있는 two-way interactions 하에서 average-case optimizer의 closed-form solution을 유도하고, 더 풍부한 model class를 위한 일반적인 gradient-based procedure를 제시한다. Outcome model에서의 uncertainty는 Delta method approximation을 통해 optimal policy와 그 value 모두에 asymptotically 전파된다. 또한 우리는 제도적 세부사항(예: primary elections)을 minimax objective 안에 모델링하고, 정당 간 전략적 divergence를 측정하는 data-driven measure를 도입한다. Synthetic data에서는 dimensionality와 $n$이 변할 때 finite-sample error와 coverage를 실증적으로 특성화한다. U.S. presidential conjoint에서는 adversarially learned policies가 비adversarial(averaging) optimizer와 뚜렷이 대비되게, 우리 데이터에서 역사적 선거 범위와 부합하는 restricted-equilibrium vote share를 산출한다."},{"id":"64110","en":"Formalizing and Falsifying Causal Pathways of Rare Events","ko":"희귀 사건의 인과 경로 형식화 및 반증","authors":"Anahita Haghighat, Dominik Janzing","pos":"#4304","link":"https://openreview.net/forum?id=QoInle2UdN","abs":"

Building on recent formalizations of root cause analysis for rare events (“outliers”) in structural equation models, we propose a formal definition of a causal pathway and discuss its testable implications. We identify conditions under which these implications depend only on a causal abstraction defined by the pathway of rare events, rather than on the full causal graph of the underlying system. Accordingly, we introduce an abstraction of causal structure to pathways of rare events that bridges simple verbal causal explanations and detailed causal modeling.

","absKo":"

희귀 사건(“outlier”)에 대한 root cause analysis의 최근 formalization을 바탕으로, 우리는 causal pathway의 형식적 정의를 제안하고 그 검증 가능한 함의를 논의한다. 우리는 이러한 함의가 underlying system의 전체 causal graph가 아니라, 희귀 사건의 pathway에 의해 정의된 causal abstraction에만 의존하는 조건을 식별한다. 이에 따라 우리는 단순한 언어적 causal explanation과 세부적인 causal modeling을 연결하는, 희귀 사건 pathway에 대한 causal structure의 abstraction을 도입한다.

"},{"id":"60651","en":"Federated Causal Inference on Multi-Site Observational Data via Propensity Score Aggregation","ko":"Propensity Score Aggregation을 통한 다기관 관측 데이터의 Federated Causal Inference","authors":"Rémi Khellaf, Aurélien Bellet, julie Josse","pos":"#4305","link":"https://openreview.net/forum?id=y3qyP3ycQi","abs":"

Causal inference typically assumes centralized access to individual-level data. Yet, in practice, data are often decentralized across multiple sites, making centralization infeasible due to privacy, logistical, or legal constraints. We address this problem by estimating the Average Treatment Effect (ATE) from decentralized observational data via a Federated Learning (FL) approach, allowing inference through the exchange of aggregate statistics rather than individual-level data. We propose a novel method to estimate propensity scores via a federated weighted average of local scores using Membership Weights (MW), defined as probabilities of site membership conditional on covariates. MW can be flexibly estimated with parametric or non-parametric classification models using standard FL algorithms. The resulting propensity scores are used to construct Federated Inverse Propensity Weighting (Fed-IPW) and Augmented IPW (Fed-AIPW) estimators. In contrast to meta-analysis methods, which fail when any site violates positivity, our approach exploits heterogeneity in treatment assignment across sites to improve overlap. We show that Fed-IPW and Fed-AIPW perform well under site-level heterogeneity in sample sizes, treatment mechanisms, and covariate distributions. Theoretical analysis and experiments on simulated and real-world data demonstrate clear advantages over meta-analysis and related approaches.

","absKo":"Causal inference는 일반적으로 individual-level data에 대한 중앙집중식 접근을 가정한다. 그러나 실제로 데이터는 여러 site에 분산되어 있는 경우가 많아, privacy, 물류, 법적 제약 때문에 중앙집중화가 불가능하다. 우리는 Federated Learning (FL) 접근을 통해 decentralized observational data로부터 Average Treatment Effect (ATE)를 추정함으로써 이 문제를 다룬다. 이 방식은 individual-level data가 아니라 aggregate statistics의 교환을 통해 추론을 가능하게 한다. 우리는 Membership Weights (MW)를 사용하여 local score의 federated weighted average로 propensity score를 추정하는 새로운 방법을 제안한다. MW는 covariate가 주어졌을 때의 site membership probability로 정의된다. MW는 표준 FL algorithm을 사용하여 parametric 또는 non-parametric classification model로 유연하게 추정할 수 있다. 이렇게 얻은 propensity score는 Federated Inverse Propensity Weighting (Fed-IPW) 및 Augmented IPW (Fed-AIPW) estimator를 구성하는 데 사용된다. 어떤 site라도 positivity를 위반하면 실패하는 meta-analysis 방법과 달리, 우리의 접근은 site 간 treatment assignment의 heterogeneity를 활용하여 overlap을 개선한다. 우리는 Fed-IPW와 Fed-AIPW가 sample size, treatment mechanism, covariate distribution의 site-level heterogeneity 하에서 우수하게 동작함을 보인다. 이론 분석과 시뮬레이션 및 실제 데이터 실험은 meta-analysis 및 관련 접근보다 분명한 우위를 입증한다.

"},{"id":"62092","en":"Doubly Robust Distributionally Robust Offline Contextual Pricing","ko":"이중 강건 분포강건 오프라인 Contextual Pricing","authors":"Min Xu, Xinyi Yin, Caihua Chen, Yuxuan Han, Houcai Shen, Yunfan Zhang","pos":"#4307","link":"https://openreview.net/forum?id=jsp5fGi1a5","abs":"Offline contextual pricing often relies on logged observational data, but faces challenges from distributional shifts between training and deployment environments. Distributionally robust optimization (DRO) provides a principled approach to off-policy evaluation and learning (OPE/L). However, existing methods are mostly limited to discrete actions. Recent work has explored DRO for continuous treatments using inverse propensity weighting (IPW), while such IPW-based estimators can be sensitive to the convergence rate of propensity score estimates, particularly when estimated nonparametrically, which may lead to larger estimation errors and regret. In this work, we develop a doubly robust (DR) framework for distributionally robust OPE/L in continuous pricing settings. For evaluation, we propose a localized DR estimator that addresses the computational challenges of worst-case expectations by fitting only a small number of regressions, comparable to standard non-robust DR, while achieving semiparametric efficiency under mild product rate conditions. For learning, we leverage the inherent smoothness of demand noise to handle pricing-specific discontinuities in revenue outcomes (e.g., threshold-based purchase decisions), establishing a finite-sample regret bound of $\\tilde{\\mathcal{O}}_p(T^{-s/(2s+1)})$ for smoothness orders $s=1,2$. This bound improves upon existing regret rates in existing DRO-based off-policy learning (OPL) for continuous treatments. Extensive experiments under various levels of distribution shift validate our proposed framework.","absKo":"오프라인 contextual pricing은 종종 logged observational data에 의존하지만, training과 deployment 환경 사이의 distribution shift로 인한 도전에 직면한다. Distributionally robust optimization (DRO)은 off-policy evaluation and learning (OPE/L)을 위한 원리적인 접근을 제공한다. 그러나 기존 방법은 대부분 discrete action에 제한된다. 최근 연구는 inverse propensity weighting (IPW)을 사용해 continuous treatment에 대한 DRO를 탐구했지만, 이러한 IPW 기반 estimator는 특히 nonparametrically 추정될 때 propensity score 추정의 convergence rate에 민감할 수 있으며, 이는 더 큰 estimation error와 regret로 이어질 수 있다. 본 연구에서는 continuous pricing setting에서의 distributionally robust OPE/L을 위한 doubly robust (DR) framework를 개발한다. 평가를 위해, 우리는 standard non-robust DR과 비슷한 소수의 regression만 적합하면서 worst-case expectation의 계산상 어려움을 해결하고, mild한 product rate 조건하에서 semiparametric efficiency를 달성하는 localized DR estimator를 제안한다. 학습을 위해서는 demand noise의 내재적 smoothness를 활용하여 threshold 기반 purchase decision과 같은 pricing 특유의 revenue outcome discontinuity를 처리하고, smoothness order가 $s=1,2$일 때 finite-sample regret bound $\\tilde{\\mathcal{O}}_p(T^{-s/(2s+1)})$를 확립한다. 이 bound는 continuous treatment에 대한 기존 DRO-based off-policy learning (OPL)의 기존 regret rate를 개선한다. 다양한 수준의 distribution shift 하에서 수행한 광범위한 실험은 제안한 framework를 검증한다."},{"id":"64421","en":"Budgeted Active Experimentation for Treatment Effect Estimation from Observational and Randomized Data","ko":"관측 및 무작위 데이터로부터 처치 효과 추정을 위한 예산 제약 능동 실험","authors":"Jiacan Gao, Xinyan Su, Mingyuan Ma, Yiyan HUANG, Xiao Xu, Xinrui Wan, Tianqi Gu, Enyun Yu, Jiecheng Guo, Zhiheng Zhang","pos":"#4313","link":"https://openreview.net/forum?id=Np7Y3AEVNq","abs":"

Estimating heterogeneous treatment effects is central to data-driven decision-making, yet industrial applications often face a fundamental tension between limited randomized controlled trial (RCT) budgets and abundant but biased observational data collected under historical targeting policies. Although observational logs offer the advantage of scale, they inherently suffer from severe policy-induced imbalance and overlap violations, rendering standalone estimation unreliable. We propose a budgeted active experimentation framework that iteratively enhances model training for causal effect estimation via active sampling. By leveraging observational priors, we develop an acquisition function targeting uplift estimation uncertainty, overlap deficits, and domain discrepancy to select the most informative units for randomized experiments. We establish finite-sample deviation bounds, asymptotic normality via martingale Central Limit Theorems (CLTs), and minimax lower bounds to prove information-theoretic optimality. Extensive experiments on industrial datasets demonstrate that our approach significantly outperforms standard randomized baselines in cost-constrained settings.

","absKo":"이질적인 treatment effect를 추정하는 일은 data-driven decision-making의 핵심이지만, 산업 응용에서는 제한된 randomized controlled trial(RCT) 예산과 역사적 targeting policy 하에서 수집된 풍부하지만 편향된 observational data 사이의 근본적인 긴장이 자주 발생한다. observational log는 규모의 장점을 제공하지만, 본질적으로 심각한 policy-induced imbalance와 overlap violation을 겪기 때문에 단독 추정은 신뢰하기 어렵다. 우리는 active sampling을 통해 causal effect estimation을 위한 model training을 반복적으로 향상시키는 budgeted active experimentation framework를 제안한다. observational prior를 활용하여, uplift estimation uncertainty, overlap deficit, domain discrepancy를 목표로 하는 acquisition function을 개발하고, 이를 통해 randomized experiment에 가장 정보가 풍부한 unit을 선택한다. 우리는 finite-sample deviation bound, martingale Central Limit Theorems(CLTS)를 통한 asymptotic normality, 그리고 minimax lower bound를 확립하여 information-theoretic optimality를 증명한다. 산업 데이터셋에 대한 광범위한 실험은, 우리의 접근법이 비용 제약이 있는 환경에서 표준 randomized baseline보다 현저히 우수함을 보여준다.

"},{"id":"67246","en":"Position: Mechanisms for Aggregated Individual Reporting Should be Established for Post-Deployment Evaluation","ko":"Position: 배포 후 평가를 위해 집계된 개별 보고를 위한 메커니즘이 확립되어야 한다","authors":"Jessica Dai, Inioluwa Raji, Benjamin Recht, Irene Y. Chen","pos":"#3111","link":"https://openreview.net/forum?id=15P3Uufs7I","abs":"

The need for developing model evaluations beyond static benchmarking, especially in the post-deployment phase, is now well-understood. At the same time, concerns about the concentration of power in deployed AI systems have sparked a keen interest in \"democratic\" or \"public\" AI. In this work, we bring these two ideas together by proposing mechanisms for aggregated individual reporting (AIR), a framework for post-deployment evaluation that relies on individual reports from the public. An AIR mechanism allows those who interact with a specific, deployed (AI) system to report when they feel that they may have experienced something problematic; these reports are then aggregated over time, with the goal of evaluating the relevant system in a fine-grained manner. This position paper argues that individual experiences should be understood as an integral part of post-deployment evaluation, and that the scope of our proposed aggregated individual reporting mechanism is a practical path to that end. On the one hand, individual reporting can identify substantively novel insights about safety and performance; on the other, aggregation can be uniquely useful for informing action. From a normative perspective, the post-deployment phase completes a missing piece in the conversation about \"democratic\" AI. As a pathway to implementation, we provide a workflow of concrete design decisions and pointers to areas for future research.

","absKo":"모델 evaluation을 static benchmarking을 넘어, 특히 deployment 이후 단계에서 발전시켜야 할 필요성은 이제 널리 이해되고 있다. 동시에 배포된 AI system의 권력 집중에 대한 우려는 \"democratic\" 또는 \"public\" AI에 대한 큰 관심을 불러일으켰다. 이 연구에서는 이러한 두 아이디어를 결합하여, 대중의 individual report에 의존하는 post-deployment evaluation framework인 aggregated individual reporting (AIR)을 제안한다. AIR mechanism은 특정 배포된 (AI) system과 상호작용한 사람이 문제를 경험했을 수 있다고 느낄 때 이를 report할 수 있게 하며, 이러한 report는 시간이 지나며 누적되어 관련 system을 세밀하게 평가하는 것을 목표로 한다. 이 position paper는 individual experience가 post-deployment evaluation의 필수적인 일부로 이해되어야 하며, 우리가 제안하는 aggregated individual reporting mechanism의 범위가 그 목표를 향한 실용적인 경로라고 주장한다. 한편으로 individual reporting은 safety와 performance에 대해 실질적으로 새로운 통찰을 밝혀낼 수 있고, 다른 한편으로 aggregation은 action을 inform하는 데 uniquely useful할 수 있다. 규범적 관점에서 보면, post-deployment phase는 \"democratic\" AI를 둘러싼 논의에서 빠져 있던 퍼즐의 한 조각을 완성한다. 구현 경로로서, 우리는 구체적인 design decision과 future research 영역에 대한 pointer로 구성된 workflow를 제공한다.

"},{"id":"61731","en":"Evaluating LLMs When They Do Not Know the Answer: Statistical Evaluation of Mathematical Reasoning via Comparative Signals","ko":"답을 모르는 LLM 평가: 비교 신호를 통한 수학적 추론의 통계적 평가","authors":"Zihan Dong, Zhixian Zhang, Yang Zhou, Can Jin, Ruijia Wu, Linjun Zhang","pos":"#4103","link":"https://openreview.net/forum?id=nOQOjKYwTM","abs":"

Evaluating mathematical reasoning in LLMs is constrained by limited benchmark sizes and inherent model stochasticity, yielding high-variance accuracy estimates and unstable rankings across platforms. On difficult problems, an LLM may fail to produce a correct final answer, yet still provide reliable pairwise comparison signals indicating which of two candidate solutions is better. We leverage this observation to design a statistically efficient evaluation framework that combines standard labeled outcomes with pairwise comparison signals obtained by having models judge auxiliary reasoning chains. Treating these comparison signals as control variates, we develop a semiparametric estimator based on the efficient influence function (EIF) for the setting where auxiliary reasoning chains are observed. This yields a one-step estimator that achieves the semiparametric efficiency bound, guarantees strict variance reduction over naive sample averaging, and admits asymptotic normality for principled uncertainty quantification. Across simulations, our one-step estimator substantially improves ranking accuracy, with gains increasing as model output noise grows. Experiments on GPQA Diamond, AIME 2025, and GSM8K further demonstrate more precise performance estimation and more reliable model rankings, especially in small-sample regimes where conventional evaluation is pretty unstable.

","absKo":"

LLM의 mathematical reasoning을 평가하는 일은 제한된 benchmark 규모와 모델 고유의 stochasticity 때문에 제약을 받으며, 그 결과 accuracy 추정치의 분산이 크고 플랫폼 간 ranking도 불안정합니다. 어려운 문제에서는 LLM이 올바른 최종 답을 내지 못하더라도, 두 후보 solution 중 어느 쪽이 더 나은지를 가리키는 신뢰할 수 있는 pairwise comparison signal을 제공할 수 있습니다. 우리는 이 관찰을 활용하여, 표준 labeled outcome과 모델이 auxiliary reasoning chain을 판정함으로써 얻는 pairwise comparison signal을 결합하는 통계적으로 효율적인 evaluation framework를 설계합니다. 이 comparison signal을 control variate로 취급하여, auxiliary reasoning chain이 관찰되는 설정에 대해 efficient influence function (EIF)에 기반한 semiparametric estimator를 개발합니다. 이를 통해 semiparametric efficiency bound를 달성하는 one-step estimator를 얻으며, 순진한 sample averaging에 비해 엄격한 variance reduction을 보장하고, 원리적인 uncertainty quantification을 위한 asymptotic normality를 허용합니다. 다양한 simulation에서 우리의 one-step estimator는 ranking accuracy를 크게 향상시키며, model output noise가 커질수록 개선 폭도 커집니다. GPQA Diamond, AIME 2025, GSM8K 실험은 더 정밀한 performance estimation과 더 신뢰할 수 있는 model ranking을 추가로 보여주며, 특히 conventional evaluation이 매우 불안정한 small-sample regime에서 그 효과가 두드러집니다.

"},{"id":"66424","en":"Genome-Factory: A Library for Tuning, Deploying, and Interpreting Genomic Foundation Models","ko":"Genome-Factory: 유전체 Foundation Model의 튜닝, 배포, 해석을 위한 라이브러리","authors":"Weimin Wu, Xuefeng Song, Yibo Wen, Qinjie Lin, Zhihan Zhou, Jerry Yao-Chieh Hu, Zhong Wang, Han Liu","pos":"#1011","link":"https://openreview.net/forum?id=3sx2WRBSSO","abs":"

We introduce Genome-Factory, the first integrated Python library for tuning, deploying, and interpreting genomic foundation models. Our core contribution is to simplify and unify the workflow for genomic model development: data collection, model tuning, inference, benchmarking, and interpretability. For data collection, Genome-Factory offers an automated pipeline to download genomic sequences and preprocess them. It also includes quality control like GC content normalization. For model tuning, Genome-Factory supports three approaches: full-parameter, low-rank adaptation, and adapter-based fine-tuning. It is compatible with a wide range of genomic models. For inference, Genome-Factory enables both embedding extraction and DNA sequence generation. For benchmarking, we include two existing benchmarks and provide a flexible interface for users to incorporate additional benchmarks. For interpretability, Genome-Factory introduces the first open-source biological interpreter based on a sparse auto-encoder. This module disentangles embeddings into sparse, near-monosemantic latent units and links them to genomic features by regressing on external readouts. To improve accessibility, Genome-Factory offers a zero-code command-line and a user-friendly web interface. We validate the utility of Genome-Factory across three dimensions: (i) Compatibility with diverse models and fine-tuning methods; (ii) Benchmarking downstream performance using two open-source benchmarks; (iii) Biological interpretation of learned representations with DNABERT-2. These results highlight its practical value for real-world genomic analysis.

","absKo":"

우리는 genomic foundation model의 튜닝, 배포, 해석을 위한 최초의 통합 Python 라이브러리 Genome-Factory를 소개한다. 우리의 핵심 기여는 genomic model 개발의 워크플로를 데이터 수집, model tuning, inference, benchmarking, interpretability 전반에서 단순화하고 통합하는 데 있다. 데이터 수집 측면에서 Genome-Factory는 genomic sequence를 다운로드하고 이를 전처리하는 자동화 파이프라인을 제공한다. 또한 GC content normalization과 같은 quality control도 포함한다. model tuning 측면에서 Genome-Factory는 full-parameter, low-rank adaptation, adapter-based fine-tuning의 세 가지 접근을 지원한다. 이는 다양한 genomic model과 호환된다. inference 측면에서 Genome-Factory는 embedding extraction과 DNA sequence generation을 모두 가능하게 한다. benchmarking 측면에서 우리는 두 개의 기존 benchmark를 포함하고, 사용자가 추가 benchmark를 손쉽게 통합할 수 있는 유연한 인터페이스를 제공한다. interpretability 측면에서 Genome-Factory는 sparse auto-encoder를 기반으로 한 최초의 오픈소스 biological interpreter를 도입한다. 이 모듈은 embedding을 sparse하고 거의 monosemantic한 latent unit으로 분해하고, 외부 readout에 대해 회귀함으로써 이를 genomic feature와 연결한다. 접근성을 높이기 위해 Genome-Factory는 zero-code command-line과 사용자 친화적인 web interface를 제공한다. 우리는 세 가지 차원에서 Genome-Factory의 유용성을 검증한다: (i) 다양한 model 및 fine-tuning 방법과의 호환성; (ii) 두 개의 오픈소스 benchmark를 사용한 downstream performance 평가; (iii) DNABERT-2를 이용한 학습된 representation의 biological interpretation. 이러한 결과는 실제 genomic analysis에서의 실용적 가치를 보여준다.

"},{"id":"62466","en":"PyHealth 2.0: A Comprehensive Open-Source Toolkit for Accessible and Reproducible Clinical Deep Learning","ko":"PyHealth 2.0: 접근성과 재현성이 높은 임상 Deep Learning을 위한 종합 오픈소스 Toolkit","authors":"John Wu, Yongda Fan, Zhenbang Wu, Paul Landes, Eric Schrock, Sayeed Sajjad Razin, Arjun Chatterjee, Naveen Baskaran, Joshua Steier, Andrea Fitzpatrick, Bilal Arif, Rian Atri, Jathurshan Pradeepkumar, Siddhartha Laghuvarapu, Junyi Gao, Adam Cross, Jimeng Sun","pos":"#1013","link":"https://openreview.net/forum?id=gMLVFN9hl8","abs":"

Difficulty replicating baselines, high computational costs, and required domain expertise create persistent barriers to clinical AI research. To address these challenges, we introduce PyHealth 2.0, an enhanced clinical deep learning toolkit that enables predictive modeling in as few as 7 lines of code. PyHealth 2.0 offers three key contributions: (1) a comprehensive toolkit addressing reproducibility and compatibility challenges by unifying 15+ datasets, 20+ clinical tasks, 25+ models, 5+ interpretability methods, and uncertainty quantification including conformal prediction within a single framework that supports diverse clinical data modalities—signals, imaging, and electronic health records—with translation of 5+ medical coding standards; (2) accessibility-focused design accommodating multimodal data and diverse computational resources with up to 39× faster processing and 20× lower memory usage, enabling work from 16GB laptops to production systems; and (3) an active open-source community of 400+ members lowering domain expertise barriers through extensive documentation, reproducible research contributions, and collaborations with academic health systems and industry partners, including multi-language support via RHealth. PyHealth 2.0 establishes an open-source foundation and community advancing accessible, reproducible healthcare AI.

","absKo":"

기준선 재현의 어려움, 높은 계산 비용, 그리고 요구되는 도메인 전문성은 clinical AI 연구에 지속적인 장벽을 만든다. 이러한 문제를 해결하기 위해 우리는 예측 모델링을 단 7줄의 코드로 가능하게 하는 향상된 clinical deep learning toolkit인 PyHealth 2.0을 소개한다. PyHealth 2.0은 세 가지 핵심 기여를 제공한다: (1) 신호, 영상, electronic health records를 아우르는 다양한 clinical data modality를 지원하는 단일 framework 안에서 15개 이상의 datasets, 20개 이상의 clinical tasks, 25개 이상의 models, 5개 이상의 interpretability methods, 그리고 conformal prediction을 포함한 uncertainty quantification을 통합함으로써 재현성과 호환성 문제를 해결하는 종합 toolkit으로서, 5개 이상의 medical coding standards 번역을 포함한다; (2) multimodal data와 다양한 computational resources를 수용하는 접근성 중심 설계로, 최대 39배 빠른 처리와 20배 낮은 메모리 사용을 제공하여 16GB 노트북부터 production system까지 작업을 가능하게 한다; 그리고 (3) 방대한 문서화, 재현 가능한 연구 기여, 학술 의료 시스템 및 산업 파트너와의 협력을 통해 도메인 전문성 장벽을 낮추는 400명 이상의 활성 open-source community로서, RHealth를 통한 다국어 지원을 포함한다. PyHealth 2.0은 접근 가능하고 재현 가능한 healthcare AI를 발전시키는 open-source 기반과 커뮤니티를 구축한다.

"},{"id":"66507","en":"Ekka: Automated Diagnosis of Silent Errors in LLM Inference","ko":"Ekka: LLM 추론에서 무증상 오류의 자동 진단","authors":"Yile Gu, Zhen Zhang, Shaowei Zhu, Xinwei Fu, Jun Wu, Yida Wang, Baris Kasikci","pos":"#2405","link":"https://openreview.net/forum?id=35GKHwXUmY","abs":"LLM serving frameworks are quickly evolving with a complex software stack and a vast number of optimizations. The rapid development process can introduce silent errors where output quality silently degrades without any explicit error signals. Diagnosing silent errors is notoriously difficult due to the substantial semantic gap between the high-level symptoms and the low-level root causes. We observe that diagnosis of silent errors can be effectively framed as a differential debugging problem by leveraging the existence of semantically correct reference implementations. We propose Ekka, an automated diagnosis system that identifies root causes by systematically aligning and comparing intermediate execution states between a target and a reference framework. We constructed a benchmark of real-world silent errors from popular serving frameworks, where Ekka shows 84\\% pass@$1$ diagnosis accuracy and 88\\% pass@$5$ diagnosis accuracy, outperforming state-of-the-art systems. Ekka also diagnoses 4 new silent errors from serving frameworks, all of which have been confirmed by the developers.","absKo":"LLM serving frameworks는 복잡한 software stack과 방대한 최적화들로 인해 빠르게 진화하고 있다. 급격한 개발 과정은 명시적인 error signal 없이 output quality가 조용히 저하되는 silent error를 유발할 수 있다. Silent error의 진단은 높은 수준의 증상과 낮은 수준의 root cause 사이의 큰 semantic gap 때문에 악명 높게 어렵다. 우리는 semantically correct한 reference implementation의 존재를 활용함으로써 silent error의 진단을 differential debugging 문제로 효과적으로 정식화할 수 있음을 관찰한다. 이에 따라 우리는 target framework와 reference framework 사이의 intermediate execution state를 체계적으로 정렬하고 비교함으로써 root cause를 식별하는 automated diagnosis system인 Ekka를 제안한다. 우리는 인기 있는 serving framework들에서 발생한 실제 silent error들로 benchmark를 구성했으며, Ekka는 pass@$1$ diagnosis accuracy 84\\%, pass@$5$ diagnosis accuracy 88\\%를 달성하여 state-of-the-art system들을 능가한다. 또한 Ekka는 serving framework에서 발견된 새로운 silent error 4개를 진단했으며, 이들 모두는 developer들에 의해 확인되었다."},{"id":"64109","en":"BubbleSpec: Turning Long-Tail Bubbles into Speculative Rollout Drafts for Synchronous Reinforcement Learning","ko":"BubbleSpec: 동기식 강화학습을 위해 롱테일 버블을 추측적 롤아웃 초안으로 전환","authors":"Yuhang Xu, Kaibin Tian, Yang Tian, Zhice Yang, Yifeng Yu, Yan Li, Shengzhong Liu, Fan Wu, Guihai Chen","pos":"#2614","link":"https://openreview.net/forum?id=QoJrOgCmC4","abs":"Reinforcement Learning (RL) has become a cornerstone for improving the performance of Large Language Models (LLMs). However, its rollout phase constitutes a significant efficiency bottleneck, mainly arising from the long-tail bubbles across data parallel ranks, particularly in long-context scenarios where faster GPUs remain idle while waiting for stragglers. Existing solutions, such as partial rollout or asynchronous RL, mitigate these bubbles by compromising the algorithm's strict synchronous nature. Instead, we propose **BubbleSpec**, a novel framework that accelerates RL rollouts while strictly keeping the mathematical exactness. Instead of attempting to eliminate bubbles, BubbleSpec exploits them. We exploit the idle time windows of faster ranks to pre-generate rollout results for subsequent steps, serving as drafts for speculative decoding. Unlike prior speculative methods that rely on historical epoch similarity and warm-ups, BubbleSpec is agnostic to dataset size and provides immediate acceleration from the onset of training. Extensive evaluations demonstrate that BubbleSpec reduces decoding steps by **$\\sim$50\\%** and increases rollout throughput by up to **1.8$\\times$**. Critically, BubbleSpec is seamlessly compatible with various RL frameworks and strategies as it sustains the strict synchronous property of RL algorithms.","absKo":"Reinforcement Learning(RL)은 Large Language Models(LLMs)의 성능을 향상시키는 핵심 요소가 되었다. 그러나 rollout 단계는 중요한 효율 병목을 구성하며, 이는 주로 data parallel rank 전반에 걸친 long-tail bubble에서 비롯된다. 특히 long-context 시나리오에서는 더 빠른 GPU가 느린 작업을 기다리며 유휴 상태가 되는 문제가 두드러진다. partial rollout이나 asynchronous RL과 같은 기존 해결책은 알고리즘의 엄격한 동기적 특성을 일부 희생함으로써 이러한 bubble을 완화한다. 이에 반해 우리는 수학적 정확성을 엄격히 유지하면서 RL rollout을 가속하는 새로운 프레임워크 **BubbleSpec**을 제안한다. bubble을 제거하려 하기보다, BubbleSpec은 이를 활용한다. 우리는 더 빠른 rank의 유휴 시간 구간을 이용해 후속 step의 rollout 결과를 미리 생성하고, 이를 speculative decoding의 draft로 사용한다. 과거 epoch 유사성과 warm-up에 의존하는 기존 speculative 방법과 달리, BubbleSpec은 dataset size에 무관하며 학습 시작 시점부터 즉시 가속을 제공한다. 광범위한 평가 결과 BubbleSpec은 decoding step 수를 **$\\sim$50\\%** 줄이고 rollout throughput을 최대 **1.8$\\times$**까지 높인다. 결정적으로, BubbleSpec은 RL 알고리즘의 엄격한 synchronous property를 유지하므로 다양한 RL framework와 strategy와 매끄럽게 호환된다."},{"id":"62347","en":"CONTINUUM: Restoring the Contiguous Tensor Abstraction Efficiently for Dynamic AI Workloads via Hardware Virtualization","ko":"CONTINUUM: 하드웨어 가상화를 통해 동적 AI 워크로드에서 Contiguous Tensor Abstraction을 효율 복원","authors":"Yangyu Zhang, shuoming zhang, Chunwei Xia, Shuaijiang Li, Zhicheng Li, Ruiyuan Xu, Zheming Yang, Lei Chen, YUAN WEN, Guangli Li, Xiaobing Feng, Huimin Cui, Jiacheng Zhao","pos":"#313","link":"https://openreview.net/forum?id=hROxrfMoXj","abs":"

Emerging LLM workloads demand extreme mem- ory agility. However, state-of-the-art inference systems (e.g., vLLM) rely on software-defined paging, which sacrifices the contiguous tensor abstraction. This rigid interface exposes fragmen- tation complexity to developers, imposing a se- vere engineering burden that stifles algorithmic innovation. We introduce CONTINUUM, a tensor memory virtualization subsystem implemented as a PyTorch extension. By bypassing serialized OS bottlenecks via a lightweight GPU driver ex- tension, CONTINUUM can significantly reduce the mapping costs by orders of magnitude—from milliseconds to microseconds. Built atop this low-latency API, CONTINUUM provides Elastic Tensor, with a set of flexible tensor operations that natively supports complex memory dynamics and zero-copy topological aliasing. Evaluations demonstrate that CONTINUUM achieves signifi- cantly higher throughput across diverse dynamic scenarios, effectively democratizing the imple- mentation of next-generation LLM applications.

","absKo":"

새롭게 등장하는 LLM workload는 극단적인 memory agility를 요구한다. 그러나 최신 inference system(예: vLLM)은 contiguous tensor abstraction을 희생하는 software-defined paging에 의존한다. 이러한 경직된 인터페이스는 fragmentation의 복잡성을 개발자에게 노출시켜, 알고리즘 혁신을 저해하는 심각한 engineering burden을 초래한다. 우리는 PyTorch extension으로 구현된 tensor memory virtualization subsystem인 CONTINUUM을 소개한다. 가벼운 GPU driver extension을 통해 직렬화된 OS bottleneck을 우회함으로써, CONTINUUM은 mapping cost를 밀리초에서 마이크로초로, 즉 수십만 배 규모로 크게 줄일 수 있다. 이 저지연 API 위에 구축된 CONTINUUM은 Elastic Tensor를 제공하며, 복잡한 memory dynamics와 zero-copy topological aliasing을 네이티브하게 지원하는 유연한 tensor operation 집합을 갖춘다. 평가 결과, CONTINUUM은 다양한 dynamic scenario에서 훨씬 더 높은 throughput을 달성하여, 차세대 LLM application 구현을 사실상 민주화한다.

"},{"id":"62545","en":"cuRegOT: A GPU-Accelerated Solver for Entropic-Regularized Optimal Transport","ko":"cuRegOT: Entropic-Regularized Optimal Transport를 위한 GPU 가속 Solver","authors":"Yixuan Qiu","pos":"#314","link":"https://openreview.net/forum?id=fgIg1oOMwk","abs":"

Optimal transport (OT) has emerged as a fundamental tool in modern machine learning, yet its computational cost remains a significant bottleneck for large-scale applications. While harnessing the massive parallelism of modern GPU hardware is critical for efficiency, the de facto standard Sinkhorn algorithm, despite its ease of parallelization, often suffers from slow convergence in challenging problems. More recently, the sparse-plus-low-rank quasi-Newton method offers a balance between convergence rate and per-iteration complexity; however, its efficiency on GPUs is severely hindered by the serial nature of sparse matrix symbolic analysis and irregular memory access patterns. To bridge this gap, we present cuRegOT, a high-performance GPU solver tailored for entropic-regularized OT. We introduce a suite of algorithmic and architectural optimizations, including an amortized symbolic analysis strategy to mitigate CPU bottlenecks, an asynchronous Sinkhorn iterates generation mechanism, and a fused kernel for bandwidth-efficient gradient evaluation. These strategies are backed by rigorous theoretical guarantees ensuring algorithmic convergence. Extensive numerical experiments demonstrate that cuRegOT achieves significant speedups over state-of-the-art GPU-based solvers across a variety of benchmark tasks.

","absKo":"

Optimal transport(OT)는 현대 machine learning의 근본적인 도구로 부상했지만, 그 계산 비용은 대규모 응용에서 여전히 중대한 병목이다. 효율성을 위해 현대 GPU 하드웨어의 막대한 병렬성을 활용하는 것이 중요하지만, 사실상 표준인 Sinkhorn algorithm은 병렬화가 용이함에도 불구하고 어려운 문제에서는 종종 느린 수렴에 시달린다. 최근에는 sparse-plus-low-rank quasi-Newton method가 수렴 속도와 반복당 복잡도 사이의 균형을 제공하지만, sparse matrix의 symbolic analysis가 순차적이라는 특성과 불규칙한 메모리 접근 패턴 때문에 GPU에서의 효율성이 심각하게 저하된다. 이러한 격차를 해소하기 위해 우리는 entropic-regularized OT에 맞춘 고성능 GPU solver인 cuRegOT를 제시한다. 우리는 CPU 병목을 완화하기 위한 amortized symbolic analysis 전략, asynchronous Sinkhorn iterates generation mechanism, bandwidth-efficient gradient evaluation을 위한 fused kernel 등을 포함한 알고리즘 및 아키텍처 최적화를 도입한다. 이러한 전략들은 알고리즘 수렴을 보장하는 엄밀한 이론적 보장으로 뒷받침된다. 광범위한 수치 실험은 cuRegOT가 다양한 benchmark task 전반에서 최신 GPU 기반 solver들보다 유의미한 속도 향상을 달성함을 보여준다.

"},{"id":"66760","en":"Solving Time-Dependent Differential Equations with Physical Dynamical Systems","ko":"물리 동역학 시스템으로 시간 의존 미분방정식 풀기","authors":"Chuan Liu, Yijie Chen, Ruibing Song, Wenhao Huang, Chunshu Wu, Deqian Kong, Ying Nian Wu, Kaiyuan Yang, Ang Li, Tony Geng","pos":"#3903","link":"https://openreview.net/forum?id=0YHSZPkMp8","abs":"Time-Dependent Differential Equations (TDDEs) model dynamical processes across science and engineering, but time-critical applications require solvers delivering high-fidelity trajectories under stringent latency constraints. Most existing TDDE solvers are limited by time discretization, forcing a latency-accuracy trade-off where smaller step sizes capture high-fidelity trajectories but incur prohibitive runtime, while larger steps meet real-time budgets at the cost of trajectory distortion. Dynamical System Machines (DSMs) offer a promising alternative by computing through continuous-time physical evolution, yet existing DSMs struggle to capture the spatiotemporal complexity of TDDEs. This work introduces DS-TS, a novel TDDE solver that achieves both high-accuracy and ultra-efficiency, leveraging the continuous-time computation of DSMs. DS-TS integrates three key innovations: (1) Excitatory-Inhibitory Inspired Coupling to better model complex spatial interactions; (2) State-aware Dynamic Non-linearity to enable rich inter-node interactions and state-dependent spatiotemporal correlations; and (3) Hierarchical Temporal Integration to capture long-range temporal dependencies. Experiments demonstrate that DS-TS achieves high-fidelity solutions while delivering orders-of-magnitude improvements in speed ($\\sim 10^3\\times$) and energy efficiency ($\\sim 10^5\\times$) compared to baseline solvers.","absKo":"Time-Dependent Differential Equations (TDDEs)는 과학과 공학 전반의 dynamical process를 모델링하지만, time-critical application에서는 엄격한 latency constraint 하에서 high-fidelity trajectory를 제공하는 solver가 필요하다. 기존 TDDE solver의 대부분은 time discretization에 의해 제한되며, 더 작은 step size는 high-fidelity trajectory를 포착하지만 실행 시간이 지나치게 늘어나고, 더 큰 step은 real-time budget을 맞추는 대신 trajectory distortion을 초래하는 latency-accuracy trade-off에 갇힌다. Dynamical System Machine (DSM)은 continuous-time physical evolution을 통해 계산함으로써 유망한 대안을 제공하지만, 기존 DSM은 TDDE의 spatiotemporal complexity를 포착하는 데 어려움을 겪는다. 본 연구는 DSM의 continuous-time computation을 활용하여 high-accuracy와 ultra-efficiency를 동시에 달성하는 새로운 TDDE solver인 DS-TS를 제안한다. DS-TS는 세 가지 핵심 혁신을 통합한다: (1) 복잡한 spatial interaction을 더 잘 모델링하기 위한 Excitatory-Inhibitory Inspired Coupling; (2) 풍부한 inter-node interaction과 state-dependent spatiotemporal correlation을 가능하게 하는 State-aware Dynamic Non-linearity; (3) 장거리 temporal dependency를 포착하는 Hierarchical Temporal Integration. 실험 결과 DS-TS는 high-fidelity solution을 제공하면서, baseline solver에 비해 속도(약 $10^3\\times$)와 energy efficiency(약 $10^5\\times$)에서 자릿수 수준의 향상을 달성함을 보인다."},{"id":"62917","en":"DuetServe: Harmonizing Prefill and Decode for LLM Serving via Adaptive GPU Multiplexing","ko":"DuetServe: Adaptive GPU Multiplexing으로 LLM Serving에서 Prefill과 Decode 조화","authors":"Lei Gao, Chaoyi Jiang, Hossein Entezari Zarch, Daniel Wong, Mark D. HIll, Murali Annavaram","pos":"#4000","link":"https://openreview.net/forum?id=bu2Pcce6Rq","abs":"

Modern LLM serving systems must sustain high throughput while meeting strict latency SLOs across two distinct inference phases: compute-intensive prefill and memory-bound decode phases. Existing approaches either (1) aggregate both phases on shared GPUs, leading to interference between prefill and decode phases, which degrades Time-Between-Tokens (TBT); or (2) disaggregate the two phases across GPUs, improving latency but wasting resources through duplicated models and KV cache transfers. We present DuetServe, a unified LLM serving framework that achieves disaggregation-level isolation within a single GPU. DuetServe operates in aggregated mode by default and dynamically activates SM-level GPU spatial multiplexing when TBT degradation is predicted. Its key idea is to decouple prefill and decode execution only when needed through fine-grained, adaptive SM partitioning that provides phase isolation only when contention threatens latency service level objectives. DuetServe integrates (1) an attention-aware roofline model to forecast iteration latency, (2) a partitioning optimizer that selects the optimal SM split to maximize throughput under TBT constraints, and (3) an interruption-free execution engine that eliminates CPU–GPU synchronization overhead. Evaluations show that DuetServe improves total throughput by up to 1.3× while maintaining low generation latency compared to state-of-the-art frameworks.

","absKo":"현대의 LLM serving system은 compute-intensive prefill과 memory-bound decode라는 두 개의 상이한 inference phase 전반에서 엄격한 latency SLO를 만족시키면서 높은 throughput을 유지해야 한다. 기존 접근은 (1) 두 phase를 shared GPU 위에 함께 집계하여 prefill과 decode phase 간 간섭을 유발하고, 그 결과 Time-Between-Tokens (TBT)를 악화시키거나, (2) 두 phase를 GPU들 사이에 disaggregate하여 latency를 개선하지만 중복 모델과 KV cache 전송으로 자원을 낭비한다. 우리는 단일 GPU 내에서 disaggregation 수준의 isolation을 달성하는 통합 LLM serving framework인 DuetServe를 제시한다. DuetServe는 기본적으로 aggregated mode로 동작하며, TBT 저하가 예측될 때 동적으로 SM-level GPU spatial multiplexing을 활성화한다. 핵심 아이디어는 세밀하고 적응적인 SM partitioning을 통해 contention이 latency service-level objective를 위협할 때에만 prefill과 decode 실행을 분리하여 phase isolation을 제공하는 것이다. DuetServe는 (1) iteration latency를 예측하는 attention-aware roofline model, (2) TBT 제약 하에서 throughput을 최대화하는 최적 SM split을 선택하는 partitioning optimizer, (3) CPU-GPU synchronization overhead를 제거하는 interruption-free execution engine을 통합한다. 평가는 DuetServe가 최신 state-of-the-art framework와 비교해 낮은 generation latency를 유지하면서 총 throughput을 최대 1.3배까지 향상함을 보여준다.

"},{"id":"66409","en":"FOCUS: DLLMs Know How to Tame Their Compute Bound","ko":"FOCUS: DLLM은 계산 병목을 길들이는 법을 안다","authors":"Kaihua Liang, Xin Tan, An Zhong, Hong Xu, Marco Canini","pos":"#4001","link":"https://openreview.net/forum?id=40fUEdwvH3","abs":"

Diffusion Large Language Models (DLLMs) offer a compelling alternative to Auto-Regressive models, but their deployment is constrained by high decoding cost. In this work, we identify a key inefficiency in DLLM decoding: while computation is parallelized over token blocks, only a small subset of tokens is decodable at each diffusion step, causing most compute to be wasted on non-decodable tokens. We further observe a strong correlation between attention-derived token importance and token-wise decoding probability. Based on this insight, we propose FOCUS—an inference system designed for DLLMs. By dynamically focusing computation on decodable tokens and evicting non-decodable ones on-the-fly, FOCUS increases the effective batch size, alleviating compute limitations and enabling scalable throughput. Empirical evaluations demonstrate that FOCUS achieves up to 3.52× throughput improvement over the production-grade engine LMDeploy, while preserving or improving generation quality across multiple benchmarks.

","absKo":"

Diffusion Large Language Models (DLLMs)은 Auto-Regressive model의 매력적인 대안이지만, 높은 decoding 비용 때문에 배포가 제약됩니다. 이 연구에서는 DLLM decoding의 핵심 비효율을 확인합니다. computation은 token block 단위로 병렬화되지만, 각 diffusion step에서 decode 가능한 token은 소수에 불과하여 대부분의 compute가 decode 불가능한 token에 낭비됩니다. 또한 attention에서 도출한 token 중요도와 token별 decoding probability 사이에 강한 상관관계가 있음을 관찰합니다. 이 통찰을 바탕으로, 우리는 FOCUS를 제안합니다. FOCUS는 DLLM을 위한 inference system입니다. decode 가능한 token에 computation을 동적으로 focus하고, decode 불가능한 token은 실행 중에 제거함으로써 effective batch size를 키우고, compute 제약을 완화하며, 확장 가능한 throughput을 가능하게 합니다. 실험 결과 FOCUS는 production-grade engine인 LMDeploy 대비 최대 3.52× throughput 향상을 달성하면서도 여러 benchmark에서 generation quality를 유지하거나 개선하는 것으로 나타났습니다.

"},{"id":"62548","en":"From Volume to Value: Preference-Aligned Memory Construction for On-Device RAG","ko":"Volume에서 Value로: 온디바이스 RAG를 위한 선호 정렬 Memory Construction","authors":"Changmin Lee, Jaemin Kim, Taesik Gong","pos":"#4002","link":"https://openreview.net/forum?id=fdrkyBHleI","abs":"With the rapid emergence of personal AI agents based on Large Language Models (LLMs), implementing them on-device has become essential for privacy and responsiveness. To handle the inherently personal and context-dependent nature of real-world requests, such agents must ground their generation in device-resident personal context. However, under tight memory budgets, the core bottleneck is *what to store* so that retrieval remains aligned with the user. We propose EPIC (Efficient Preference-aligned Index Construction), which focuses on user preferences as a compact and stable form of personal context and integrates them throughout the RAG pipeline. EPIC selectively retains preference-relevant information from raw data and aligns retrieval toward preference-aligned contexts. Across four benchmarks covering conversations, debates, explanations, and recommendations, EPIC reduces indexing memory by 2,404$\\times$, improves preference-following accuracy by 20.17\\%p, and achieves 33.33$\\times$ lower retrieval latency over the best-performing baseline. In our on-device experiment, EPIC maintains a memory footprint under 1 MB with 27.9 ms/query retrieval latency in streaming updates. The code is available at","absKo":"대규모 언어 모델(LLMs) 기반 personal AI agents가 빠르게 등장함에 따라, privacy와 responsiveness를 위해 이를 on-device로 구현하는 것이 필수적이 되었다. 실제 요청의 본질인 개인적이고 context-dependent한 성격을 다루기 위해, 이러한 agent는 자신의 생성을 device-resident personal context에 기반해야 한다. 그러나 memory budget이 매우 제한된 상황에서 핵심 병목은 사용자와의 정렬을 유지하면서 retrieval이 가능하도록 *무엇을 저장할 것인가*이다. 우리는 개인적 context의 compact하고 stable한 형태로 user preferences에 주목하고 이를 RAG pipeline 전반에 통합하는 EPIC(Efficient Preference-aligned Index Construction)를 제안한다. EPIC는 raw data에서 preference와 관련된 정보를 선택적으로 보존하고, retrieval을 preference-aligned context 쪽으로 정렬한다. 대화, 토론, 설명, 추천을 포괄하는 네 가지 benchmark에서 EPIC는 indexing memory를 2,404$\\times$ 줄이고, preference-following accuracy를 20.17\\%p 향상시키며, 최고 성능 baseline 대비 retrieval latency를 33.33$\\times$ 낮춘다. on-device 실험에서 EPIC는 streaming update 시 27.9 ms/query retrieval latency와 함께 1 MB 미만의 memory footprint를 유지한다. 코드는 다음에서 사용할 수 있다"},{"id":"64036","en":"Not All Prefills Are Equal: PPD Disaggregation for Multi-turn LLM Serving","ko":"모든 Prefill이 동등하지는 않다: 멀티턴 LLM 서빙을 위한 PPD 분리","authors":"Zongze Li, Jingyu Liu, Zach Xu, Yineng Zhang, Tahseen Rabbani, Ce Zhang","pos":"#4003","link":"https://openreview.net/forum?id=RW23qIUb5f","abs":"

Prefill-Decode (PD) disaggregation has become the standard architecture for modern LLM inference engines, which alleviates the interference of two distinctive workloads. With the growing demand for multi-turn interactions in chatbots and agentic systems, we re-examined PD in this case and found two fundamental inefficiencies: (1) every turn requires prefilling the new prompt and response from the last turn, and (2) repeated KV transfers between prefill and decode nodes saturate the bandwidth, leading to high latency and even service degradation. Our key insight is that not all prefill operations are equally disruptive: append-prefill---processing only the new input tokens while reusing cached KV states---incurs substantially less decoding slowdown than full prefill. This motivates routing append-prefill to decode nodes locally. However, through comprehensive analysis, we show that no single fixed routing strategy satisfies all Service Level Objectives (SLOs) simultaneously. Based on this insight, we propose Prefill Prefill-capable Decode (PPD) disaggregation, a dynamic routing system that decides when to process Turn 2+ requests locally on decode nodes using cached KV states. PPD adapts to varying SLOs via configurable weights and seamlessly integrates with traditional PD deployments. With extensive evaluations, we show that PPD reduces Turn 2+ time-to-first-token (TTFT) by 68% while maintaining competitive time-per-output-token (TPOT), effectively alleviating KV transfer congestion under high load. We believe PPD represents a flexible and efficient paradigm for multi-turn LLM serving.

","absKo":"

Prefill-Decode (PD) disaggregation은 현대 LLM inference engine의 표준 아키텍처가 되었으며, 서로 다른 두 workload의 간섭을 완화한다. chatbot과 agentic system에서 multi-turn interaction에 대한 수요가 증가함에 따라, 우리는 이 경우의 PD를 재검토했고 두 가지 근본적인 비효율을 발견했다: (1) 매 turn마다 새 prompt와 이전 turn의 response를 다시 prefill해야 하며, (2) prefill node와 decode node 사이의 반복적인 KV transfer가 bandwidth를 포화시켜 높은 latency와 심지어 service degradation까지 유발한다. 우리의 핵심 통찰은 모든 prefill operation이 동일하게 disruptive한 것은 아니라는 점이다. append-prefill---즉 cached KV state를 재사용하면서 새로운 input token만 처리하는 방식---은 full prefill에 비해 decoding slowdown을 훨씬 덜 유발한다. 이는 append-prefill을 decode node로 local하게 routing해야 한다는 동기를 제공한다. 그러나 포괄적인 분석을 통해, 어떤 단일 고정 routing strategy도 모든 Service Level Objective (SLO)를 동시에 만족시킬 수는 없음을 보인다. 이러한 통찰에 기반하여, 우리는 Prefill Prefill-capable Decode (PPD) disaggregation을 제안한다. 이는 cached KV state를 사용해 Turn 2+ request를 decode node에서 local하게 처리할 시점을 결정하는 dynamic routing system이다. PPD는 configurable weight를 통해 다양한 SLO에 적응하며, 기존 PD deployment와도 매끄럽게 통합된다. 광범위한 평가를 통해, PPD가 경쟁력 있는 time-per-output-token (TPOT)을 유지하면서 Turn 2+의 time-to-first-token (TTFT)을 68% 줄이고, high load에서 KV transfer congestion을 효과적으로 완화함을 보인다. 우리는 PPD가 multi-turn LLM serving을 위한 유연하고 효율적인 paradigm이라고 믿는다.

"},{"id":"64145","en":"Two-Stage Unit Tying for Simplifying Differentiable Logic Gate Networks","ko":"미분 가능 논리 게이트 네트워크 단순화를 위한 2단계 유닛 결합","authors":"Seungheon Lee, Jeongmin Sun, Jaeyong Chung","pos":"#4004","link":"https://openreview.net/forum?id=QVt9dgerpS","abs":"

Differentiable logic gate networks map learned models directly to gate-level circuits, enabling ultra-low-latency inference, yet their logic footprint often exceeds FPGA capacity budgets. Tightly fitting a trained model to a target FPGA requires a post-training mechanism to trade off network complexity and accuracy—analogous to pruning in standard neural networks. To this end, we introduce unit tying: a simplification that forces selected gates to constants (0 or 1), enabling constant propagation and downstream logic elimination. However, we observe that naively extending pruning criteria to logic networks is unreliable under such near-discrete modifications. We therefore propose a two-stage algorithm for unit tying: (i) a fast Gauss–Newton screening step under a teacher-referenced logit-distortion objective that constructs a high-recall overshoot set and (ii) a refinement step that corrects approximation and interaction-driven errors using a small number of finite-difference evaluations. On CIFAR-10 and MNIST, our method consistently improves the accuracy–area trade-off over common saliency baselines, yielding substantial post-synthesis LUT reductions of up to 48% on CIFAR-10 and 43% on MNIST, with modest accuracy degradation.

","absKo":"

Differentiable logic gate network는 학습된 model을 gate-level circuit으로 직접 매핑하여 초저지연 inference를 가능하게 하지만, 그 logic footprint가 종종 FPGA capacity budget을 초과한다. 학습된 model을 target FPGA에 촘촘하게 맞추려면, 표준 neural network의 pruning과 유사하게 network complexity와 accuracy를 절충하는 post-training mechanism이 필요하다. 이를 위해 우리는 unit tying을 도입한다. 이는 선택된 gate를 상수(0 또는 1)로 고정하여 constant propagation과 downstream logic elimination을 가능하게 하는 단순화 기법이다. 그러나 pruning criterion을 logic network에 단순 확장하는 방식은 이러한 준이산적 수정 하에서 신뢰할 수 없음을 관찰했다. 따라서 우리는 unit tying을 위한 두 단계 알고리즘을 제안한다. (i) teacher-referenced logit-distortion objective 하에서 빠른 Gauss–Newton screening 단계로 high-recall overshoot set을 구성하고, (ii) 소수의 finite-difference evaluation을 사용해 approximation error와 interaction-driven error를 보정하는 refinement 단계이다. CIFAR-10과 MNIST에서 우리의 방법은 일반적인 saliency baseline보다 accuracy-area trade-off를 일관되게 개선하며, CIFAR-10에서는 최대 48%, MNIST에서는 43%에 달하는 상당한 post-synthesis LUT reduction을 적은 정확도 저하와 함께 달성한다.

"},{"id":"63992","en":"SPARe: Stacked Parallelism with Adaptive Reordering for Fault-Tolerant LLM Pretraining Systems with 100k+ GPUs","ko":"SPARe: 100k+ GPUs를 사용하는 오류 허용 LLM 사전학습 시스템을 위한 적응적 재배치 기반 스택 병렬화","authors":"Jin Lee, Zhonghao Chen, Xuhang He, Robert Underwood, Bogdan Nicolae, Franck Cappello, Xiaoyi Lu, Sheng Di, Zheng Zhang","pos":"#4101","link":"https://openreview.net/forum?id=RyaQldE2Gh","abs":"In large-scale LLM pretraining systems with $100\\mathrm{k}+$ GPUs, failures become the norm rather than the exception, and restart costs can dominate wall-clock training time. However, existing fault-tolerance mechanisms are largely unprepared for this restart-dominant regime. To address this challenge, we propose SPARe—Stacked Parallelism with Adaptive Reordering—a fault-tolerance framework that masks node failures during gradient synchronization by stacking redundant data shards across parallelism groups and adaptively reordering execution. SPARe achieves availability comparable to traditional replication while maintaining near-constant computation overhead of only $2\\sim3\\times$, even under high redundancy where traidional replication would require linearly inflating overhead. We derive closed-form expressions for endurable failure count and computation overhead, validate them via SimGrid-based discrete-event simulation, and jointly optimize redundancy and checkpointing to minimize training time. At extreme scale with up to $600\\mathrm{k}$ GPUs, SPARe reduces time-to-train by $40\\sim50$% compared to traditional replication.","absKo":"$100\\mathrm{k}+$ GPUs를 사용하는 대규모 LLM pretraining 시스템에서는 실패가 예외가 아니라 규범이 되며, 재시작 비용이 wall-clock 학습 시간을 지배할 수 있다. 그러나 기존의 fault-tolerance 메커니즘은 이러한 재시작 지배적(regime) 환경에 충분히 대비되어 있지 않다. 이 문제를 해결하기 위해 우리는 SPARe—Stacked Parallelism with Adaptive Reordering—를 제안한다. 이는 병렬성 그룹 전반에 걸쳐 redundant data shard를 적층하고 실행 순서를 적응적으로 재배열함으로써 gradient synchronization 중의 노드 장애를 가리는 fault-tolerance 프레임워크이다. SPARe는 전통적인 replication과 유사한 가용성을 달성하면서도, 기존 replication이 선형적으로 오버헤드를 늘려야 하는 높은 redundancy 상황에서도 계산 오버헤드를 단 $2\\sim3\\times$ 수준으로 거의 일정하게 유지한다. 우리는 견딜 수 있는 실패 개수와 계산 오버헤드에 대한 closed-form 표현을 유도하고, SimGrid 기반 discrete-event simulation으로 이를 검증했으며, 학습 시간을 최소화하도록 redundancy와 checkpointing을 공동 최적화한다. 최대 $600\\mathrm{k}$ GPUs에 이르는 극단적 규모에서 SPARe는 전통적 replication 대비 time-to-train을 $40\\sim50$% 줄인다."},{"id":"63613","en":"Prefix cache aware data reordering for LLM augmented database analytics","ko":"LLM 증강 데이터베이스 분석을 위한 prefix cache 인지형 데이터 재정렬","authors":"Yingze Li, dong wang, Yiming Guo, Yao Chen, Hongzhi Wang, Bingsheng He","pos":"#4106","link":"https://openreview.net/forum?id=VSY1nFjumI","abs":"LLM-augmented database analytics face a major bottleneck in the costly prefill phase. Although relational tables inherently contain repeated attribute values, standard row-by-row processing produces fragmented prompt layouts that obscure shared prefixes, thereby minimizing opportunities for prefix KV cache reuse and constraining system efficiency. Existing solutions typically employ heuristic or exhaustive search methods to reorder prompt layouts, but these approaches can be inefficient and may not leverage the structural properties of relational tables. We address this challenge by formulating prefix-cache-aware prompt layout optimization as a problem rooted in the isomorphism between prefix-cache reuse and the radix tree topology induced by the relational data distribution. Building on this perspective, we introduce a practical greedy tree-shaping algorithm that efficiently selects row and column orderings to maximize prefix overlap. Our approach, SOLO, improves prefill throughput by up to 90.3\\% under fixed prefix-cache budget. Moreover, it reduces planning overhead by up to 242$\\times$ compared to state-of-the-art baselines.","absKo":"

LLM을 활용한 database analytics는 비용이 큰 prefill 단계에서 주요 병목을 겪는다. 관계형 테이블은 본질적으로 반복되는 attribute 값을 포함하지만, 표준적인 row-by-row 처리 방식은 공유 prefix를 가리는 단편화된 prompt layout을 만들어 prefix KV cache 재사용 기회를 최소화하고 시스템 효율을 제한한다. 기존 해결책은 일반적으로 prompt layout을 재배열하기 위해 heuristic 또는 exhaustive search method를 사용하지만, 이러한 접근은 비효율적일 수 있으며 관계형 테이블의 구조적 성질을 충분히 활용하지 못한다. 우리는 prefix-cache reuse와 관계형 데이터 분포가 유도하는 radix tree topology 사이의 isomorphism에 뿌리를 둔 문제로 prefix-cache-aware prompt layout optimization을 정식화함으로써 이 과제를 해결한다. 이러한 관점에 기반하여, 우리는 prefix overlap을 최대화하도록 row와 column ordering을 효율적으로 선택하는 실용적인 greedy tree-shaping algorithm을 도입한다. 우리의 접근법인 SOLO는 고정된 prefix-cache budget 하에서 prefill throughput을 최대 90.3\\%까지 향상시킨다. 또한 state-of-the-art baseline과 비교해 planning overhead를 최대 242$\\times$까지 줄인다.

"},{"id":"63365","en":"HelioX: A GPU-Native Framework for Simulation and Training of Biophysically Detailed Networks","ko":"HelioX: 생물물리학적으로 상세한 네트워크의 시뮬레이션과 학습을 위한 GPU 네이티브 프레임워크","authors":"Junfeng Lu, Zijie Yu, SHAOYANG CUI, Gan He, Ruiqin Xiong, Kai Du, Tiejun Huang","pos":"#915","link":"https://openreview.net/forum?id=XjTRtfxK9g","abs":"

Biophysically detailed neural networks represent a promising frontier for brain-inspired AI, offering intrinsic spatio-temporal dynamics to enhance the expressivity and computational density of deep learning systems. However, general-purpose deep learning frameworks suffer from a fundamental mismatch between their dense parallel optimizations and the irregular, tree-structured complexity of biological mechanisms. In this work, we propose HelioX, a GPU-native framework designed to unify high-performance simulation with scalable training. Unlike approaches that adapt biology to existing deep learning tools, HelioX adopts a \"GPU-to-Biophysics\" paradigm. We tailor the underlying GPU parallelism to biological structures by implementing custom-fused CUDA kernels for both the Dendritic Hierarchical Scheduling (DHS) algorithm and its gradient propagation. This design eliminates the runtime overhead of generic automatic differentiation and enables multi-stream concurrency for spike generation and equation assembly. Experimental results demonstrate that HelioX outperforms standard simulators (NEURON) by orders of magnitude and surpasses prior GPU-based solvers in both speed and scalability. We successfully train deep biophysical MLPs and organism-scale biophysical neural networks (e.g., the BAAIWorm C. elegans model) on a single consumer-grade GPU. HelioX establishes a new standard for computational efficiency, enabling the training of biophysically detailed models at scales previously unattainable.

","absKo":"

Biophysically detailed neural network는 brain-inspired AI를 위한 유망한 최전선으로, deep learning system의 expressivity와 computational density를 높이기 위한 본질적인 spatio-temporal dynamics를 제공한다. 그러나 범용 deep learning framework는 밀집된 병렬 최적화와 biological mechanism의 불규칙하고 tree-structured한 복잡성 사이의 근본적인 불일치로 어려움을 겪는다. 본 연구에서는 고성능 시뮬레이션과 확장 가능한 학습을 통합하도록 설계된 GPU-native framework인 HelioX를 제안한다. 기존 deep learning tool에 biology를 맞추는 접근과 달리, HelioX는 \"GPU-to-Biophysics\" paradigm을 채택한다. 우리는 Dendritic Hierarchical Scheduling(DHS) algorithm과 그 gradient propagation 모두에 대해 custom-fused CUDA kernel을 구현하여, 기반이 되는 GPU parallelism을 biological structure에 맞게 조정했다. 이 설계는 generic automatic differentiation의 runtime overhead를 제거하고 spike generation과 equation assembly를 위한 multi-stream concurrency를 가능하게 한다. 실험 결과, HelioX는 표준 simulator(NEURON)를 orders of magnitude 수준으로 능가하며, 이전 GPU 기반 solver보다 속도와 확장성 모두에서 우수함을 보였다. 우리는 single consumer-grade GPU에서 deep biophysical MLP와 organism-scale biophysical neural network(예: BAAIWorm C. elegans model)를 성공적으로 학습했다. HelioX는 computational efficiency의 새로운 기준을 확립하며, 이전에는 달성 불가능했던 규모의 biophysically detailed model 학습을 가능하게 한다.

"},{"id":"62731","en":"Graph is a Natural Regularization: Revisiting Vector Quantization for Graph Representation Learning","ko":"그래프는 자연스러운 Regularization: Graph Representation Learning의 Vector Quantization 재조명","authors":"Zian Zhai, Fan Li, Xingyu Tan, Xiaoyang Wang, Wenjie Zhang","pos":"#4012","link":"https://openreview.net/forum?id=dsPE185aHi","abs":"

Vector Quantization (VQ) has recently emerged as a promising approach for learning discrete representations of graph-structured data. However, a fundamental challenge, i.e., codebook collapse, remains underexplored in the graph domain, significantly limiting the expressiveness and generalization of graph tokens. In this paper, we present the first empirical study and observe that codebook collapse consistently occurs when applying VQ to graph data, even with mitigation strategies proposed in vision or language domains. Moreover, we provide a diagnosis of collapse from data and optimization perspectives, showing that collapse is associated with graph data properties such as feature redundancy and connectivity density, and is further reinforced by the training dynamics of deterministic hard assignment. To address these issues, we propose RGVQ, a novel framework that integrates graph topology and feature similarity as explicit regularization signals to enhance codebook utilization and promote token diversity. RGVQ introduces soft assignments via Gumbel-Softmax reparameterization, ensuring that all codewords receive gradient updates. In addition, RGVQ incorporates a structure-aware contrastive regularization to penalize assigning the same token to dissimilar node pairs. Extensive experiments demonstrate that RGVQ substantially improves codebook utilization and consistently boosts the performance of state-of-the-art graph VQ backbones across multiple downstream tasks, enabling more expressive and transferable graph token representations.

","absKo":"

Vector Quantization (VQ)은 최근 graph-structured data의 discrete representation을 학습하는 유망한 접근법으로 부상했다. 그러나 codebook collapse라는 근본적인 문제는 graph 영역에서 충분히 탐구되지 않았으며, 이는 graph token의 expressiveness와 generalization을 크게 제한한다. 본 논문에서는 첫 번째 실증 연구를 제시하고, vision이나 language domain에서 제안된 완화 전략을 적용하더라도 VQ를 graph data에 적용할 때 codebook collapse가 일관되게 발생함을 관찰한다. 더 나아가, 우리는 data 및 optimization 관점에서 collapse를 진단하여, collapse가 feature redundancy와 connectivity density 같은 graph data 특성과 연관되어 있으며 deterministic hard assignment의 training dynamics에 의해 추가로 강화됨을 보인다. 이러한 문제를 해결하기 위해 우리는 graph topology와 feature similarity를 명시적인 regularization signal로 통합하여 codebook utilization을 높이고 token diversity를 촉진하는 새로운 framework인 RGVQ를 제안한다. RGVQ는 Gumbel-Softmax reparameterization을 통한 soft assignment를 도입하여 모든 codeword가 gradient update를 받도록 보장한다. 또한 RGVQ는 구조를 인식하는 contrastive regularization을 포함하여, 서로 다른 node pair에 동일한 token을 할당하는 것을 penalize한다. 광범위한 실험은 RGVQ가 codebook utilization을 크게 개선하고, 여러 downstream task에서 state-of-the-art graph VQ backbone의 성능을 일관되게 향상시켜, 더 표현력 있고 transferable한 graph token representation을 가능하게 함을 보여준다.

"},{"id":"60692","en":"ElicitR: Unlocking Latent Reasoning in Dense Retrievers via Generative Regularization","ko":"ElicitR: Generative Regularization을 통한 Dense Retriever의 잠재 Reasoning 해제","authors":"Fengyu Cai, Iryna Gurevych, Heinz Koeppl","pos":"#206","link":"https://openreview.net/forum?id=xWHzF6WRGE","abs":"

Reasoning-intensive retrieval is increasingly important for downstream applications, requiring more than lexical overlap or coarse semantic matching. While prior work mainly relies on Language Models (LMs) to synthesize reasoning-oriented supervision, we posit that it is already latent in LM-based retrievers but suppressed by contrastive overfitting. To elicit this latent reasoning, we introduce ElicitR, a retriever–LM framework with generative regularization that captures nuanced relationships among a query and its candidate documents beyond binary relevance. Concretely, alongside contrastive learning, we regularize the retriever by co-training a small LM on query–positive–negative batches. Next token prediction (NTP) for each text is conditioned on its prefix and the other in-batch texts, with cross-text conditioning weighted by retriever-computed similarities. Using MS MARCO as the only paired query-document supervision and a 135M LM for generative regularization with unlabeled raw-text initialization, ElicitR consistently improves BRIGHT by 16-29% relative across 0.1B–3B retriever scales while maintaining performance on BEIR. At 3B, ElicitR reaches an nDCG@10 of 23.1, substantially outperforming larger models trained with far more curated pairs and proprietary APIs. Further analyses show that ElicitR prevents overfitting, improves retrieval calibration, and remains robust to batch sizes, supporting its practicality.

","absKo":"

Reasoning-intensive retrieval은 downstream application에서 점점 더 중요해지고 있으며, 단순한 lexical overlap이나 거친 semantic matching 이상의 것을 요구한다. 기존 연구는 주로 Language Models (LMs)에 의존해 reasoning-oriented supervision을 합성하지만, 우리는 그것이 이미 LM 기반 retriever 내부에 잠재되어 있으나 contrastive overfitting에 의해 억제되고 있다고 본다. 이러한 잠재 reasoning을 끌어내기 위해, 우리는 이진 relevance를 넘어 query와 candidate document 사이의 미묘한 관계를 포착하는 generative regularization을 갖춘 retriever–LM framework인 ElicitR을 제안한다. 구체적으로, contrastive learning과 함께 query–positive–negative batch로 작은 LM을 co-training하여 retriever를 regularize한다. 각 text에 대한 Next token prediction (NTP)은 해당 prefix와 batch 내 다른 text에 조건화되며, text 간 conditioning은 retriever가 계산한 similarity로 가중된다. MS MARCO를 유일한 paired query-document supervision으로 사용하고, unlabeled raw-text initialization과 함께 135M LM을 generative regularization에 활용했을 때, ElicitR은 BEIR 성능을 유지하면서 0.1B–3B retriever 규모 전반에서 BRIGHT를 상대적으로 16-29% 일관되게 향상시킨다. 3B에서는 ElicitR이 nDCG@10 23.1을 달성하여, 훨씬 더 많은 curated pair와 proprietary API로 학습된 더 큰 모델들을 크게 능가한다. 추가 분석은 ElicitR이 overfitting을 방지하고 retrieval calibration을 개선하며 batch size에 대해 견고함을 유지함을 보여 주어, 그 실용성을 뒷받침한다.

"},{"id":"61761","en":"HiPPO Zoo: Making Implicit State Space Memory Explicit","ko":"HiPPO Zoo: 암묵적 State Space Memory의 명시화","authors":"Jack Goffinet, Casey Hanks, David Carlson","pos":"#1607","link":"https://openreview.net/forum?id=nB0TrIRAs1","abs":"

Representing the past in a compressed, efficient, and informative manner is a central problem for systems trained on sequential data. The HiPPO framework, originally proposed by Gu & Dao et al., provides a principled approach to sequential compression by projecting signals onto orthogonal polynomial (OP) bases via structured linear ordinary differential equations. Subsequent works have embedded these dynamics in state space models (SSMs), where HiPPO structure serves as an initialization. Nonlinear successors of these SSM methods such as Mamba are state of the art for many tasks with long-range dependencies, but the mechanisms by which they represent and prioritize history remain largely implicit. In this work, we revisit the HiPPO framework with the goal of making these mechanisms explicit. We show how polynomial representations of history can be extended to support capabilities of modern SSMs such as adaptive allocation of memory, and input-dependent state updates, and associative memory, while retaining direct interpretability in the OP basis. We introduce a unified framework comprising five such extensions, which we collectively refer to as a \"HiPPO zoo.\" Each extension exposes a specific modeling capability as an explicit modification of the underlying measure or dynamics governing the polynomial coefficients, rather than as an opaque learned transformation. The resulting models adapt their memory online and train in streaming settings with efficient updates. We illustrate the behaviors and advantages of these extensions through a range of synthetic sequence modeling tasks, highlighting how explicit polynomial memories can recover and clarify mechanisms implicit in SSMs.

","absKo":"압축적이면서 효율적이고 정보량이 풍부한 방식으로 과거를 표현하는 것은 sequential data로 학습하는 시스템에서 핵심 문제다. 원래 Gu & Dao et al.이 제안한 HiPPO framework는 structured linear ordinary differential equation을 통해 신호를 orthogonal polynomial (OP) basis에 투영함으로써 sequential compression에 대한 원리적 접근을 제공한다. 이후 연구들은 이러한 dynamics를 state space model (SSM)에 내장했으며, 여기서 HiPPO 구조는 initialization으로 사용된다. Mamba와 같은 이러한 SSM method의 비선형 후속 기법들은 긴 범위 의존성을 갖는 많은 작업에서 state of the art이지만, history를 표현하고 우선순위를 부여하는 메커니즘은 여전히 대부분 암묵적이다. 본 연구에서는 이러한 메커니즘을 명시적으로 드러내는 것을 목표로 HiPPO framework를 재검토한다. 우리는 역사에 대한 polynomial representation이 적응적 memory allocation, input-dependent state update, associative memory와 같은 현대 SSM의 기능을 지원하도록 어떻게 확장될 수 있는지 보이면서, OP basis에서 직접적인 해석 가능성은 유지함을 보인다. 우리는 이러한 확장 다섯 가지를 아우르는 통합 framework를 제안하며, 이를 collectively \"HiPPO zoo\"라고 부른다. 각 확장은 기본 measure 또는 polynomial coefficient를 지배하는 dynamics에 대한 명시적 수정으로서 특정 modeling capability를 드러내며, 불투명한 learned transformation으로서가 아니다. 그 결과 모델은 memory를 online으로 적응시키고 efficient update를 통해 streaming setting에서 학습한다. 우리는 다양한 synthetic sequence modeling task를 통해 이러한 확장의 행동과 장점을 설명하며, 명시적 polynomial memory가 SSM에 암묵적으로 존재하는 메커니즘을 어떻게 복원하고 명확히 하는지 강조한다.

"},{"id":"64378","en":"MuLoCo: Muon is a Practical Inner Optimizer for DiLoCo","ko":"MuLoCo: Muon은 DiLoCo를 위한 실용적 내부 옵티마이저다","authors":"Benjamin Thérien, Xiaolong Huang, Aaron Defazio, Irina Rish, Eugene Belilovsky","pos":"#3600","link":"https://openreview.net/forum?id=OBXBXSgwcD","abs":"DiLoCo is a powerful framework for training large language models (LLMs) under networking constraints, allowing for increased parallelism and accelerator utilization in data center settings. A critical but often overlooked factor in DiLoCo’s behavior is the choice of inner optimizer, which shapes the pseudogradient used by the outer optimizer. Given the recent success of Muon relative to AdamW for data parallel (DP) training, in this work, we examine how Muon's normalized optimizer steps can affect the pseudogradient's quality. Empirically, we find that, relative to AdamW, Muon yields more \\emph{directionally correct} pseudogradients as the number of workers ($K$) is increased. In our experiments pre-training language models, we conduct extensive hyperparameter tuning across 150M, 416M, 914M, 1.76B, and 3.1B models for DiLoCo, MuLoCo, AdamW DP, and Muon DP. Consistently across all scales, we find that with $K\\geq1$ workers, MuLoCo (Muon inner optimizer DiLoCo) achieves superior performance to DiLoCo in absolute terms and for $K>2$ it outperforms DiLoCo relative to their data parallel baselines, while being compatible with quantization, streaming, and long synchronization intervals. At $K=1$, we find that MuLoCo can even outperform the data-parallel gold standard while having larger optimal and critical batch sizes.","absKo":"DiLoCo는 networking 제약 하에서 large language models(LLMs)을 학습하기 위한 강력한 framework로, data center 환경에서 병렬성과 accelerator 활용도를 높일 수 있게 해준다. DiLoCo의 동작에서 중요하지만 자주 간과되는 요소는 inner optimizer의 선택이며, 이는 outer optimizer가 사용하는 pseudogradient를 형성한다. 최근 data parallel(DP) 학습에서 AdamW에 비해 Muon이 성공을 거둔 점에 착안하여, 본 연구에서는 Muon의 normalized optimizer step이 pseudogradient의 품질에 어떤 영향을 미치는지 살펴본다. 실험적으로, AdamW와 비교했을 때 worker 수($K$)가 증가할수록 Muon이 더 \\emph{방향적으로 올바른} pseudogradient를 생성함을 확인했다. language model pre-training 실험에서는 DiLoCo, MuLoCo, AdamW DP, Muon DP에 대해 150M, 416M, 914M, 1.76B, 3.1B 모델 전반에 걸쳐 광범위한 hyperparameter tuning을 수행했다. 모든 scale에서 일관되게, $K\\geq1$ worker일 때 MuLoCo(Muon inner optimizer DiLoCo)가 절대 성능 면에서 DiLoCo보다 우수했으며, $K>2$에서는 data parallel baseline 대비 성능에서도 DiLoCo를 능가했다. 동시에 quantization, streaming, 긴 synchronization interval과도 호환된다. $K=1$일 때는 MuLoCo가 더 큰 optimal batch size와 critical batch size를 가지면서도 data-parallel gold standard를 능가할 수 있음도 확인했다."},{"id":"66436","en":"Rethinking the Design Space of Reinforcement Learning for Diffusion Models: On the Importance of Likelihood Estimation Beyond Loss Design","ko":"Diffusion Model을 위한 강화 학습의 설계 공간 재고: Loss 설계를 넘어선 우도 추정의 중요성","authors":"Jaemoo Choi, Yuchen Zhu, Wei Guo, Petr Molodyk, Bo Yuan, Jinbin Bai, Yi Xin, Molei Tao, Yongxin Chen","pos":"#2714","link":"https://openreview.net/forum?id=3mJ9eTU319","abs":"Reinforcement learning has been widely applied to diffusion and flow models for visual tasks such as text-to-image generation. However, these tasks remain challenging because diffusion models have intractable likelihoods, which creates a barrier for directly applying popular policy-gradient type methods. Existing approaches primarily focus on crafting new objectives built on already heavily engineered LLM objectives, using ad hoc estimators for likelihood, without a thorough investigation into how such estimation affects overall algorithmic performance. In this work, we provide a systematic analysis of the RL design space by disentangling three factors: i) policy-gradient objectives, ii) likelihood estimators, and iii) rollout sampling schemes. We show that adopting an evidence lower bound (ELBO) based model likelihood estimator, computed only from the final generated sample, is the dominant factor enabling effective, efficient, and stable RL optimization, outweighing the impact of the specific policy-gradient loss functional. We validate our findings across multiple reward benchmarks using SD 3.5 Medium, and observe consistent trends across all tasks. Our method improves the GenEval score from $0.24$ to $0.95$ in $90$ GPU hours, which is $4.6\\times$ more efficient than FlowGRPO and $2\\times$ more efficient than the SOTA method DiffusionNFT without reward hacking.","absKo":"텍스트-투-이미지 생성과 같은 visual task에 대해 reinforcement learning은 diffusion 및 flow model에 널리 적용되어 왔다. 그러나 diffusion model은 intractable likelihood를 가지므로, 널리 사용되는 policy-gradient 유형 방법을 직접 적용하는 데 장벽이 된다. 기존 접근법은 주로 이미 많이 엔지니어링된 LLM objective 위에 구축된 새로운 objective를 만드는 데 초점을 맞추며, likelihood를 위한 ad hoc estimator를 사용하지만, 이러한 estimation이 전체 algorithmic performance에 미치는 영향을 충분히 조사하지는 않았다. 본 연구에서는 i) policy-gradient objective, ii) likelihood estimator, iii) rollout sampling scheme의 세 가지 요소를 분리하여 RL design space를 체계적으로 분석한다. 우리는 최종 생성 샘플만으로 계산한 evidence lower bound (ELBO) 기반 model likelihood estimator를 채택하는 것이 효과적이고 효율적이며 안정적인 RL 최적화를 가능하게 하는 지배적인 요인임을 보이며, 특정 policy-gradient loss function의 영향보다 더 중요함을 보인다. 우리는 SD 3.5 Medium을 사용한 여러 reward benchmark에서 이러한 발견을 검증했으며, 모든 task에서 일관된 경향을 관찰했다. 우리의 방법은 $90$ GPU hour 만에 GenEval score를 $0.24$에서 $0.95$로 향상시키며, reward hacking 없이 FlowGRPO보다 $4.6\\times$ 더 효율적이고 SOTA method DiffusionNFT보다 $2\\times$ 더 효율적이다."},{"id":"62783","en":"Avoid What You Know: Divergent Trajectory Balance for GFlowNets","ko":"이미 알고 있는 것은 피하라: GFlowNets를 위한 Divergent Trajectory Balance","authors":"Pedro Dall’Antonia, Tiago Silva, Daniel Csillag, Salem Lahlou, Diego Mesquita","pos":"#3301","link":"https://openreview.net/forum?id=dPDOQs7qHl","abs":"

Generative Flow Networks (GFlowNets) are a flexible family of amortized samplers trained to generate discrete and compositional objects with probability proportional to a reward function. To this end, they learn a policy function over an intractably large state graph by minimizing a stochastic objective over sampled trajectories. However, learning efficiency is constrained by the model’s ability to rapidly explore diverse high-probability regions during training. To mitigate this issue, recent works have focused on incentivizing the exploration of unvisited and valuable states via curiosity-driven search and self-supervised random network distillation, which tend to waste samples on already well-approximated regions of the state space. In this context, we propose Adaptive Complementary Exploration (ACE), a principled algorithm for the effective exploration of novel and high-probability regions when learning GFlowNets. To achieve this, ACE introduces an exploration GFlowNet explicitly trained to search for high-reward states in regions underexplored by the canonical GFlowNet, which learns to sample from the target distribution. Through extensive experiments, we show that ACE consistently and significantly improves upon prior work in terms of approximation accuracy to the target distribution and discovery rate of diverse high-reward states.

","absKo":"

Generative Flow Networks (GFlowNets)는 reward function에 비례하는 확률로 discrete하고 compositional한 object를 생성하도록 학습되는 amortized sampler의 유연한 계열이다. 이를 위해, 이들은 sampled trajectory에 대한 stochastic objective를 최소화함으로써 계산적으로 다루기 어려울 정도로 큰 state graph 위에서 policy function을 학습한다. 그러나 학습 효율성은 training 동안 다양한 high-probability region을 얼마나 빠르게 탐색할 수 있는지에 의해 제약된다. 이 문제를 완화하기 위해 최근 연구들은 curiosity-driven search와 self-supervised random network distillation을 통해 방문하지 않은 가치 있는 state의 탐색을 장려하는 데 초점을 맞추어 왔지만, 이러한 방법들은 이미 잘 근사된 state space 영역에 샘플을 낭비하는 경향이 있다. 이러한 맥락에서, 우리는 GFlowNet 학습 시 새로운 high-probability region을 효과적으로 탐색하기 위한 원칙적인 알고리즘인 Adaptive Complementary Exploration (ACE)을 제안한다. 이를 위해 ACE는 target distribution에서 샘플링하도록 학습되는 canonical GFlowNet이 충분히 탐색하지 못한 영역에서 high-reward state를 찾도록 명시적으로 학습된 exploration GFlowNet을 도입한다. 광범위한 실험을 통해, ACE가 target distribution에 대한 approximation accuracy와 다양한 high-reward state의 발견율 측면에서 기존 연구보다 일관되고 유의미하게 향상됨을 보인다.

"},{"id":"60826","en":"On the Effect of Misspecifying the Embedding Dimension in Low-rank Network Models","ko":"Low-Rank Network Model에서 Embedding Dimension 오지정의 영향","authors":"Roddy Taing, Keith Levin","pos":"#3303","link":"https://openreview.net/forum?id=wIMGGV9l1i","abs":"

As network data has become ubiquitous in the sciences, there has been growing interest in network models whose structure is driven by latent node-level variables in a (typically low-dimensional) latent geometric space. These \"latent positions\" are often estimated via embeddings, whereby the nodes of a network are mapped to points in Euclidean space so that \"similar\" nodes are mapped to nearby points. Under certain model assumptions, these embeddings are consistent estimates of the latent positions, but most such results require the embedding dimension to be chosen correctly. Methods for choosing the embedding dimension have been studied extensively, but little is known about the behavior of embeddings when the dimension is misspecified. In this work, we provide a theoretical description of the effects of dimension misspecification under the random dot product graph, a class of latent space network models that includes several widely-used network models, most notably the stochastic blockmodel, as special cases. We show that when the dimension is chosen too large, consistent estimation still holds, albeit at a slower rate than when the embedding dimension is chosen correctly. On the other hand, when the dimension is chosen too small, there is a fundamental estimation error lower bound that need not go to zero in the large-network limit. A range of synthetic data experiments support our theoretical results. Our main technical result, which may be of independent interest, is a generalization of earlier work in random matrix theory showing that all non-signal eigenvectors of a low-rank matrix subject to additive noise are delocalized.

","absKo":"

network data가 과학 전반에서 보편화되면서, 구조가 latent geometric space의 latent node-level variable에 의해 주도되는 network model에 대한 관심이 커지고 있다. 이러한 \"latent position\"은 종종 embedding을 통해 추정되는데, 여기서 network의 node는 Euclidean space의 점으로 매핑되어 \"similar\"한 node가 서로 가까운 점으로 배치된다. 특정 model assumption 하에서는 이러한 embedding이 latent position의 consistent estimate가 되지만, 대부분의 결과는 embedding dimension을 정확히 선택해야 한다는 전제를 둔다. embedding dimension을 선택하는 방법은 광범위하게 연구되어 왔지만, dimension이 misspecified될 때 embedding이 어떻게 거동하는지는 거의 알려져 있지 않다. 이 연구에서는 stochastic blockmodel을 포함한 여러 널리 사용되는 network model을 special case로 포함하는 latent space network model의 한 부류인 random dot product graph에서 dimension misspecification의 효과를 이론적으로 설명한다. 우리는 dimension을 너무 크게 선택하더라도 consistent estimation이 여전히 성립하지만, embedding dimension을 올바르게 선택했을 때보다 느린 rate를 갖는다는 것을 보인다. 반면 dimension을 너무 작게 선택하면, large-network limit에서 0으로 가지 않아도 되는 근본적인 estimation error lower bound가 존재한다. 다양한 synthetic data 실험이 우리의 이론 결과를 뒷받침한다. 독립적으로도 흥미로운 우리의 주요 technical result는, additive noise가 있는 low-rank matrix에서 signal이 아닌 모든 eigenvector가 delocalized됨을 보인 이전 random matrix theory 연구를 일반화한 것이다.

"},{"id":"63387","en":"Functional Adjoint Sampler: Scalable Sampling on Infinite Dimensional Spaces","ko":"Functional Adjoint Sampler: 무한 차원 공간에서의 확장 가능한 샘플링","authors":"Byoungwoo Park, Juho Lee, Guan-Horng Liu","pos":"#3408","link":"https://openreview.net/forum?id=XcJk9un0Tb","abs":"

Learning-based methods for sampling from the Gibbs distribution in finite-dimensional spaces have progressed quickly, yet theory and algorithmic design for infinite-dimensional function spaces remain limited. This gap persists despite their strong potential for sampling the paths of conditional diffusion processes, enabling efficient simulation of trajectories of diffusion processes that respect rare events or boundary constraints. In this work, we present the adjoint sampler for infinite-dimensional function spaces, a stochastic optimal control-based diffusion sampler that operates in function space and targets Gibbs-type distributions on infinite-dimensional Hilbert spaces. Our Functional Adjoint Sampler (FAS) generalizes Adjoint Sampling (Havens et al., 2025) to Hilbert spaces based on a SOC theory called stochastic maximum principle, yielding a simple and scalable matching-type objective for a functional representation. We show that FAS achieves superior transition path sampling performance across synthetic potential and real molecular systems, including Alanine Dipeptide and Chignolin.

","absKo":"

유한차원 공간에서 Gibbs distribution으로부터 sampling하는 learning-based method는 빠르게 발전해 왔지만, infinite-dimensional function space를 위한 theory와 algorithmic design은 여전히 제한적이다. 이 간극은 conditional diffusion process의 path를 sampling하는 강력한 잠재력에도 불구하고 남아 있으며, rare event나 boundary constraint를 따르는 diffusion process의 trajectory를 효율적으로 simulation할 수 있게 한다. 본 연구에서는 infinite-dimensional function space를 위한 adjoint sampler를 제시한다. 이는 function space에서 동작하며 infinite-dimensional Hilbert space 위의 Gibbs-type distribution을 목표로 하는 stochastic optimal control 기반 diffusion sampler이다. 우리의 Functional Adjoint Sampler(FAS)는 stochastic maximum principle이라 불리는 SOC theory를 바탕으로 Havens et al., 2025의 Adjoint Sampling을 Hilbert space로 일반화하며, functional representation을 위한 간단하고 확장 가능한 matching-type objective를 도출한다. 우리는 FAS가 Alanine Dipeptide와 Chignolin을 포함한 합성 potential 및 실제 molecular system 전반에서 우수한 transition path sampling 성능을 달성함을 보인다.

"},{"id":"63536","en":"Trajectory-Aware Heuristic Learning for Combinatorial Search","ko":"조합 탐색을 위한 궤적 인지 휴리스틱 학습","authors":"Mustafa Seddiqi, Tiberiu Popa, Marta Kersten-Oertel","pos":"#3503","link":"https://openreview.net/forum?id=WKPfaCG2FJ","abs":"

Learning effective value heuristics for combinatorial search is difficult, as prior methods rely on surrogate supervision or costly downstream search to assess progress. We introduce a trajectory-aware probabilistic framework that models uncertainty in cost-to-go labels instead of treating them as fixed targets. Heuristic learning is cast as inference over state trajectories using an HMM-style model (Rabiner, 1989), where estimated depth-change dynamics define transitions and forward–backward inference yields soft supervision. To evaluate heuristic quality without search, we propose a large-scale local ranking metric that measures a model’s ability to order neighboring states. On the Rubik’s Cube, our approach consistently improves local ranking accuracy and downstream search performance under matched computational budgets.

","absKo":"조합 탐색을 위한 효과적인 value heuristic을 학습하는 일은 어렵다. 기존 방법이 progress를 평가하기 위해 surrogate supervision이나 비용이 큰 downstream search에 의존하기 때문이다. 우리는 cost-to-go label을 고정된 target으로 취급하지 않고 그 불확실성을 모델링하는 trajectory-aware probabilistic framework를 제안한다. heuristic learning은 state trajectories 위의 inference로 정식화되며, HMM-style model (Rabiner, 1989)을 사용해 estimated depth-change dynamics가 transitions를 정의하고 forward–backward inference가 soft supervision을 제공한다. search 없이 heuristic quality를 평가하기 위해, 이웃한 states를 올바르게 순서화하는 model의 능력을 측정하는 대규모 local ranking metric을 제안한다. Rubik's Cube에서 우리의 접근법은 matched computational budgets 하에서 local ranking accuracy와 downstream search performance를 일관되게 향상시킨다.

"},{"id":"66606","en":"Maximum-Likelihood Learning of Latent Dynamics Without Reconstruction","ko":"재구성 없는 잠재 동역학의 최대 우도 학습","authors":"Samo Hromadka, Kai Biegun, Lior Fox, James Heald, Maneesh Sahani","pos":"#3504","link":"https://openreview.net/forum?id=2CliXIoJNZ","abs":"

We address the challenge of uncovering systematic, and potentially controllable, dynamical structure underlying complex high-dimensional time series data. Existing generative and autoregressive models have difficulty telling systematic content apart from distractors, while contrastive methods struggle to learn accurate dynamics. To address these shortcomings, we develop the Recognition-Parametrized Gaussian State Space Model (RP-GSSM), a probabilistic framework that infers accurate latent dynamics without relying on a parametrized decoder. By eliminating explicit generative parameters, the model directs its entire representational capacity to encoding dynamically relevant state; and, being fully probabilistic, learns via maximum likelihood without auxiliary objectives or ad-hoc regularization. Combining the expressive power of a neural network encoder with exact inference under a jointly Gaussian prior allows the RP-GSSM to embed a broad class of intrinsically nonlinear dynamical systems. We show that the RP-GSSM recovers physically meaningful latent states from noisy video more faithfully than competing methods, more reliably identifies underlying controllable nonlinear dynamics, and remains substantially more robust to visual distractors.

","absKo":"

우리는 복잡한 고차원 시계열 데이터에 내재한 체계적이며 잠재적으로 제어 가능한 dynamical structure를 밝혀내는 문제를 다룬다. 기존 generative 및 autoregressive model은 체계적 content와 distractor를 구분하는 데 어려움을 겪는 반면, contrastive method는 정확한 dynamics를 학습하는 데 한계가 있다. 이러한 단점을 해결하기 위해, 우리는 parameterized decoder에 의존하지 않고 정확한 latent dynamics를 추론하는 probabilistic framework인 Recognition-Parametrized Gaussian State Space Model (RP-GSSM)을 개발한다. 명시적인 generative parameter를 제거함으로써 모델은 모든 representational capacity를 dynamically relevant state를 인코딩하는 데 집중할 수 있고, 완전한 probabilistic 모델이므로 auxiliary objective나 ad-hoc regularization 없이 maximum likelihood로 학습된다. neural network encoder의 표현력과 jointly Gaussian prior 하에서의 exact inference를 결합함으로써 RP-GSSM은 광범위한 intrinsically nonlinear dynamical system을 임베딩할 수 있다. 우리는 RP-GSSM이 노이즈가 있는 video로부터 physically meaningful latent state를 경쟁 방법보다 더 충실하게 복원하고, underlying controllable nonlinear dynamics를 더 신뢰성 있게 식별하며, visual distractor에 대해 훨씬 더 강건함을 보인다.

"},{"id":"64171","en":"Exact and Approximate Algorithms for Polytree Learning","ko":"폴리트리 학습을 위한 정확 및 근사 알고리즘","authors":"Juha Harviainen, Frank Sommer, Manuel Sorge","pos":"#3505","link":"https://openreview.net/forum?id=QIBu6vtRMI","abs":"Polytrees are a subclass of Bayesian networks that seek to capture the conditional dependencies between a set of $n$ variables as a directed forest and are motivated by their more efficient inference and improved interpretability. Since the problem of learning the best polytree is NP-hard, we study which restrictions make it more tractable by considering for example in-degree bounds, properties of score functions measuring the quality of a polytree, and approximation algorithms. We devise an algorithm that finds the optimal polytree in time $\\mathcal{O}\\big((2+\\epsilon)^n\\big)$ for arbitrarily small $\\epsilon > 0$ and any constant in-degree bound $k$, improving over the fastest previously known algorithm of time complexity $\\mathcal{O}\\big(3^n\\big)$. We further give polynomial-time algorithms for finding a polytree whose score is within a factor of $k$ from the optimal one for arbitrary scores and a factor of $2$ for additive ones. Many of the results are complemented by (nearly) tight lower bounds for either the time complexity or the approximation factors.","absKo":"폴리트리(polytrees)는 Bayesian network의 부분 클래스로, $n$개의 변수 집합 사이의 conditional dependency를 directed forest로 포착하고자 하며, 더 효율적인 inference와 향상된 interpretability를 장점으로 한다. 최적의 polytree를 학습하는 문제는 NP-hard이므로, 우리는 in-degree bound, polytree의 품질을 측정하는 score function의 성질, approximation algorithm 등을 고려하여 어떤 제약이 이를 더 tractable하게 만드는지 연구한다. 우리는 임의로 작은 $\\epsilon > 0$과 어떤 상수 in-degree bound $k$에 대해서도 시간복잡도 $\\mathcal{O}\\big((2+\\epsilon)^n\\big)$에 최적의 polytree를 찾는 알고리즘을 고안했으며, 이는 이전에 알려진 가장 빠른 $\\mathcal{O}\\big(3^n\\big)$ 알고리즘을 개선한 것이다. 또한 arbitrary score에 대해서는 최적값의 $k$배 이내, additive score에 대해서는 $2$배 이내의 score를 갖는 polytree를 다항시간에 찾는 알고리즘도 제시한다. 많은 결과는 시간복잡도 또는 approximation factor에 대한 (거의) tight한 lower bound로 보완된다."},{"id":"65544","en":"Transformed Latent Variable Multi-Output Gaussian Processes","ko":"변환된 잠재 변수 Multi-Output Gaussian Process","authors":"Xiaoyu Jiang, Xinxing Shi, Sokratia Georgaka, Magnus Rattray, Mauricio Álvarez","pos":"#3506","link":"https://openreview.net/forum?id=CXKhhN5j1N","abs":"Multi-Output Gaussian Processes (MOGPs) provide a principled probabilistic framework for modelling correlated outputs but face scalability bottlenecks when applied to datasets with high-dimensional output spaces. To maintain tractability, existing methods typically resort to restrictive assumptions, such as employing low-rank or sum-of-separable kernels, which can limit expressiveness. We propose the Transformed Latent Variable MOGP (T-LVMOGP), a novel framework that scales MOGPs to a massive number of outputs while preserving the capacity to capture meaningful inter-output dependencies. T-LVMOGP constructs a flexible multi-output deep kernel by mapping inputs and output-specific latent variables into an embedding space using a Lipschitz-regularised neural network. Combined with stochastic variational inference, our model effectively scales to high-dimensional output settings. Across diverse benchmarks, including climate modelling with over $10,000$ outputs and zero-inflated spatial transcriptomics data, T-LVMOGP outperforms baselines in both predictive accuracy and computational efficiency.","absKo":"Multi-Output Gaussian Processes (MOGPs)는 상관된 출력을 모델링하는 데 있어 원리적인 확률적 프레임워크를 제공하지만, 고차원 출력 공간을 가진 데이터셋에 적용할 때 확장성 병목에 직면한다. 계산 가능성을 유지하기 위해 기존 방법들은 보통 저랭크(low-rank) 또는 sum-of-separable kernel과 같은 제한적인 가정에 의존하는데, 이는 표현력을 제한할 수 있다. 우리는 T-LVMOGP(Transformed Latent Variable MOGP)를 제안한다. 이는 의미 있는 출력 간 의존성을 포착하는 능력을 유지하면서 MOGP를 대규모 출력 수로 확장하는 새로운 프레임워크이다. T-LVMOGP는 Lipschitz-regularised neural network를 사용해 입력과 출력별 latent variable을 embedding space로 매핑함으로써 유연한 multi-output deep kernel을 구성한다. 여기에 stochastic variational inference를 결합하여, 본 모델은 고차원 출력 설정으로 효과적으로 확장된다. 10,000개가 넘는 출력을 가진 climate modelling과 zero-inflated spatial transcriptomics data를 포함한 다양한 benchmark에서, T-LVMOGP는 예측 정확도와 계산 효율성 모두에서 baseline을 능가한다."},{"id":"61407","en":"SCORE: A Unified Framework for Overshoot Refund in Online FDR Control","ko":"SCORE: Online FDR Control에서 Overshoot Refund를 위한 통합 Framework","authors":"Qi Kuang, Bowen Gang, Yin Xia","pos":"#3507","link":"https://openreview.net/forum?id=qX4Nm7eNM5","abs":"We propose a unified framework to enhance the power of online multiple hypothesis testing procedures based on $e$-values. While $e$-value-based methods offer robust online False Discovery Rate (FDR) control under minimal assumptions, they often suffer from power loss by discarding evidence that exceeds the rejection threshold. We address this inefficiency via the **S**equential **C**ontrol with **O**vershoot **R**efund for **E**-values (SCORE) framework, which leverages the inequality $\\mathbb{I}(y \\ge 1) \\le y - (y-1)_+$ to reclaim this otherwise ``wasted'' evidence. This simple yet powerful insight yields a unified principle for improving a broad class of online testing algorithms. Building on this framework, we develop SCORE-enhanced versions of several state-of-the-art procedures, including SCORE-LOND, SCORE-LORD, and SCORE-SAFFRON, all of which strictly dominate their original counterparts while preserving valid finite-sample FDR control. Furthermore, under mild assumptions, SCORE permits retroactive updates of alpha‑wealth by using the latest decision twice: first to determine its reward or loss, and then to refresh past wealth. Such a mechanism enables more aggressive testing strategies while maintaining valid FDR control, thereby further improving statistical power. The effectiveness of the proposed methods is validated through extensive simulation and real-data experiments.","absKo":"우리는 $e$-value에 기반한 online multiple hypothesis testing procedure의 power를 향상시키기 위한 통합 framework를 제안한다. $e$-value 기반 방법은 최소한의 가정 하에서 강건한 online False Discovery Rate (FDR) 제어를 제공하지만, rejection threshold를 초과한 evidence를 버리기 때문에 종종 power 손실을 겪는다. 우리는 $\\mathbb{I}(y \\ge 1) \\le y - (y-1)_+$라는 부등식을 활용해 이처럼 ``낭비''되는 evidence를 회수하는, **S**equential **C**ontrol with **O**vershoot **R**efund for **E**-values (SCORE) framework를 통해 이러한 비효율을 해결한다. 이 단순하지만 강력한 통찰은 광범위한 online testing algorithm의 성능을 향상시키는 통합 원리를 제공한다. 이 framework를 바탕으로 우리는 SCORE-LOND, SCORE-LORD, SCORE-SAFFRON을 포함한 여러 state-of-the-art procedure의 SCORE-enhanced 버전을 개발했으며, 이들은 모두 유효한 finite-sample FDR 제어를 유지하면서 기존 버전을 엄격하게 지배한다. 또한 완만한 가정 하에서, SCORE는 최신 decision을 두 번 사용함으로써 alpha‑wealth를 소급적으로 업데이트할 수 있다. 첫째는 reward 또는 loss를 결정하는 데, 둘째는 과거 wealth를 갱신하는 데 사용한다. 이러한 메커니즘은 유효한 FDR 제어를 유지하면서 더 공격적인 testing strategy를 가능하게 하여, statistical power를 추가로 향상시킨다. 제안한 방법의 효과는 광범위한 simulation과 real-data experiment를 통해 검증되었다."},{"id":"62194","en":"Rethinking the Flow-based Gradual Domain Adaption: A Semi-Dual Optimal Transport Perspective","ko":"Flow 기반 Gradual Domain Adaptation 재고찰: Semi-Dual Optimal Transport 관점","authors":"Zhichao Chen, Zhan Zhuang, Yunfei Teng, Eric Wang, Fangyikang Wang, Zhengnan Li, Tianqiao Liu, Haoxuan Li, Zhouchen Lin","pos":"#3508","link":"https://openreview.net/forum?id=iqXzDUd36x","abs":"Gradual Domain Adaption (GDA) aims to mitigate domain shift by progressively adapting models from the source domain to the target domain via intermediate domains. However, real intermediate domains are often unavailable or ineffective, necessitating the synthesis of intermediate samples. Flow-based models are recently used for this purpose by interpolating between source and target distributions, but their training typically resorts to sample-based log-likelihood estimation, which can discard useful information and thus degrade GDA performance. The key to addressing this limitation is constructing the intermediate domains via samples directly. To this end, we propose an $\\underline{\\text{E}}$ntropy-regularized $\\underline{\\text{S}}$emi-dual $\\underline{\\text{U}}$nbalanced $\\underline{\\text{O}}$ptimal $\\underline{\\text{T}}$ransport (E-SUOT) framework to construct intermediate domains. Specifically, we reformulate flow-based GDA as a Lagrangian dual problem and derive an equivalent objective that circumvents the needs for likelihood estimation. However, the dual problem results in the unstable min–max training procedure. To alleviate this issue, we further introduce entropy regularization to convert it into a more stable alternative optimization procedure. Based on this, we propose a novel GDA training framework and provide theoretical analysis in terms of stability and generalization. Finally, extensive experiments are conducted to demonstrate the efficacy of the E-SUOT framework.","absKo":"Gradual Domain Adaption (GDA)는 중간 도메인을 통해 소스 도메인에서 타깃 도메인으로 모델을 점진적으로 적응시켜 domain shift를 완화하는 것을 목표로 한다. 그러나 실제 중간 도메인은 종종 존재하지 않거나 효과적이지 않기 때문에, 중간 샘플을 합성할 필요가 있다. 최근에는 flow-based model이 소스와 타깃 분포 사이를 보간하는 방식으로 이러한 목적에 활용되고 있지만, 그 학습은 대개 sample-based log-likelihood estimation에 의존하며, 이 과정에서 유용한 정보를 버릴 수 있어 GDA 성능이 저하된다. 이러한 한계를 해결하는 핵심은 샘플을 직접 사용해 중간 도메인을 구성하는 것이다. 이를 위해 우리는 중간 도메인을 구성하기 위한 $\\underline{\\text{E}}$ntropy-regularized $\\underline{\\text{S}}$emi-dual $\\underline{\\text{U}}$nbalanced $\\underline{\\text{O}}$ptimal $\\underline{\\text{T}}$ransport (E-SUOT) framework를 제안한다. 구체적으로, flow-based GDA를 Lagrangian dual problem으로 재구성하고 likelihood estimation의 필요를 우회하는 동등한 objective를 도출한다. 그러나 dual problem은 불안정한 min–max training procedure를 초래한다. 이 문제를 완화하기 위해, 우리는 추가로 entropy regularization을 도입하여 이를 더 안정적인 alternative optimization procedure로 변환한다. 이를 바탕으로 새로운 GDA training framework를 제안하고, 안정성과 generalization 측면에서 이론적 분석을 제공한다. 마지막으로, E-SUOT framework의 효용을 입증하기 위해 광범위한 실험을 수행한다."},{"id":"65107","en":"Learning to Bet for Horizon-Aware Anytime-Valid Testing","ko":"지평선 인지 Anytime-Valid 검정을 위한 베팅 학습","authors":"Ege Onur Taga, Samet Oymak, Shubhanshu Shekhar","pos":"#3509","link":"https://openreview.net/forum?id=GiSCgWW0Ze","abs":"We develop horizon-aware anytime-valid tests and confidence sequences for bounded means under a strict deadline $N$. Using the betting/e-process framework, we cast horizon-aware betting as a finite-horizon optimal control problem with state space $(t, \\log W_t)$, where $t$ is the time and $W_t$ is the test martingale value. We first show that in certain interior regions of the state space, policies that deviate significantly from Kelly betting are provably suboptimal, while Kelly betting reaches the threshold with high probability. We then identify sufficient conditions showing that outside this region, more aggressive betting than Kelly can be better if log-wealth is low or time is short, and less aggressive can be better if log-wealth is high. Taken together these results suggest a simple phase diagram in the $(t, \\log W_t)$ plane, delineating regions where Kelly, fractional Kelly, and aggressive betting may be preferable. Guided by this phase diagram, we introduce a Deep Reinforcement Learning approach based on a universal Deep Q-Network (DQN) agent that learns a single policy from synthetic experience and maps simple statistics of past observations to bets across horizons and null values. In limited-horizon experiments, the learned DQN policy outperforms state-of-the-art baselines.","absKo":"엄격한 마감 $N$ 아래에서 bounded mean에 대한 horizon-aware anytime-valid test와 confidence sequence를 개발한다. betting/e-process framework를 사용하여, 우리는 horizon-aware betting을 state space $(t, \\log W_t)$를 갖는 finite-horizon optimal control problem으로 정식화하는데, 여기서 $t$는 시간이고 $W_t$는 test martingale 값이다. 먼저 상태 공간의 특정 내부 영역에서는 Kelly betting에서 크게 벗어나는 policy가 증명 가능하게 비최적이며, Kelly betting은 높은 확률로 threshold에 도달함을 보인다. 이어서 이 영역 밖에서는 log-wealth가 낮거나 시간이 짧을 때는 Kelly보다 더 공격적인 betting이 더 나을 수 있고, log-wealth가 높을 때는 덜 공격적인 betting이 더 나을 수 있음을 보여주는 충분 조건을 식별한다. 종합하면 이러한 결과는 $(t, \\log W_t)$ 평면에서 Kelly, fractional Kelly, 그리고 aggressive betting이 각각 선호될 수 있는 영역을 구분하는 간단한 phase diagram을 시사한다. 이 phase diagram에 따라 우리는 synthetic experience로부터 단일 policy를 학습하고, 과거 관측의 간단한 statistic을 horizon과 null value에 걸친 bet로 매핑하는 universal Deep Q-Network(DQN) agent에 기반한 Deep Reinforcement Learning 접근을 제안한다. 제한된 horizon 실험에서 학습된 DQN policy는 state-of-the-art baseline을 능가한다."},{"id":"63332","en":"Colorful Pinball: Density-Weighted Quantile Regression for Conditional Guarantee of Conformal Prediction","ko":"Colorful Pinball: Conformal Prediction의 조건부 보장을 위한 밀도 가중 분위 회귀","authors":"Qianyi Chen, Bo Li","pos":"#3510","link":"https://openreview.net/forum?id=Y2lYfMD0z8","abs":"While conformal prediction provides robust marginal coverage guarantees, achieving reliable conditional coverage for specific inputs remains challenging. Although exact distribution-free conditional coverage is impossible with finite samples, recent work has focused on improving the conditional coverage of standard conformal procedures. Distinct from approaches that target relaxed notions of conditional coverage, we directly minimize the mean squared error of conditional coverage by refining the quantile regression components that underpin many conformal methods. Leveraging a Taylor expansion, we derive a sharp surrogate objective for quantile regression: a density-weighted pinball loss, where the weights are given by the conditional density of the conformity score evaluated at the true quantile. We propose a three-headed quantile network that estimates these weights via finite differences using auxiliary quantile levels at $1-\\alpha \\pm \\delta$, subsequently fine-tuning the central quantile by optimizing the weighted loss. We provide a theoretical analysis with exact non-asymptotic guarantees characterizing the resulting excess risk. Extensive experiments on diverse high-dimensional real-world datasets demonstrate remarkable improvements in conditional coverage performance.","absKo":"Conformal prediction은 강건한 marginal coverage 보장을 제공하지만, 특정 input에 대해 신뢰할 수 있는 conditional coverage를 달성하는 것은 여전히 어렵다. 정확한 distribution-free conditional coverage는 finite sample에서 불가능하지만, 최근 연구는 표준 conformal procedure의 conditional coverage를 개선하는 데 집중해 왔다. 완화된 conditional coverage 개념을 목표로 하는 접근과는 달리, 우리는 conformal method의 기반이 되는 quantile regression component를 정교하게 다듬어 conditional coverage의 mean squared error를 직접 최소화한다. Taylor expansion을 활용해, 우리는 quantile regression을 위한 sharp surrogate objective를 유도한다. 이는 density-weighted pinball loss이며, 가중치는 true quantile에서 평가한 conformity score의 conditional density로 주어진다. 우리는 auxiliary quantile level $1-\\alpha \\pm \\delta$를 사용해 finite difference로 이러한 가중치를 추정하고, 이후 weighted loss를 최적화하여 central quantile을 fine-tuning하는 three-headed quantile network를 제안한다. 우리는 그 결과로 얻어지는 excess risk를 특성화하는 exact non-asymptotic guarantee를 포함한 이론적 분석을 제공한다. 다양한 고차원 실제 데이터셋에 대한 광범위한 실험은 conditional coverage 성능에서 놀라운 향상을 보여준다."},{"id":"64297","en":"Identifying Common Hubs in Multiple Gaussian Graphical Models","ko":"다중 가우시안 그래프 모델에서 공통 허브 식별","authors":"José Á. Sánchez Gómez, Weibin Mo, Junlong Zhao, Yufeng Liu","pos":"#3706","link":"https://openreview.net/forum?id=P0lcBIhzv0","abs":"

The Gaussian graphical model (GGM) is a useful tool to represent relationships of conditional dependence among variables. In many real-world applications, datasets often contain multiple related sub-populations, whose associated GGMs may have common structure, as well as large structural differences. In such cases, it is useful to recover common hub variables, which are the highly connected variables in the GGMs of all sub-populations. In this paper, we propose the Joint Inverse Components for Hub Detection (JIC-HD) method to recover the common hubs across multiple GGMs without the need to estimate all subpopulation GGMs. To this end, we introduce joint minimax eigenspaces, and show that these can be leveraged for the recovery of common hubs. We establish theoretical guarantees for the recovery of common hubs. Additionally, our numerical simulation studies confirm superior performance of our JIC-HD in detecting common hubs compared to the existing methods in the literature. Our method is especially advantageous when the multiple GGMs have both common and individual hubs across sub-populations. Finally, we analyze cancer gene-expression datasets and identify biologically meaningful common hub genes across cancer subtypes.

","absKo":"

Gaussian graphical model(GGM)은 변수들 사이의 조건부 의존 관계를 나타내는 데 유용한 도구이다. 많은 실제 응용에서 데이터셋은 종종 서로 관련된 여러 부분집단을 포함하며, 이들에 대응하는 GGM은 공통 구조를 가지는 동시에 큰 구조적 차이도 보일 수 있다. 이러한 경우, 모든 부분집단의 GGM에서 높은 연결성을 지닌 변수들인 공통 hub 변수를 복원하는 것이 유용하다. 본 논문에서는 모든 subpopulation GGM을 추정할 필요 없이 여러 GGM 전반의 공통 hub를 복원하기 위한 Joint Inverse Components for Hub Detection(JIC-HD) 방법을 제안한다. 이를 위해 joint minimax eigenspaces를 도입하고, 이것이 공통 hub 복원에 활용될 수 있음을 보인다. 또한 공통 hub 복원에 대한 이론적 보장을 확립한다. 추가로, 수치 시뮬레이션 연구를 통해 기존 문헌의 방법들과 비교했을 때 JIC-HD가 공통 hub 검출에서 더 우수한 성능을 보임을 확인한다. 우리의 방법은 여러 GGM이 subpopulation 간에 공통 hub와 개별 hub를 모두 가질 때 특히 유리하다. 마지막으로, cancer gene-expression datasets를 분석하여 cancer subtype 전반에 걸친 생물학적으로 의미 있는 공통 hub gene을 식별한다.

"},{"id":"66099","en":"On the Identifiability of Poisson Branching Structural Causal Model Under Latent Confounding","ko":"잠재 교란 하에서 Poisson 분기 구조적 인과 모델의 식별 가능성에 관하여","authors":"Jie Qiao, Zihuai Zeng, Ruichu Cai, Zhengming Chen, Zhifeng Hao","pos":"#3805","link":"https://openreview.net/forum?id=73YmKB7KpW","abs":"

Causal discovery from observational count data poses unique challenges, particularly when the data exhibit inherent branching structures, e.g., an upstream event (e.g., an ad impression) triggers a downstream event (e.g., a purchase) with a certain probability. Such branching dynamics are naturally captured by thinning operators (for the branching structure) and an independent Poisson distribution (for exogenous noise), constituting the Poisson Branching Structural Causal Model (PB-SCM). However, existing approaches based on PB-SCM rely on the restrictive assumption of causal sufficiency, failing to account for ubiquitous latent confounders that can bias estimation. In this work, we propose the Latent Confounding Poisson Branching Structural Causal Model (LC-PB-SCM) to bridge this gap. We leverage Probability Generating Functions (PGFs) to characterize the complex dependencies introduced by latent confounding. Then, we establish a Trie representation theorem that maps the branching causal mechanisms to the algebraic properties of PGF monomials. Based on local PGFs, we establish a complete identifiability condition for local 3-variables that covers all causal patterns distinguishable up to monomial equivalence. Finally, we propose a practical algorithm to learn causal structures under latent confounding and demonstrate its effectiveness through experiments on both synthetic and real-world datasets.

","absKo":"

관측 count data로부터의 causal discovery는 고유한 도전 과제를 지니며, 특히 데이터가 본질적인 branching 구조를 보일 때 더욱 그렇다. 예를 들어, upstream event(예: 광고 노출)가 일정 확률로 downstream event(예: 구매)를 유발하는 경우가 이에 해당한다. 이러한 branching dynamics는 thinning operator(branching 구조)와 독립적인 Poisson 분포(exogenous noise)에 의해 자연스럽게 포착되며, 이는 Poisson Branching Structural Causal Model (PB-SCM)을 이룬다. 그러나 PB-SCM에 기반한 기존 접근법은 causal sufficiency라는 제한적인 가정에 의존하여, 추정에 편향을 초래할 수 있는 널리 존재하는 latent confounder를 반영하지 못한다. 본 연구에서는 이 간극을 메우기 위해 Latent Confounding Poisson Branching Structural Causal Model (LC-PB-SCM)을 제안한다. 우리는 Probability Generating Functions (PGFs)를 활용하여 latent confounding이 도입하는 복잡한 의존성을 특성화한다. 그런 다음 branching causal mechanism을 PGF monomial의 대수적 성질에 대응시키는 Trie representation theorem을 정립한다. local PGF를 바탕으로, monomial equivalence까지 구분 가능한 모든 causal pattern을 포괄하는 local 3-variable에 대한 완전한 identifiability 조건을 확립한다. 마지막으로, latent confounding 하에서 causal structure를 학습하는 실용적 알고리즘을 제안하고, 합성 및 실제 데이터셋 모두에서의 실험을 통해 그 효과를 입증한다.

"},{"id":"65697","en":"Conformal Risk-Averse Decision Making with Action Conditional Guarantee","ko":"행동 조건부 보장을 갖춘 Conformal 위험 회피 의사결정","authors":"Zihan Zhu, Shayan Kiyani, George Pappas, Hamed Hassani","pos":"#3809","link":"https://openreview.net/forum?id=B5RiXueKqb","abs":"

Reliable decision making pipelines powered by machine learning models require uncertainty quantification (UQ) methods that come with explicit safety guarantees. Conformal prediction provides such UQ by wrapping ML predictions into prediction sets, and recent work by \\cite{kiyani2025decision} established that these sets can be translated into optimal risk-averse decision policies—yet only inheriting marginal safety guarantees. We generalize and strengthen their results by (i) introducing action-conditional conformal prediction, which yields safety guarantees conditioned explicitly on each action taken by the decision maker, (ii) showing that action-conditional prediction sets serve as a proxy for the feasible decision space for risk-averse decision makers aiming to optimize action-conditional value-at-risk, and (iii) proposing a principled finite-sample algorithm based on pinball-loss minimization, connecting the framework of \\cite{gibbs2025conformal} to action-conditional guarantees. Experiments on two real-world datasets confirm that our approach significantly improves action-conditional performance over several conformal baselines.

","absKo":"

machine learning model로 구동되는 신뢰할 수 있는 decision making pipeline은 명시적 safety guarantee를 갖춘 uncertainty quantification(UQ) method를 필요로 한다. Conformal prediction은 ML prediction을 prediction set으로 감싸는 방식으로 이러한 UQ를 제공하며, \\cite{kiyani2025decision}의 최근 연구는 이 set을 optimal risk-averse decision policy로 변환할 수 있음을 보였다. 그러나 그 경우에도 marginal safety guarantee만을 상속한다. 우리는 다음을 통해 그 결과를 일반화하고 강화한다. (i) action-conditional conformal prediction을 도입하여 decision maker가 취한 각 action에 조건부인 safety guarantee를 제공하고, (ii) action-conditional prediction set이 action-conditional value-at-risk를 최적화하려는 risk-averse decision maker에게 feasible decision space의 proxy로 기능함을 보이며, (iii) pinball-loss minimization에 기반한 원칙적인 finite-sample algorithm을 제안하여 \\cite{gibbs2025conformal}의 framework를 action-conditional guarantee와 연결한다. 두 개의 real-world dataset에서 수행한 실험은 우리의 접근법이 여러 conformal baseline에 비해 action-conditional performance를 유의하게 향상시킴을 확인한다.

"},{"id":"65391","en":"Epistemic Uncertainty Quantification for Pre-trained VLMs via Riemannian Flow Matching","ko":"Riemannian Flow Matching을 통한 사전학습 VLM의 인식론적 불확실성 정량화","authors":"Li Ju, Mayank Nautiyal, Andreas Hellander, Ekta Vats, Prashant Singh","pos":"#4010","link":"https://openreview.net/forum?id=E8Q093jDug","abs":"

Vision-Language Models (VLMs) are typically deterministic in nature and lack intrinsic mechanisms to quantify epistemic uncertainty, which reflects the model’s lack of knowledge or ignorance of its own representations. We theoretically motivate negative log-density of an embedding as a proxy for the epistemic uncertainty, where low-density regions signify model ignorance. The proposed method REPVLM computes the probability density on the hyperspherical manifold of the VLM embeddings using Riemannian Flow Matching. We empirically demonstrate that REPVLM achieves near-perfect correlation between uncertainty and prediction error, significantly outperforming existing baselines. Beyond classification, we also demonstrate that the model also provides a scalable metric for out-of-distribution detection and automated data curation.

","absKo":"

Vision-Language Models (VLMs)는 일반적으로 deterministic하며, 모델이 자신의 representation에 대해 얼마나 모르고 있는지를 반영하는 epistemic uncertainty를 정량화할 내재적 mechanism이 부족하다. 우리는 embedding의 negative log-density를 epistemic uncertainty의 proxy로 이론적으로 동기화하며, 낮은 density region은 model ignorance를 의미한다. 제안하는 방법 REPVLM은 Riemannian Flow Matching을 사용해 VLM embedding의 hyperspherical manifold 위에서 probability density를 계산한다. 실험적으로 REPVLM이 uncertainty와 prediction error 사이에서 거의 완벽에 가까운 상관을 달성하며, 기존 baseline을 크게 능가함을 보인다. classification을 넘어, 이 model이 out-of-distribution detection과 automated data curation을 위한 확장 가능한 metric도 제공함을 보인다.

"},{"id":"62798","en":"Geometry-Aware Probabilistic Circuits via Voronoi Tessellations","ko":"Voronoi Tessellations 기반 Geometry-Aware Probabilistic Circuits","authors":"Sahil Sidheekh, Sriraam Natarajan","pos":"#413","link":"https://openreview.net/forum?id=d9LsOCOB9z","abs":"

Probabilistic circuits (PCs) enable exact and tractable inference but employ data independent mixture weights that limit their ability to capture local geometry of the data manifold. We propose Voronoi tessellations (VT) as a natural way to incorporate geometric structure directly into the sum nodes of a PC. However, naïvely introducing such structure breaks tractability. We formalize this incompatibility and develop two complementary solutions: (1) an approximate inference framework that provides guaranteed lower and upper bounds for inference, and (2) a structural condition for VT under which exact tractable inference is recovered. Finally, we introduce a differentiable relaxation for VT that enables gradient-based learning and empirically validate the resulting approach on standard density estimation tasks.

","absKo":"

Probabilistic circuits (PCs)는 정확하고 tractable한 inference를 가능하게 하지만, 데이터와 무관한 mixture weight를 사용하기 때문에 데이터 manifold의 local geometry를 포착하는 능력이 제한됩니다. 우리는 geometric structure를 PC의 sum node에 직접 통합하는 자연스러운 방법으로 Voronoi tessellation (VT)을 제안합니다. 그러나 이러한 구조를 단순히 도입하면 tractability가 깨집니다. 우리는 이 비양립성을 formalize하고, 상호 보완적인 두 가지 해법을 개발합니다: (1) inference에 대해 보장된 lower bound와 upper bound를 제공하는 approximate inference framework, 그리고 (2) exact tractable inference를 복원할 수 있는 VT의 structural condition입니다. 마지막으로, gradient-based learning을 가능하게 하는 VT의 differentiable relaxation을 도입하고, 표준 density estimation task에서 그 결과 방법을 실증적으로 검증합니다.

"},{"id":"60976","en":"Optimal structure learning and conditional independence testing","ko":"최적 Structure Learning과 Conditional Independence Testing","authors":"Ming Gao, Yuhao Wang, Bryon Aragam","pos":"#4204","link":"https://openreview.net/forum?id=uurjpxrLc5","abs":"

We establish a fundamental connection between optimal structure learning and optimal conditional independence testing by showing that the minimax optimal rate for structure learning problems is determined by the minimax rate for conditional independence testing in these problems. This is accomplished by establishing a general reduction between these two problems in the case of poly-forests, and demonstrated by deriving optimal rates for several examples, including Bernoulli, Gaussian and nonparametric models. Furthermore, we show that the optimal algorithm in these settings is a suitable modification of the PC algorithm. This theoretical finding provides a unified framework for analyzing the statistical complexity of structure learning through the lens of minimax testing.

","absKo":"

우리는 구조 학습 문제의 minimax 최적 속도가 이들 문제에서 conditional independence testing의 minimax rate로 결정된다는 것을 보임으로써, optimal structure learning과 optimal conditional independence testing 사이의 근본적인 연결을 확립한다. 이는 poly-forests의 경우 이 두 문제 사이의 일반적인 reduction을 확립함으로써 달성되었고, Bernoulli, Gaussian 및 nonparametric model을 포함한 여러 예제에 대해 optimal rate를 도출함으로써 입증된다. 더 나아가, 이러한 설정에서의 optimal algorithm이 PC algorithm의 적절한 수정 버전임을 보인다. 이 이론적 결과는 minimax testing의 관점에서 structure learning의 통계적 복잡성을 분석하기 위한 통합적 framework를 제공한다.

"},{"id":"64392","en":"Prediction-Powered Risk Monitoring of Deployed Models for Detecting Harmful Distribution Shifts","ko":"배포된 모델의 예측 기반 위험 모니터링을 통한 해로운 분포 변화 탐지","authors":"Guangyi Zhang, Yunlong Cai, Guanding Yu, Osvaldo Simeone","pos":"#4413","link":"https://openreview.net/forum?id=O57TtEyCrW","abs":"

We study the problem of monitoring model performance in dynamic environments where labeled data are limited. To this end, we propose prediction-powered risk monitoring (PPRM), a semi-supervised risk-monitoring approach based on prediction-powered inference (PPI). PPRM constructs anytime-valid lower bounds on the running risk by combining synthetic labels with a small set of true labels. Harmful shifts are detected via a threshold-based comparison with an upper bound on the nominal risk, satisfying assumption-free finite-sample guarantees in the probability of false alarm. We demonstrate the effectiveness of PPRM through extensive experiments on image classification, large language model (LLM), and telecommunications monitoring tasks.

","absKo":"

라벨이 제한된 동적 환경에서 모델 성능을 모니터링하는 문제를 연구한다. 이를 위해 prediction-powered inference (PPI)를 기반으로 한 반지도 위험 모니터링 기법인 prediction-powered risk monitoring (PPRM)을 제안한다. PPRM은 합성 라벨과 소수의 실제 라벨을 결합하여 running risk에 대한 anytime-valid lower bound를 구성한다. 유해한 변화는 명목 위험(nominal risk)의 upper bound와의 임계값 기반 비교를 통해 탐지되며, 오경보 확률에 대해 가정 없는 유한표본 보장을 만족한다. 이미지 분류, large language model (LLM), 통신 모니터링 태스크에 대한 광범위한 실험을 통해 PPRM의 효과를 입증한다.

"},{"id":"67109","en":"Position: Epistemic uncertainty estimation methods are fundamentally incomplete","ko":"Position: 인식적 불확실성 추정 방법은 근본적으로 불완전하다","authors":"Sebastian Jimenez, Mira Juergens, Willem Waegeman","pos":"#1612","link":"https://openreview.net/forum?id=g598HZM6ib","abs":"

Identifying and disentangling sources of predictive uncertainty is essential for trustworthy supervised learning. We argue that widely used second-order decomposition-based approaches to uncertainty quantification are fundamentally incomplete. First, we show that unaccounted bias contaminates uncertainty estimates by overestimating aleatoric (data-related) uncertainty and underestimating the epistemic (model-related) counterpart, leading to systematically incorrect uncertainty quantification. Second, we demonstrate that existing methods capture only partial contributions to the variance-driven part of epistemic uncertainty; different approaches account for different variance sources, yielding estimates that are incomplete and difficult to interpret. Together, these results highlight that current epistemic uncertainty estimates can only be used in safety-critical and high-stakes decision-making when limitations are fully understood by end users and acknowledged by AI developers.

","absKo":"

예측 불확실성의 source를 식별하고 분리하는 것은 신뢰할 수 있는 supervised learning에 필수적이다. 우리는 널리 사용되는 second-order decomposition 기반 uncertainty quantification 접근법이 근본적으로 불완전하다고 주장한다. 첫째, 고려되지 않은 bias가 uncertainty estimate를 오염시켜 aleatoric (data-related) uncertainty를 과대평가하고 epistemic (model-related) counterpart를 과소평가함으로써, 체계적으로 잘못된 uncertainty quantification을 초래함을 보인다. 둘째, 기존 방법은 variance-driven part에 기반한 epistemic uncertainty의 부분적인 기여만 포착한다는 점을 보여 준다. 서로 다른 접근법은 서로 다른 variance source를 반영하므로, 추정치는 불완전하고 해석하기 어렵다. 이 결과를 종합하면, 현재의 epistemic uncertainty estimate는 안전이 중요한 safety-critical 및 high-stakes decision-making에서 end user가 그 한계를 완전히 이해하고 AI 개발자가 이를 인정할 때에만 사용할 수 있음을 시사한다.

"},{"id":"61682","en":"ANCHOR: Abductive Network Construction with Hierarchical Orchestration for Reliable Probability Inference in Large Language Models","ko":"ANCHOR: Large Language Model에서 신뢰할 수 있는 확률 추론을 위한 계층적 Orchestration 기반 Abductive Network Construction","authors":"Wentao Qiu, Guanran Luo, Zhongquan Jian, Jingqi Gao, Meihong Wang, Qingqiang Wu","pos":"#2701","link":"https://openreview.net/forum?id=no1Tgzkw0x","abs":"

A central challenge in large-scale decision-making under incomplete information is estimating reliable probabilities. Recent approaches leverage Large Language Models (LLMs) to generate explanatory factors and elicit coarse-grained probability estimates. Typically, an LLM performs forward abduction to propose factors, each paired with two mutually exclusive attributes, and a Naïve Bayes model is trained over factor combinations to refine the final probabilities. However, the induced factor space is often sparse, leading to frequent ''unknown'' outcomes when the system cannot map a query context to any supported factor configuration. Simply expanding the factor set to increase coverage is ineffective: it amplifies statistical noise and introduces spurious correlations that violate the conditional-independence assumption, ultimately degrading stability and reliability.To address these limitations, we propose Anchor, an inference framework that orchestrates aggregated Bayesian inference over a hierarchically structured factor space. Anchor first constructs a dense and organized factor space via iterative generation and hierarchical clustering. It then performs context-aware mapping through hierarchical retrieval and refinement, substantially reducing ''unknown'' predictions. Finally, Anchor augments Naïve Bayes with a Causal Bayesian Network to capture latent dependencies among factors, relaxing the strict independence assumption. Experiments show that Anchor markedly reduces ''unknown'' predictions and produces more reliable probability estimates than direct LLM baselines, achieving state-of-the-art performance while significantly reducing time and token overhead.

","absKo":"불완전한 정보 하에서의 대규모 의사결정에서 핵심 과제는 신뢰할 수 있는 확률을 추정하는 일이다. 최근 접근법은 Large Language Models(LLMs)를 활용해 설명적 요인을 생성하고, 거친 수준의 확률 추정을 유도한다. 일반적으로 LLM은 forward abduction을 수행해 두 개의 상호 배타적인 attribute와 쌍을 이루는 요인을 제안하고, Naïve Bayes model은 factor 조합 위에서 학습되어 최종 확률을 정교화한다. 그러나 이렇게 유도된 factor space는 종종 sparse하여, 시스템이 query context를 지원되는 어떤 factor configuration에도 매핑하지 못할 때 ''unknown'' 결과가 자주 발생한다. coverage를 늘리기 위해 factor set을 단순히 확장하는 것은 효과적이지 않다. 이는 statistical noise를 증폭시키고 conditional-independence assumption을 위반하는 가짜 상관관계를 도입하여, 궁극적으로 stability와 reliability를 저하시킨다. 이러한 한계를 해결하기 위해 우리는 Anchor를 제안한다. Anchor는 계층적으로 구조화된 factor space 위에서 aggregated Bayesian inference를 조율하는 inference framework이다. Anchor는 먼저 iterative generation과 hierarchical clustering을 통해 dense하고 정돈된 factor space를 구성한다. 그런 다음 hierarchical retrieval과 refinement를 통해 context-aware mapping을 수행하여 ''unknown'' prediction을 크게 줄인다. 마지막으로 Anchor는 Naïve Bayes에 Causal Bayesian Network를 결합해 factor 간 latent dependency를 포착함으로써, 엄격한 independence assumption을 완화한다. 실험 결과, Anchor는 직접적인 LLM baseline보다 ''unknown'' prediction을 현저히 줄이고 더 신뢰할 수 있는 probability estimate를 생성했으며, time과 token overhead를 크게 줄이면서 state-of-the-art 성능을 달성했다.

"},{"id":"65055","en":"The surprising strength of weak classifiers for validating neural posterior estimates","ko":"Neural Posterior 추정 검증을 위한 약한 분류기의 놀라운 강력함","authors":"Vansh Bansal, Tianyu Chen, James Scott","pos":"#2710","link":"https://openreview.net/forum?id=HDoWJi2jlj","abs":"The two-sample testing problem, a fundamental task in statistics and machine learning, seeks to determine whether two sets of samples, drawn from underlying distributions $p$ and $q$, are in fact identically distributed (i.e.~whether $p=q$). A popular and intuitive approach is the classifier two-sample test (C2ST), where a classifier is trained to distinguish between samples from $p$ and $q$. Yet despite simplicity of the C2ST, its reliability hinges on access to a near-Bayes-optimal classifier, a requirement that is rarely met and difficult to verify. This raises a major open question: can a weak classifier still be useful for two-sample testing? We show that the answer is a definitive yes. Building on the work of Hu & Lei (2024), we analyze two conformal variants of the C2ST that convert the scores from any trained classifier---even if weak, biased, or overfit---into exact, finite-sample p-values. We establish two key theoretical properties of the conformal C2ST: (i) finite-sample Type-I error control, and (ii) non-trivial power that degrades gently in tandem with the error of the trained classifier. The upshot is that even poorly performing classifiers can yield powerful and reliable two-sample tests. This general framework finds a powerful application in Bayesian inference, particularly for validating Neural Posterior Estimation (NPE) models, where the task of comparing a learned posterior approximation $q(\\theta \\mid y)$ to the true posterior $p(\\theta \\mid y)$ can be framed as a two-sample test. Empirically, the Conformal C2ST outperforms classical discriminative tests across a wide range of benchmarks for this task. Our results establish the conformal C2ST as a practical, theoretically grounded diagnostic tool.","absKo":"두 표본 검정(two-sample testing) 문제는 통계와 머신러닝의 기본적인 task로, underlying distribution $p$와 $q$에서 추출된 두 샘플 집합이 실제로는 동일한 분포를 따르는지, 즉 $p=q$인지 판별하는 것을 목표로 한다. 널리 쓰이고 직관적인 접근법은 classifier two-sample test(C2ST)로, 분류기를 학습시켜 $p$와 $q$의 샘플을 구분한다. 그러나 C2ST는 단순함에도 불구하고 near-Bayes-optimal classifier에 대한 접근성에 신뢰성이 달려 있으며, 이러한 조건은 실제로는 거의 충족되지 않고 검증도 어렵다. 이는 중요한 열린 질문을 낳는다. 약한 분류기라도 two-sample testing에 유용할 수 있는가? 우리는 그 답이 분명히 그렇다고 보인다. Hu & Lei(2024)의 연구를 바탕으로, 우리는 어떤 학습된 분류기라도, 심지어 약하거나 편향되었거나 overfit된 경우에도, 그 score를 정확한 finite-sample p-value로 변환하는 C2ST의 두 가지 conformal variant를 분석한다. 우리는 conformal C2ST의 두 가지 핵심 이론적 성질을 확립한다. (i) finite-sample Type-I error control, (ii) 학습된 분류기의 error와 함께 완만하게 악화되는 비자명한 power. 요컨대, 성능이 좋지 않은 분류기라도 강력하고 신뢰할 수 있는 two-sample test를 제공할 수 있다. 이 일반적인 framework는 Bayesian inference에서 특히 유용한데, learned posterior approximation $q(\\theta \\mid y)$와 true posterior $p(\\theta \\mid y)$를 비교하는 task를 two-sample test로 정식화할 수 있기 때문이다. 실험적으로, Conformal C2ST는 이 task를 위한 다양한 benchmark 전반에서 classical discriminative test를 능가한다. 우리의 결과는 conformal C2ST를 실용적이면서도 이론적으로 정당화된 진단 도구로 확립한다."},{"id":"60908","en":"Disentangling Latent Risk Pathways via Bayesian Hypergraph Inference","ko":"Bayesian Hypergraph Inference를 통한 잠재 Risk Pathway 분리","authors":"Shengxian Ding, Haonan Gao, Pangpang Liu, Xinyuan Tian, Yize Zhao","pos":"#3302","link":"https://openreview.net/forum?id=vNfbqRzash","abs":"

Electronic health records (EHR) pose large-scale multi-disease modeling problems in which many outcomes are rare and strongly influenced by shared risk factors. While modern approaches achieve strong predictive performance, they often treat diseases independently or rely on black-box architectures, offering limited insight into how risk factors organize disease risk and little principled uncertainty quantification. We introduce a Bayesian hypergraph inference framework that reframes multi-disease modeling around latent, risk-factor-modulated disease pathways. Risk factors act on hyperedges, latent disease subsets with shared risk patterns, allowing diseases to participate in multiple distinct pathways and enabling interpretable, higher-order structure beyond pairwise associations. A repulsion prior encourages parsimonious and identifiable structure, while posterior inference provides calibrated uncertainty over both disease groupings and risk-factor influence. To enable scalable inference on large EHR datasets, we develop a structured variational inference algorithm that preserves logical dependencies among hyperedge existence, disease membership, and pathway-level effects. Experiments on simulated data and UK Biobank demonstrate stable and interpretable disease pathway structure, well-calibrated uncertainty, improved estimation for rare diseases, and competitive predictive performance.

","absKo":"

Electronic health records(EHR)는 다수의 결과가 희귀하고 공유된 위험 요인의 강한 영향을 받는 대규모 다질병 모델링 문제를 제기한다. 현대적 접근법들은 강력한 예측 성능을 달성하지만, 질병을 서로 독립적으로 취급하거나 black-box 아키텍처에 의존하는 경우가 많아, 위험 요인이 질병 위험을 어떻게 조직화하는지에 대한 통찰이 제한적이고 원칙적인 불확실성 정량화도 부족하다. 우리는 다질병 모델링을 latent, risk-factor-modulated disease pathways 중심으로 재구성하는 Bayesian hypergraph inference framework를 제안한다. 위험 요인은 hyperedge에 작용하며, 이는 공유된 위험 패턴을 갖는 잠재적 질병 부분집합으로서 질병이 여러 서로 다른 pathway에 동시에 참여할 수 있게 하고 pairwise association을 넘어서는 해석 가능하고 더 높은 차수의 구조를 가능하게 한다. 반발(prurulsion) prior는 간결하고 식별 가능한 구조를 장려하며, posterior inference는 질병 군집과 위험 요인 영향 모두에 대해 보정된 불확실성을 제공한다. 대규모 EHR 데이터셋에서 확장 가능한 추론을 가능하게 하기 위해, 우리는 hyperedge 존재, 질병 membership, pathway-level 효과 사이의 논리적 의존성을 보존하는 structured variational inference algorithm을 개발한다. 시뮬레이션 데이터와 UK Biobank에서의 실험은 안정적이고 해석 가능한 disease pathway 구조, 잘 보정된 불확실성, 희귀 질병에 대한 향상된 추정, 그리고 경쟁력 있는 예측 성능을 보여 준다.

"},{"id":"63152","en":"Variational Routing: A Scalable Bayesian Framework for Calibrated Mixture-of-Experts Transformers","ko":"Variational Routing: 보정된 Mixture-of-Experts Transformer를 위한 확장 가능한 베이지안 프레임워크","authors":"Albus Li, Matthew Wicker","pos":"#3512","link":"https://openreview.net/forum?id=ZZZkLWphuO","abs":"

Foundation models are increasingly being deployed in contexts where understanding the uncertainty of their outputs is critical to ensuring responsible deployment. While Bayesian methods offer a principled approach to uncertainty quantification, their computational overhead renders their use impractical for training or inference at foundation model scale. State-of-the-art models achieve parameter counts in the trillions through carefully engineered sparsity including Mixture-of-Experts (MoE) layers. In this work, we demonstrate calibrated uncertainty at scale by introducing Variational Mixture-of-Experts Routing (VMoER), a structured Bayesian approach for modelling uncertainty in MoE layers. VMoER confines Bayesian inference to the expert-selection stage which is typically done by a deterministic routing network. We instantiate VMoER using two inference strategies: amortised variational inference over routing logits and inferring a temperature parameter for stochastic expert selection. Across tested foundation models, VMoER improves routing stability under noise by 38%, reduces calibration error by 94%, and increases out-of-distribution AUROC by 12%, while incurring less than 1% additional FLOPs. These results suggest VMoER offers a scalable path toward robust and uncertainty-aware foundation models.

","absKo":"

Foundation model은 출력의 불확실성을 이해하는 것이 책임 있는 배포를 보장하는 데 중요해지는 상황에서 점점 더 많이 사용되고 있다. Bayesian method는 uncertainty quantification에 원칙적인 접근법을 제공하지만, 계산 오버헤드가 커서 foundation model 규모에서의 학습이나 추론에 사용하기에는 비현실적이다. 최신 state-of-the-art model은 Mixture-of-Experts (MoE) layer를 포함한 정교한 sparsity를 통해 수조 개의 parameter에 이른다. 이 연구에서는 MoE layer의 불확실성을 모델링하기 위한 구조화된 Bayesian 접근법인 Variational Mixture-of-Experts Routing (VMoER)을 도입하여, 대규모에서 calibrated uncertainty를 달성함을 보인다. VMoER는 Bayesian inference를 일반적으로 deterministic routing network가 수행하는 expert selection 단계에 국한한다. 우리는 routing logit에 대한 amortised variational inference와 stochastic expert selection을 위한 temperature parameter 추정을 사용하는 두 가지 inference 전략으로 VMoER를 구현한다. 테스트한 모든 foundation model에서 VMoER는 noise 하에서 routing stability를 38% 향상시키고, calibration error를 94% 감소시키며, out-of-distribution AUROC를 12% 높이는 동시에 추가 FLOPs는 1% 미만에 그친다. 이러한 결과는 VMoER가 강건하고 uncertainty-aware한 foundation model을 향한 확장 가능한 경로를 제공함을 시사한다.

"},{"id":"64500","en":"Tractable Expected Information Gains for Exponential Family Posteriors","ko":"지수족 사후 분포를 위한 다루기 쉬운 기대 정보 이득","authors":"Rik Knowles, Tom Rainforth","pos":"#3513","link":"https://openreview.net/forum?id=Mx1j6e6amP","abs":"We investigate which models admit a collapse of the expected information gain (EIG) and its derivative from a doubly intractable to a singly intractable expression. We prove that a sufficient condition is that the posterior distribution belongs to an exponential family (EF) that depends on the experimental design and data only through its natural parameters, and derive corresponding singly intractable and unbiased estimators for the $\\operatorname{EIG}$ and its (reparameterised) gradient. We further derive necessary conditions on the likelihood to obtain an EF posterior of the required form, showing that this does not necessarily require the prior to be conjugate. This is complemented by a theoretical analysis of certain degenerate behaviors that may arise when optimizing the $\\operatorname{EIG}$ for EF-modeled experiments. Finally, we empirically demonstrate the benefits of our singly intractable estimators, showing substantial performance gains over standard nested estimators.","absKo":"우리는 expected information gain (EIG)와 그 derivative가 doubly intractable한 expression에서 singly intractable한 expression으로 collapse될 수 있는 model이 무엇인지 조사한다. 우리는 posterior distribution이 experimental design과 data에 대해 오직 natural parameter를 통해서만 의존하는 exponential family (EF)에 속하는 것이 충분 조건임을 증명하고, 이에 대응하는 singly intractable하고 unbiased한 $\\operatorname{EIG}$ 및 그 (reparameterised) gradient 추정량을 유도한다. 또한 required form의 EF posterior를 얻기 위한 likelihood에 대한 필요 조건을 추가로 유도하여, prior가 반드시 conjugate일 필요는 없음을 보인다. 여기에 더해, EF-modeled experiment에서 $\\operatorname{EIG}$를 최적화할 때 나타날 수 있는 특정 degenerate behavior에 대한 이론적 분석을 제공한다. 마지막으로, 우리는 singly intractable estimator의 이점을 실증적으로 보여주며, 표준 nested estimator보다 상당한 성능 향상을 확인한다."},{"id":"61437","en":"Bayesian Tensor Decomposition with Diffusion Model Prior","ko":"Diffusion Model Prior를 이용한 Bayesian Tensor Decomposition","authors":"Zerui Tao, Qibin Zhao","pos":"#3601","link":"https://openreview.net/forum?id=q806xA8NPR","abs":"

Low-rank tensor decomposition (TD) is usually effective on clean, fully observed data, but it often degrades under severe missingness or noise. The low-rank constraint alone provides a weak inductive bias, while common handcrafted priors (e.g., sparsity or smoothness) fail to capture rich real-world structures. To compensate for this weak inductive bias under heavy corruption, one would like to inject a learned, data-driven prior; however, the state-of-the-art diffusion models are not readily compatible with current TD and tractable posterior inference. To address these challenges, we introduce DiffBCP, a Bayesian CP decomposition framework that combines a cumulative shrinkage process prior for automatic rank selection with an off-the-shelf pre-trained diffusion model as an implicit prior on the reconstructed tensor. To make posterior inference tractable despite the coupling among the likelihood, low-rank constraint, and diffusion prior, we develop a split Gibbs sampler: CP factors admit conjugate updates, while the diffusion block is sampled via low-rank-guided denoising. A noise-adaptive coupling schedule further reduces sensitivity to hand-tuned annealing. Experiments on image inpainting and denoising, including high-resolution out-of-distribution images, show consistent gains over Bayesian, nonlinear, and plug-and-play TD baselines.

","absKo":"

Low-rank tensor decomposition(TD)은 보통 깨끗하고 완전히 관측된 데이터에서는 효과적이지만, 심한 missingness나 noise 하에서는 성능이 자주 저하된다. low-rank constraint만으로는 약한 inductive bias만 제공하며, sparsity나 smoothness 같은 일반적인 handcrafted prior는 풍부한 실제 구조를 포착하지 못한다. 이러한 강한 corruption 하에서 약한 inductive bias를 보완하려면 learned, data-driven prior를 주입하고 싶지만, 최신 diffusion model은 현재의 TD 및 tractable posterior inference와 쉽게 결합되지 않는다. 이러한 문제를 해결하기 위해 우리는 Bayesian CP decomposition framework인 DiffBCP를 제안한다. 이 방법은 automatic rank selection을 위한 cumulative shrinkage process prior와 reconstructed tensor에 대한 implicit prior로서 off-the-shelf pre-trained diffusion model을 결합한다. likelihood, low-rank constraint, diffusion prior 사이의 coupling에도 불구하고 posterior inference를 tractable하게 만들기 위해, 우리는 split Gibbs sampler를 개발한다. CP factor는 conjugate update를 허용하고, diffusion block은 low-rank-guided denoising으로 샘플링된다. 또한 noise-adaptive coupling schedule은 손으로 조정하는 annealing에 대한 민감도를 추가로 줄인다. 이미지 inpainting과 denoising, including high-resolution out-of-distribution image에 대한 실험은 Bayesian, nonlinear, plug-and-play TD baseline보다 일관된 성능 향상을 보인다.

"},{"id":"63002","en":"Bayesian model selection and misspecification testing in imaging inverse problems only from noisy and partial measurements","ko":"잡음과 부분 측정치만으로 영상 inverse problems에서 Bayesian model selection과 misspecification 테스트","authors":"Tom Sprunck, Marcelo Pereyra, Tobías Liaudat","pos":"#3602","link":"https://openreview.net/forum?id=b1qoHUQVNy","abs":"

Modern imaging techniques heavily rely on Bayesian statistical models to address difficult image reconstruction and restoration tasks. This paper addresses the objective evaluation of such models in settings where ground truth is unavailable, with a focus on model selection and misspecification diagnosis. Existing unsupervised model evaluation methods are often unsuitable for computational imaging due to their high computational cost and incompatibility with modern image priors defined implicitly via machine learning models. We herein propose a general methodology for unsupervised model selection and misspecification detection in Bayesian imaging sciences, based on a novel combination of Bayesian cross-validation and data fission, a randomized measurement splitting technique. The approach is compatible with any Bayesian imaging sampler, including diffusion and plug-and-play samplers. We demonstrate the methodology through experiments involving various scoring rules and types of model misspecification, where we achieve excellent selection and detection accuracy with a low computational cost.

","absKo":"

현대 imaging technique은 어려운 image reconstruction 및 restoration task를 해결하기 위해 Bayesian statistical model에 크게 의존한다. 이 논문은 ground truth가 없는 환경에서 이러한 model을 객관적으로 평가하는 문제를 다루며, model selection과 misspecification diagnosis에 초점을 둔다. 기존의 unsupervised model evaluation 방법은 높은 계산 비용과 machine learning model을 통해 암묵적으로 정의되는 현대 image prior와의 비호환성 때문에 computational imaging에 적합하지 않은 경우가 많다. 우리는 Bayesian cross-validation과 randomized measurement splitting 기법인 data fission의 새로운 결합에 기반한, Bayesian imaging science에서의 unsupervised model selection 및 misspecification detection을 위한 일반적 methodology를 제안한다. 이 접근법은 diffusion과 plug-and-play sampler를 포함한 모든 Bayesian imaging sampler와 호환된다. 우리는 다양한 scoring rule과 model misspecification 유형을 포함한 실험을 통해 이 methodology를 입증하며, 낮은 계산 비용으로 매우 우수한 selection 및 detection 정확도를 달성함을 보인다.

"},{"id":"65695","en":"Cooperative variance estimation and Bayesian neural networks disentangle aleatoric and epistemic uncertainties","ko":"협력적 분산 추정과 Bayesian 신경망이 우연적 불확실성과 인식적 불확실성을 분리한다","authors":"Jiaxiang Yi, Miguel Bessa","pos":"#3603","link":"https://openreview.net/forum?id=B5umYh4CGr","abs":"

Real-world data contains aleatoric uncertainty -- irreducible noise arising from imperfect measurements or from incomplete knowledge about the data generation process. Mean variance estimation (MVE) networks can learn this type of uncertainty but require ad-hoc regularization strategies to avoid overfitting and are unable to predict epistemic uncertainty (model uncertainty). Conversely, Bayesian neural networks predict epistemic uncertainty but are notoriously difficult to train due to the approximate nature of Bayesian inference. We propose to cooperatively train a variance estimation network with a Bayesian neural network and empirically demonstrate that the resulting model disentangles aleatoric and epistemic uncertainties while improving the mean estimation. We demonstrate the effectiveness and scalability of this method across a diverse range of datasets, including a time-dependent heteroscedastic regression dataset we created where the aleatoric uncertainty is known, used to assess estimation accuracy. The proposed method is straightforward to implement, robust, and adaptable to various model architectures.

","absKo":"

실세계 데이터에는 aleatoric uncertainty, 즉 불완전한 측정이나 데이터 생성 과정에 대한 불완전한 지식에서 비롯되는 비가역적 잡음이 포함된다. Mean variance estimation(MVE) network는 이러한 종류의 uncertainty를 학습할 수 있지만, overfitting을 피하기 위해 ad-hoc regularization strategy를 필요로 하며 epistemic uncertainty(model uncertainty)를 예측할 수 없다. 반대로 Bayesian neural network는 epistemic uncertainty를 예측하지만, Bayesian inference의 근사적 성격 때문에 학습이 매우 어렵다. 우리는 variance estimation network와 Bayesian neural network를 협력적으로 학습하는 방법을 제안하며, 그 결과 모델이 aleatoric uncertainty와 epistemic uncertainty를 분리해 내면서 mean estimation도 향상됨을 실증적으로 보인다. 우리는 정확한 aleatoric uncertainty를 아는, 직접 생성한 time-dependent heteroscedastic regression dataset을 포함해 다양한 dataset 전반에서 이 방법의 효과와 확장성을 입증한다. 제안 방법은 구현이 간단하고 robust하며, 다양한 model architecture에 적응 가능하다.

"},{"id":"64517","en":"Is the Last Layer Sufficient for Uncertainty Quantification?","ko":"마지막 층만으로 불확실성 정량화에 충분한가?","authors":"Joseph Wilson, Chris van der Heide, Liam Hodgkinson, Fred Roosta","pos":"#3604","link":"https://openreview.net/forum?id=MocMOzHHcV","abs":"

Epistemic uncertainty quantification (UQ) for deep neural networks (DNNs) is a requirement for safe adoption of AI in mission-critical settings. Several leading methods for UQ linearize DNNs to form Bayesian Generalized Linear Models (GLMs), where epistemic uncertainty is modeled via the predictive posterior distribution. Linearizing around the parameters of the final connected layer of a DNN is a commonly used approximation for reducing the computational burden of such GLMs, though it is often believed to come at the cost of degraded performance. In this work, we compare GLMs arising from full-network and last-layer linearization using both theoretical and empirical approaches. We first employ tools from random matrix theory to conduct a theoretical comparison; this analysis reveals no meaningful improvement in the UQ capabilities of full linearization. Coupled with a large-scale empirical evaluation across a range of modern machine learning tasks, we arrive at the following conclusion: a last-layer approximation yields comparable UQ performance while offering substantially improved computational efficiency.

","absKo":"

deep neural network (DNN)의 epistemic uncertainty quantification (UQ)은 mission-critical 환경에서 AI를 안전하게 채택하기 위한 필수 요건이다. 여러 선도적 UQ 방법은 DNN을 linearize하여 Bayesian Generalized Linear Models (GLMs)를 구성하며, 여기서 epistemic uncertainty는 predictive posterior distribution을 통해 모델링된다. DNN의 final connected layer 파라미터 주변에서 linearizing하는 것은 이러한 GLM의 계산 부담을 줄이기 위해 널리 사용되는 근사이지만, 일반적으로 성능 저하를 수반한다고 여겨져 왔다. 본 연구에서는 full-network linearization과 last-layer linearization으로부터 유도되는 GLM을 이론적 및 경험적 방법으로 비교한다. 먼저 random matrix theory의 도구를 사용해 이론적 비교를 수행했으며, 이 분석은 full linearization의 UQ 성능이 의미 있게 향상되지 않음을 보여준다. 이어서 다양한 현대 machine learning task에 대한 대규모 경험적 평가를 통해, last-layer approximation이 상당히 향상된 계산 효율성을 제공하면서도 비슷한 UQ 성능을 낸다는 결론에 도달한다.

"},{"id":"65077","en":"Local MAP Sampling for Diffusion Models","ko":"Diffusion Model을 위한 Local MAP 샘플링","authors":"Shaorong Zhang, Rob Brekelmans, Greg Ver Steeg","pos":"#3605","link":"https://openreview.net/forum?id=GzunaSZ8jC","abs":"Diffusion Posterior Sampling (DPS) provides a principled Bayesian approach to inverse problems by sampling from $p(x_0 \\mid y)$. While posterior sampling is valuable for capturing uncertainty and multi-modality, many classical and practical inverse problem settings ultimately prioritize accurate point estimation—most notably the MAP estimator, which has long served as a standard reconstruction objective in imaging and scientific applications. We introduce Local MAP Sampling (LMAPS), a new inference framework that iteratively solving local MAP subproblems along the diffusion trajectory. This perspective clarifies their connection to global MAP and DPS, offering a unified probabilistic interpretation for optimization-based methods. Building on this foundation, we develop practical algorithms with a covariance approximation motivated by Gaussian prior assumption, a reformulated objective for stability and interpretability. Across a broad set of image restoration and scientific tasks, LMAPS achieves state-of-the-art performance.","absKo":"Diffusion Posterior Sampling(DPS)은 $p(x_0 \\mid y)$에서 sampling함으로써 inverse problem에 대한 원칙적인 Bayesian 접근을 제공한다. posterior sampling은 uncertainty와 multi-modality를 포착하는 데 유용하지만, 많은 고전적이고 실용적인 inverse problem 설정에서는 결국 정확한 point estimation을 우선시한다. 특히 MAP estimator는 오랫동안 imaging 및 scientific application에서 표준 reconstruction objective로 사용되어 왔다. 우리는 diffusion trajectory를 따라 local MAP subproblem을 반복적으로 푸는 새로운 inference framework인 Local MAP Sampling(LMAPS)를 소개한다. 이러한 관점은 global MAP 및 DPS와의 연결을 명확히 하며, optimization-based method에 대한 통합된 probabilistic interpretation을 제공한다. 이 토대 위에서 우리는 Gaussian prior assumption에서 동기를 얻은 covariance approximation과 안정성과 해석 가능성을 위한 reformulated objective를 갖는 실용적 algorithm을 개발한다. 광범위한 image restoration 및 scientific task 전반에서 LMAPS는 state-of-the-art 성능을 달성한다."},{"id":"61074","en":"Multi-Objective Bayesian Optimization via Adaptive $\\varepsilon$-Constraint Decomposition","ko":"Adaptive $\\varepsilon$-Constraint Decomposition을 통한 Multi-Objective Bayesian Optimization","authors":"Yaohong Yang, Sammie Katt, Samuel Kaski","pos":"#3606","link":"https://openreview.net/forum?id=tu7eFWXBiN","abs":"Multi-objective Bayesian optimization (MOBO) provides a principled framework for optimizing expensive black-box functions with multiple objectives. However, existing MOBO methods often struggle with coverage, scalability with respect to the number of objectives, and integrating constraints and preferences. In this work we propose STAGE-BO, Sequential Targeting Adaptive Gap-Filling $\\varepsilon$-Constraint Bayesian Optimization, that explicitly targets under-explored regions of the Pareto front. By analyzing the coverage of the approximate Pareto front, our method identifies the largest geometric gaps. These gaps are then used as constraints, which transforms the problem into a sequence of inequality-constrained subproblems, efficiently solved via constrained expected improvement acquisition. Our approach provides a uniform Pareto coverage without hypervolume computation and naturally applies to constrained, and preference-based settings. Experiments on synthetic and real-world benchmarks demonstrate superior coverage and competitive hypervolume performance against state-of-the-art baselines.","absKo":"Multi-objective Bayesian optimization(MOBO)은 여러 objective를 가진 비용이 큰 black-box function을 최적화하기 위한 원칙적인 framework를 제공한다. 그러나 기존 MOBO 방법은 Pareto front의 coverage, objective 수에 대한 scalability, constraint와 preference의 통합 측면에서 종종 어려움을 겪는다. 본 연구에서는 STAGE-BO, 즉 Sequential Targeting Adaptive Gap-Filling $\\\\varepsilon$-Constraint Bayesian Optimization을 제안하며, 이는 탐색이 부족한 Pareto front 영역을 명시적으로 목표로 삼는다. approximate Pareto front의 coverage를 분석함으로써, 우리의 방법은 가장 큰 geometric gap을 식별한다. 이 gap들은 이후 constraint로 사용되어, 문제를 일련의 inequality-constrained subproblem으로 변환하고, constrained expected improvement acquisition을 통해 효율적으로 해결된다. 우리의 접근법은 hypervolume computation 없이도 균일한 Pareto coverage를 제공하며, constrained setting과 preference-based setting에 자연스럽게 적용된다. synthetic benchmark와 real-world benchmark에서의 실험은 state-of-the-art baseline에 비해 더 우수한 coverage와 경쟁력 있는 hypervolume 성능을 입증한다."},{"id":"62773","en":"Predictive variational inference: Learn the predictively optimal posterior distribution","ko":"Predictive variational inference: 예측적으로 최적의 posterior 분포 학습","authors":"Jinlin Lai, Antonio Linero, Yuling Yao","pos":"#3607","link":"https://openreview.net/forum?id=dVIts0fNb4","abs":"

Vanilla variational inference finds an optimal approximation to the Bayesian posterior distribution, but even the exact Bayesian posterior is often not meaningful under model misspecification. We propose predictive variational inference (PVI): a general inference framework that seeks and samples from an optimal posterior density such that the resulting posterior predictive distribution is as close to the true data generating process as possible, while this closeness is measured by multiple scoring rules. By optimizing the objective, the predictive variational inference is generally not the same as, or even attempting to approximate, the Bayesian posterior, even asymptotically. Rather, we interpret it as implicit hierarchical expansion. Further, the learned posterior uncertainty detects heterogeneity of parameters among the population, enabling automatic model diagnosis. This framework applies to both likelihood-exact and likelihood-free models. We demonstrate its application in real data examples.

","absKo":"

Vanilla variational inference는 Bayesian posterior distribution에 대한 최적의 approximation을 찾지만, model misspecification 하에서는 정확한 Bayesian posterior조차 종종 의미가 없다. 우리는 predictive variational inference (PVI)를 제안한다. 이는 여러 scoring rule로 측정할 때 true data generating process에 가능한 한 가깝도록 posterior predictive distribution이 되게 하는 optimal posterior density를 찾고 그로부터 sample하는 일반적인 inference framework이다. objective를 최적화함으로써, predictive variational inference는 일반적으로 Bayesian posterior와 같지도 않고, 심지어 이를 근사하려는 시도도 아니며, asymptotically도 마찬가지이다. 오히려 우리는 이를 implicit hierarchical expansion으로 해석한다. 더 나아가, 학습된 posterior uncertainty는 population 내 parameter heterogeneity를 탐지하여 자동 model diagnosis를 가능하게 한다. 이 framework는 likelihood-exact 모델과 likelihood-free 모델 모두에 적용된다. 우리는 실제 data example에서의 적용을 보여준다.

"},{"id":"62746","en":"Richer Bayesian Last Layers with Subsampled NTK Features","ko":"하위 샘플링된 NTK Features를 활용한 강화된 Bayesian Last Layers","authors":"Sergio Calvo Ordoñez, Jonathan Plenk, Richard Bergna, Alvaro Cartea, Yarin Gal, Jose Miguel Hernandez-Lobato, Kamil Ciosek","pos":"#3608","link":"https://openreview.net/forum?id=djASMk0bzO","abs":"

Bayesian last layers (BLLs) provide a convenient and computationally efficient way to estimate uncertainty in neural networks. However, they underestimate epistemic uncertainty because they apply a Bayesian treatment only to the final layer, ignoring uncertainty induced by earlier layers. We propose a method that improves BLLs by leveraging a projection of Neural Tangent Kernel (NTK) features onto the space spanned by the last-layer features. This enables posterior inference that accounts for variability of the full network while retaining the low computational cost of a standard BLL. We show that our method yields posterior variances that are provably greater or equal to those of a standard BLL, correcting its tendency to underestimate epistemic uncertainty. To further reduce computational cost, we introduce a uniform subsampling scheme for estimating the projection matrix and for posterior inference. We derive approximation bounds for both types of sub-sampling. Empirical evaluations on UCI regression, contextual bandits, and out-of-distribution detection tasks in image and tabular datasets, demonstrate improved calibration and uncertainty estimates compared to standard BLLs and competitive baselines, while reducing computational cost.

","absKo":"Bayesian last layers (BLLs)는 neural network의 uncertainty를 추정하는 데 편리하고 계산 효율적인 방법을 제공한다. 그러나 Bayesian 처리를 마지막 layer에만 적용하고 앞선 layer들에서 유도되는 uncertainty는 무시하기 때문에 epistemic uncertainty를 과소추정한다. 우리는 Neural Tangent Kernel (NTK) feature를 last-layer feature가 span하는 공간에 projection하는 방법을 활용하여 BLL을 개선하는 방법을 제안한다. 이를 통해 표준 BLL의 낮은 계산 비용을 유지하면서도 전체 network의 변동성을 반영하는 posterior inference가 가능해진다. 우리는 이 방법이 표준 BLL보다 provably greater or equal한 posterior variance를 산출함을 보여주며, epistemic uncertainty를 과소추정하는 경향을 교정한다. 계산 비용을 더 줄이기 위해, 우리는 projection matrix 추정과 posterior inference를 위한 uniform subsampling scheme을 도입한다. 또한 두 종류의 sub-sampling에 대한 approximation bound를 유도한다. UCI regression, contextual bandit, 그리고 image 및 tabular dataset에서의 out-of-distribution detection task에 대한 실증 평가에서, 우리의 방법은 표준 BLL과 경쟁 baseline에 비해 calibration과 uncertainty estimate를 향상시키면서 계산 비용을 줄임을 보여준다.

"},{"id":"61793","en":"Robust Stochastic Gradient Posterior Sampling with Lattice Based Discretisation","ko":"Lattice 기반 이산화를 활용한 Robust Stochastic Gradient Posterior Sampling","authors":"Zier Mensch, Lars Holdijk, Samuel Duffield, Maxwell Aifer, Patrick Coles, Max Welling, Miranda C. N. Cheng","pos":"#3609","link":"https://openreview.net/forum?id=mtRgrp0u7Z","abs":"

Stochastic-gradient MCMC methods enable scalable Bayesian posterior sampling but often suffer from sensitivity to minibatch size and gradient noise. To address this, we propose Stochastic Gradient Lattice Random Walk (SGLRW), an extension of the Lattice Random Walk discretization. Unlike conventional Stochastic Gradient Langevin Dynamics (SGLD), SGLRW introduces stochastic noise only through the off-diagonal elements of the update covariance; this yields greater robustness to minibatch size while retaining asymptotic correctness. Furthermore, as comparison we analyze a natural analogue of SGLD utilizing gradient clipping. Experimental validation on Bayesian regression and classification demonstrates that SGLRW remains stable in regimes where SGLD fails, including in the presence of heavy-tailed gradient noise, and matches or improves predictive performance.

","absKo":"

Stochastic-gradient MCMC method는 scalable Bayesian posterior sampling을 가능하게 하지만, minibatch size와 gradient noise에 민감하다는 문제를 자주 겪는다. 이를 해결하기 위해 우리는 Lattice Random Walk discretization의 확장인 Stochastic Gradient Lattice Random Walk (SGLRW)를 제안한다. 기존의 Stochastic Gradient Langevin Dynamics (SGLD)와 달리, SGLRW는 stochastic noise를 update covariance의 off-diagonal element를 통해서만 도입한다. 이는 asymptotic correctness를 유지하면서 minibatch size에 대한 더 강한 robustness를 제공한다. 또한 비교 대상으로 gradient clipping을 활용한 SGLD의 자연스러운 analogue도 분석한다. Bayesian regression과 classification에서의 실험적 검증은 SGLRW가 heavy-tailed gradient noise가 존재하는 경우를 포함해 SGLD가 실패하는 regime에서도 안정적으로 동작하며, predictive performance를 맞추거나 개선함을 보여준다.

"},{"id":"64819","en":"SURGE:Unbiased Data Assimilation for Diffusion Model via Particle Filtering","ko":"SURGE: Particle Filtering을 통한 Diffusion Model을 위한 편향 없는 Data Assimilation","authors":"Lifu Wei, Yinuo Ren, Naichen Shi, Yiping Lu","pos":"#3610","link":"https://openreview.net/forum?id=JRsJLPr85W","abs":"

Data assimilation (DA) addresses the problem of sequentially estimating the state of a dynamical system from noisy and incomplete observations. In this work, we employ a diffusion model as a world model to simulate and predict the system’s dynamics. Recently, score-based diffusion models have learned global diffusion priors that effectively model (stochastic) dynamics, revealing strong potential for data assimilation. In this paper, we investigate how information from noisy observations can be incorporated to enable continuous correction and refinement of the predicted system state when using a diffusion prior. Motivated by particle filtering methods, we represent the posterior distribution using a set of particles. After receiving noisy observations, the diffusion model is guided using the observation likelihood to steer the generation process toward observation-consistent states. Nevertheless, such guidance does not guarantee sampling from the true posterior. We therefore employ a Sequential Monte Carlo approach over the diffusion trajectory, viewed as a path measure, to reweight and resample particles, thereby correcting the generation process and ensuring convergence toward the desired posterior distribution. This leads to an unbiased particle filtering method that rigorously fuses observational data with diffusion model simulations.

","absKo":"

Data assimilation(DA)은 noisy하고 incomplete한 observation으로부터 dynamical system의 state를 순차적으로 추정하는 문제를 다룬다. 본 연구에서는 diffusion model을 world model로 사용하여 시스템의 dynamics를 시뮬레이션하고 예측한다. 최근 score-based diffusion model은 (stochastic) dynamics를 효과적으로 모델링하는 global diffusion prior를 학습해 왔으며, data assimilation에 대한 강한 잠재력을 보여 주었다. 이 논문에서는 diffusion prior를 사용할 때 noisy observation으로부터의 정보를 어떻게 통합하여 예측된 system state를 지속적으로 보정하고 정교화할 수 있는지 살펴본다. particle filtering method에 착안하여, posterior distribution을 particle 집합으로 표현한다. noisy observation을 받은 뒤에는 diffusion model이 observation likelihood를 이용해 생성 과정을 observation과 일치하는 상태로 유도한다. 그러나 이러한 guidance만으로는 true posterior로부터의 sampling이 보장되지 않는다. 따라서 우리는 diffusion trajectory를 path measure로 보고 그 위에 Sequential Monte Carlo 접근을 적용하여 particle을 reweight하고 resample함으로써 생성 과정을 교정하고 원하는 posterior distribution으로의 수렴을 보장한다. 이는 observational data와 diffusion model simulation을 엄밀하게 결합하는 unbiased particle filtering method로 이어진다.

"},{"id":"66481","en":"Sparse Bayesian Deep Functional Learning with Structured Region Selection","ko":"구조화된 영역 선택을 갖는 희소 Bayesian 심층 함수 학습","authors":"Xiaoxian Zhu, Yingmeng Li, Shuangge Ma, Mengyun Wu","pos":"#3611","link":"https://openreview.net/forum?id=3IFIedDIoN","abs":"

In modern applications such as ECG monitoring, neuroimaging, wearable sensing, and industrial equipment diagnostics, complex and continuously structured data are ubiquitous, presenting both challenges and opportunities for functional data analysis. However, existing methods face a critical trade-off: conventional functional models are limited by linearity, whereas deep learning approaches lack interpretable region selection for sparse effects. To bridge these gaps, we propose a sparse Bayesian functional deep neural network (sBayFDNN). It learns adaptive functional embeddings through a deep Bayesian architecture to capture complex nonlinear relationships, while a structured prior enables interpretable, region-wise selection of influential domains with quantified uncertainty. Theoretically, we establish rigorous approximation error bounds, posterior consistency, and region selection consistency. These results provide the first theoretical guarantees for a Bayesian deep functional model, ensuring its reliability and statistical rigor. Empirically, comprehensive simulations and real‑world studies confirm the effectiveness and superiority of sBayFDNN. Crucially, sBayFDNN excels in recognizing intricate dependencies for accurate predictions and more precisely identifies functionally meaningful regions, capabilities fundamentally beyond existing approaches.

","absKo":"

ECG monitoring, neuroimaging, wearable sensing, 산업 장비 진단과 같은 현대 응용에서는 복잡하고 연속적인 구조를 가진 데이터가 어디에나 존재하며, functional data analysis에 도전과 기회를 동시에 제공한다. 그러나 기존 방법은 중요한 trade-off를 보인다. 전통적인 functional model은 linearity에 의해 제약되고, deep learning 접근법은 sparse effect에 대한 해석 가능한 region selection이 부족하다. 이러한 간극을 메우기 위해 우리는 sparse Bayesian functional deep neural network(sBayFDNN)를 제안한다. 이 모델은 deep Bayesian architecture를 통해 adaptive functional embedding을 학습하여 복잡한 비선형 관계를 포착하는 한편, structured prior를 통해 영향력 있는 domain을 정량화된 uncertainty와 함께 해석 가능한 방식으로 region-wise selection할 수 있게 한다. 이론적으로 우리는 엄밀한 approximation error bound, posterior consistency, region selection consistency를 확립한다. 이 결과는 Bayesian deep functional model에 대한 첫 이론적 보장을 제공하며, 그 신뢰성과 statistical rigor를 보장한다. 실험적으로는 포괄적인 시뮬레이션과 실제 연구가 sBayFDNN의 효과성과 우수성을 확인한다. 특히 sBayFDNN은 정확한 예측을 위해 복잡한 의존성을 인식하는 데 뛰어나며, 기능적으로 의미 있는 region을 더 정밀하게 식별하는데, 이러한 능력은 기존 접근법의 본질적 한계를 넘는다.

"},{"id":"61741","en":"TabMGP: Martingale Posterior with TabPFN","ko":"TabMGP: TabPFN을 활용한 Martingale Posterior","authors":"Kenyon Ng, Edwin Fong, David Frazier, Jeremias Knoblauch, Susan Wei","pos":"#3612","link":"https://openreview.net/forum?id=nJ3yX2ZnOJ","abs":"Bayesian inference provides principled uncertainty quantification but is often limited by challenges of prior and likelihood elicitation. The martingale posterior (MGP) (Fong et al, 2023) offers an alternative by replacing these requirements with a predictive rule. Additionally MGP focuses inference on parameters defined through a loss function. This framework is especially resonant in the era of foundation transformers; practitioners increasingly leverage models like TabPFN for their state-of-the-art capabilities, yet often require epistemic uncertainty for a scientific estimand $\\theta$ that need not parametrize the model’s implicit latent model. The MGP provides the mechanism to recover these posterior distributions. We introduce TabMGP, an MGP built on TabPFN for tabular data. TabMGP produces credible sets with near-nominal coverage and often outperforms both handcrafted MGP constructions and standard Bayesian baselines.","absKo":"Bayesian inference는 원리적인 uncertainty quantification을 제공하지만, prior와 likelihood elicitation의 어려움 때문에 종종 제약을 받는다. Martingale posterior(MGP) (Fong et al, 2023)은 이러한 요구를 predictive rule로 대체함으로써 대안을 제공한다. 또한 MGP는 loss function을 통해 정의된 parameter에 inference를 집중한다. 이 framework는 foundation transformer의 시대에 특히 잘 맞는다. 실무자들은 TabPFN과 같은 model을 최첨단 성능 때문에 점점 더 활용하지만, model의 implicit latent model을 parameterize할 필요는 없는 과학적 estimand $\\theta$에 대한 epistemic uncertainty를 종종 요구한다. MGP는 이러한 posterior distribution을 복원하는 메커니즘을 제공한다. 우리는 tabular data를 위한 TabPFN 기반 MGP인 TabMGP를 소개한다. TabMGP는 near-nominal coverage를 갖는 credible set을 생성하며, 종종 직접 설계한 MGP 구성과 표준 Bayesian baseline 모두를 능가한다."},{"id":"62764","en":"Tokenised Flow Matching for Hierarchical Simulation Based Inference","ko":"계층적 Simulation Based Inference를 위한 Tokenised Flow Matching","authors":"Giovanni Charles, Cosmo Santoni, Seth Flaxman, Elizaveta Semenova","pos":"#3613","link":"https://openreview.net/forum?id=dblO6FgIvc","abs":"

The cost of simulator evaluations is a key practical bottleneck for Simulation Based Inference (SBI). In hierarchical settings with shared global parameters and exchangeable site-level parameters and observations, this structure can be exploited to improve simulation efficiency. Existing hierarchical SBI approaches factorise the posterior yet still simulate across multiple sites per training sample; We instead explore likelihood factorisation (LF) to train from single-site simulations. In LF sampling we learn a per-site (factorised) likelihood and then assemble synthetic multi-site observations to amortise inference for the full hierarchical posterior. Building on this, we propose Tokenised Flow Matching for Posterior Estimation (TFMPE), a tokenised flow matching approach that supports function-valued observations through likelihood factorisation. To enable systematic evaluation, we introduce a benchmark for hierarchical SBI. We validate TFMPE on this benchmark and on realistic infectious disease and computational fluid dynamics models, finding well-calibrated posteriors while reducing computational cost.

","absKo":"

시뮬레이터 평가 비용은 Simulation Based Inference(SBI)의 핵심적인 실용적 병목이다. shared global parameters와 exchangeable site-level parameters 및 observations를 갖는 hierarchical setting에서는 이러한 구조를 활용해 simulation 효율을 높일 수 있다. 기존의 hierarchical SBI 방법들은 posterior를 factorize하지만, 여전히 training sample마다 여러 site에 대해 simulation을 수행한다. 우리는 대신 single-site simulation으로부터 학습하는 likelihood factorization(LF)을 탐구한다. LF sampling에서는 site별(factorized) likelihood를 학습한 뒤 synthetic multi-site observations를 조립하여 full hierarchical posterior에 대한 inference를 amortize한다. 이를 바탕으로, 우리는 likelihood factorization을 통해 function-valued observations를 지원하는 tokenized flow matching 방식인 Tokenised Flow Matching for Posterior Estimation(TFMPE)을 제안한다. 체계적인 평가를 위해 hierarchical SBI용 benchmark도 도입한다. 우리는 이 benchmark와 현실적인 infectious disease 및 computational fluid dynamics model에서 TFMPE를 검증했으며, 계산 비용을 줄이면서도 잘 보정된 posterior를 확인했다.

"},{"id":"64431","en":"Efficient Bayesian Inference from Noisy Pairwise Comparisons","ko":"노이즈가 있는 쌍별 비교로부터의 효율적 베이지안 추론","authors":"Till Aczel, Lucas Theis, Roger Wattenhofer","pos":"#4104","link":"https://openreview.net/forum?id=NinueNAODD","abs":"

Evaluating generative models is challenging because standard metrics often fail to reflect human preferences. Human evaluations are more reliable but costly and noisy, as participants vary in expertise, attention, and diligence. Pairwise comparisons improve consistency, yet aggregating them into overall quality scores requires careful modeling. Bradley-Terry-based methods update item scores from comparisons, but existing approaches either ignore rater variability or lack convergence guarantees, limiting robustness and interpretability. We introduce BBQ, a Bayesian Bradley-Terry variant that explicitly models rater quality, downweighting or removing unreliable participants, and provides guaranteed monotonic likelihood convergence through an Expectation-Maximization algorithm. Empirical results show that BBQ achieves faster convergence, well-calibrated uncertainty estimates, and more robust, interpretable rankings compared to baseline Bradley-Terry models, even with noisy or crowdsourced raters. This framework enables more reliable and cost-effective human evaluation of generative models.

","absKo":"생성 모델을 평가하는 것은 표준 metric이 인간의 선호를 제대로 반영하지 못하는 경우가 많기 때문에 어렵다. 인간 평가는 더 신뢰할 수 있지만, 참여자마다 전문성, 주의력, 성실성이 달라 비용이 많이 들고 노이즈도 크다. pairwise comparison은 일관성을 높이지만, 이를 전체 품질 점수로 통합하려면 신중한 modeling이 필요하다. Bradley-Terry 기반 방법은 comparison으로부터 item score를 갱신하지만, 기존 접근법은 rater variability를 무시하거나 수렴 보장이 없어 robustness와 interpretability가 제한된다. 우리는 rater quality를 명시적으로 modeling하여 신뢰할 수 없는 참여자를 downweight하거나 제거하고, Expectation-Maximization algorithm을 통해 monotonic likelihood convergence를 보장하는 Bayesian Bradley-Terry 변형인 BBQ를 제안한다. 실험 결과, BBQ는 노이즈가 있거나 crowdsourced rater가 있는 경우에도 baseline Bradley-Terry model보다 더 빠른 수렴, 잘 보정된 uncertainty estimate, 더 견고하고 해석 가능한 ranking을 달성한다. 이 framework는 생성 모델에 대한 보다 신뢰 가능하고 비용 효율적인 인간 평가를 가능하게 한다.

"},{"id":"60984","en":"An Evidential Route to Asymptotic Bayes Optimality under Sparsity","ko":"Sparsity 하에서 Asymptotic Bayes Optimality로 가는 Evidential Route","authors":"Qiaoyu Liang, Haohua Chen, Zihan Zhu, Michael Evans","pos":"#4412","link":"https://openreview.net/forum?id=urcZXdxMEm","abs":"

From a statistical evidence perspective, we establish some asymptotic optimality properties of certain multiple testing rules based on the relative belief ratio (Evans, 2015). Under the two-groups model with an additive 0-1 loss and within a Bayesian decision theoretic asymptotic framework of Bogdan et al. (2011), we show that relative belief multiple testing rules induced by a simple one-group light-tailed normal prior with a single hyperparameter achieve the same asymptotic Bayes risk as the Bayes oracle benchmark. This risk is the minimum achievable in this asymptotic framework. Despite originating from a different starting point, the evidential relative belief approach enjoys oracle properties. The relative belief multiple testing approach is fundamentally different from existing Bayesian multiple testing procedures, virtually all induced by more complex heavy-tailed one-group global-local shrinkage priors using purely posterior-based inferences (Datta & Ghosh, 2013; Ghosh et al., 2016; Bhadra et al., 2017; Ghosh & Chakrabarti, 2017; Qin & Ghosh, 2025). By measuring statistical evidence via both the prior and posterior, the relative belief approach reveals an alternative new inferential paradigm for attaining asymptotic Bayes optimality under sparsity, one that does not rely on developing increasingly elaborate priors.

","absKo":"

statistical evidence의 관점에서, 우리는 relative belief ratio(Evans, 2015)에 기반한 특정 multiple testing rule의 asymptotic optimality 성질을 확립한다. additive 0-1 loss를 갖는 two-groups model과 Bogdan et al. (2011)의 Bayesian decision theoretic asymptotic framework 하에서, 단일 hyperparameter를 갖는 간단한 one-group light-tailed normal prior에 의해 유도된 relative belief multiple testing rule이 Bayes oracle benchmark와 동일한 asymptotic Bayes risk를 달성함을 보인다. 이 risk는 이 asymptotic framework에서 달성 가능한 최소값이다. 서로 다른 출발점에서 비롯되었음에도 불구하고, evidential relative belief approach는 oracle 성질을 지닌다. relative belief multiple testing approach는 기존의 Bayesian multiple testing procedure와 근본적으로 다르며, 사실상 이들 대부분은 더 복잡한 heavy-tailed one-group global-local shrinkage prior에 의해 유도되고 오로지 posterior-based inference만을 사용하는 방식이다(Datta & Ghosh, 2013; Ghosh et al., 2016; Bhadra et al., 2017; Ghosh & Chakrabarti, 2017; Qin & Ghosh, 2025). prior와 posterior 모두를 통해 statistical evidence를 측정함으로써, relative belief approach는 sparsity 하에서 asymptotic Bayes optimality를 달성하기 위한 대안적인 새로운 inferential paradigm을 드러내며, 이는 점점 더 정교한 prior를 개발하는 데 의존하지 않는다.

"},{"id":"68827","en":"Efficient and Unbiased Sampling from Boltzmann Distributions via Variance-Tuned Diffusion Models","ko":"Variance-Tuned Diffusion 모델을 통한 Boltzmann 분포로부터의 효율적이고 불편향된 샘플링","authors":"Fengzhe Zhang, Laurence Midgley, Jose Miguel Hernandez-Lobato","pos":"#1508","abs":"Score-based diffusion models (SBDMs) are powerful amortized samplers for Boltzmann distributions; however, imperfect score estimates bias downstream Monte Carlo estimates. Classical importance sampling (IS) can correct this bias, but computing exact likelihoods requires solving the probability-flow ordinary differential equation (PF–ODE), a procedure that is prohibitively costly and scales poorly with dimensionality. We introduce Variance-Tuned Diffusion Importance Sampling (VT-DIS), a lightweight post-training method that adapts the per-step noise covariance of a pretrained SBDM by minimizing the $\\alpha$-divergence $(\\alpha=2)$ between its forward diffusion and reverse denoising trajectories. VT-DIS assigns a single trajectory-wise importance weight to the joint forward–reverse process, yielding unbiased expectation estimates at test time with negligible inference-time overhead compared to standard sampling. On the DW-4, LJ-13, and alanine-dipeptide benchmarks, VT-DIS achieves effective sample sizes of approximately 80%, 35%, and 3.5%, respectively, while using only a fraction of the computational budget required by vanilla diffusion + IS or PF-ODE–based IS.","absKo":"score-based diffusion model(SBDM)은 Boltzmann distribution을 위한 강력한 amortized sampler이지만, 완벽하지 않은 score estimate는 downstream Monte Carlo estimate에 bias를 유발합니다. 고전적인 importance sampling(IS)은 이러한 bias를 보정할 수 있지만, 정확한 likelihood를 계산하려면 probability-flow ordinary differential equation(PF–ODE)을 풀어야 하며, 이는 계산 비용이 지나치게 크고 차원 증가에 따라 확장성이 좋지 않습니다. 우리는 Variance-Tuned Diffusion Importance Sampling(VT-DIS)을 소개합니다. 이는 사전 학습된 SBDM의 step별 noise covariance를 조정하는 경량 post-training 방법으로, forward diffusion과 reverse denoising trajectory 사이의 $\\alpha$-divergence($(\\alpha=2)$)를 최소화합니다. VT-DIS는 joint forward–reverse process에 대해 하나의 trajectory-wise importance weight를 부여하여, 표준 sampling과 비교했을 때 추론 시 오버헤드는 거의 증가시키지 않으면서 test time에서 unbiased expectation estimate를 제공합니다. DW-4, LJ-13, alanine-dipeptide benchmark에서 VT-DIS는 각각 약 80%, 35%, 3.5%의 effective sample size를 달성하며, vanilla diffusion + IS 또는 PF-ODE 기반 IS에 필요한 계산 예산의 일부만 사용합니다.

"},{"id":"64104","en":"HypoSpace: A Diagnostic Benchmark for Set-Valued Hypothesis Generation under Underdetermination and Sublinear Coverage Bounds","ko":"HypoSpace: 미결정성 및 준선형 커버리지 한계 하에서 집합값 가설 생성을 위한 진단 벤치마크","authors":"Tingting Chen, Beibei Lin, Zifeng Yuan, Qiran Zou, Hongyu He, Anirudh Goyal, Yew Soon ONG, Dianbo Liu","pos":"#2302","link":"https://openreview.net/forum?id=QpjtK65JHO","abs":"

Many scientific problems are underdetermined: multiple distinct hypotheses are equally consistent with the same observations. In such settings, effective inference requires not only producing valid explanations, but also systematically exploring and covering the admissible hypothesis set. We introduce HypoSpace, a benchmark that treats large language models (LLMs) as samplers over finite hypothesis spaces and evaluates them on three metrics: Validity, Uniqueness, and Recovery. HypoSpace spans three structured domains (causal graph inference, gravity-constrained 3D voxel reconstruction, and Boolean genetic interaction modeling) with deterministic validators and exactly enumerable solution spaces, plus real-world anchored case studies. Empirically, frontier LLMs exhibit a consistent failure mode: high Validity but sharp degradation in Uniqueness and Recovery as hypothesis spaces grow. We further show that stratified decoding partially mitigates this collapse, demonstrating HypoSpace's utility as a diagnostic benchmark for set-valued inference.

","absKo":"

많은 scientific problem은 underdetermined되어 있다. 즉, 여러 서로 다른 hypothesis가 동일한 observation과 모두 일치한다. 이러한 설정에서는 효과적인 추론이 유효한 explanation을 생성하는 것뿐 아니라, 허용 가능한 hypothesis 집합을 체계적으로 탐색하고 포괄하는 것까지 요구한다. 우리는 대규모 language model (LLM)을 finite hypothesis space 위의 sampler로 취급하고, Validity, Uniqueness, Recovery의 세 가지 metric으로 평가하는 benchmark인 HypoSpace를 소개한다. HypoSpace는 세 개의 structured domain(causal graph inference, gravity-constrained 3D voxel reconstruction, Boolean genetic interaction modeling)을 포함하며, deterministic validator와 정확히 열거 가능한 solution space, 그리고 실제 세계에 고정된 case study를 갖춘다. 실증적으로 frontier LLM은 일관된 실패 양상을 보이는데, hypothesis space가 커질수록 Validity는 높지만 Uniqueness와 Recovery는 급격히 저하된다. 우리는 또한 stratified decoding이 이러한 붕괴를 부분적으로 완화함을 보이며, set-valued inference를 위한 진단 benchmark로서 HypoSpace의 유용성을 입증한다.

"},{"id":"60507","en":"Initialization is Half the Battle: Generating Diverse Images from a Guidance Potential Posterior","ko":"Initialization이 절반이다: Guidance Potential Posterior에서 다양한 이미지 생성","authors":"Xiang Li, Dianbo Liu, Kenji Kawaguchi","pos":"#2804","link":"https://openreview.net/forum?id=zXJjT0slV7","abs":"

Despite the remarkable fidelity of generative models, they frequently suffer from mode collapse. Existing strategies for enhancing diversity predominantly focus on intervening during the generation trajectory. We identify a critical oversight that the standard Gaussian initialization often causes trajectories to collapse into dominant modes because it is agnostic to the guidance potential landscape. In this work, we formulate selecting the initial noise from a guidance potential posterior, which effectively re-weights the prior towards diversity-rich regions. To sample from this distribution efficiently, we introduce Diversity-inducing Initialization (DivIn), which leverages Langevin dynamics to actively navigate the initialization landscape, steering initial noise away from collapsing regions while anchoring them to the valid data manifold. Our method serves as an inference-time diversity enhancement compatible with both diffusion and flow matching models. Extensive experiments show that DivIn exhibits a superior performance in both class-to-image and text-to-image scenarios. Furthermore, we highlight that as DivIn is orthogonal to trajectory-based methods, combining them significantly expands the diversity-quality Pareto frontier beyond what either achieves in isolation.

","absKo":"

생성 모델의 놀라운 fidelity에도 불구하고, 이들은 자주 mode collapse를 겪는다. 다양성 향상을 위한 기존 전략은 주로 generation trajectory 중 개입하는 데 초점을 맞춘다. 우리는 표준 Gaussian initialization이 guidance potential landscape에 무관심하기 때문에 trajectory가 지배적인 mode로 붕괴되는 중요한 간과점을 식별한다. 본 연구에서는 guidance potential posterior로부터 초기 noise를 선택하는 문제를 정식화하며, 이는 prior를 다양성이 풍부한 영역 쪽으로 효과적으로 재가중한다. 이 분포를 효율적으로 샘플링하기 위해, 우리는 Langevin dynamics를 활용하여 initialization landscape를 능동적으로 탐색하고, 초기 noise가 붕괴 영역에서 멀어지도록 유도하는 동시에 유효한 data manifold에 고정시키는 Diversity-inducing Initialization(DivIn)을 제안한다. 우리의 방법은 diffusion 및 flow matching model 모두와 호환되는 inference-time diversity enhancement로 작동한다. 광범위한 실험은 DivIn이 class-to-image와 text-to-image 시나리오 모두에서 우수한 성능을 보임을 보여준다. 또한 DivIn이 trajectory-based method와 직교적이므로, 이를 결합하면 어느 한 방법만으로 달성하는 것보다 diversity-quality Pareto frontier를 크게 확장할 수 있음을 강조한다.

"},{"id":"60561","en":"Mitigating the Contractivity Trap in Diffusion ODEs via Stein Stabilization","ko":"Stein 안정화를 통한 Diffusion ODE의 수축성 함정 완화","authors":"Shigui Li, Delu Zeng","pos":"#2810","link":"https://openreview.net/forum?id=z0CHTAHame","abs":"

A fundamental tension exists in the large-step inference of diffusion models via their deterministic probability flow ordinary differential equation (PF-ODE) paths, which we formally identify as the contractivity trap: efficient inference favors large step sizes, while stable convergence requires strong contractivity that limits expressiveness. To address this, we propose SteinDiff, an inference-time stabilization framework based on reference-free Stein corrections. Specifically, SteinDiff introduces a geometry-aware correction mechanism that stabilizes PF-ODE inference trajectories. To this end, we contribute closed-form correction estimators via Stein's identity in the continuous-time setting, enabling the method to adapt to local data geometry. We theoretically demonstrate that SteinDiff reduces integration error even when contractivity is violated and establishes its robustness against discretization-induced distributional shifts. Our analysis further reveals that these corrections act as persistent geometric anchors, providing new insights into the stability of SOTA EDM parameterizations. Extensive experiments demonstrate that SteinDiff significantly mitigates mode collapse and improves generative quality in large-step inference.

","absKo":"

Diffusion model의 deterministic probability flow ordinary differential equation(PF-ODE) 경로를 통한 large-step inference에는 근본적인 긴장이 존재하며, 우리는 이를 형식적으로 contractivity trap이라고 명명한다. 효율적인 inference는 큰 step size를 선호하는 반면, 안정적인 수렴은 표현력을 제한하는 강한 contractivity를 요구한다. 이를 해결하기 위해 우리는 reference-free Stein correction에 기반한 inference-time stabilization framework인 SteinDiff를 제안한다. 구체적으로 SteinDiff는 PF-ODE inference trajectory를 안정화하는 geometry-aware correction mechanism을 도입한다. 이를 위해 우리는 continuous-time 설정에서 Stein's identity를 이용한 closed-form correction estimator를 제공하여, method가 local data geometry에 적응할 수 있게 한다. 이론적으로 우리는 SteinDiff가 contractivity가 위배되더라도 integration error를 줄이며, discretization-induced distributional shift에 대해 강건함을 보인다는 점을 입증한다. 또한 우리의 분석은 이러한 correction이 지속적인 geometric anchor로 작용함을 보여주며, SOTA EDM parameterization의 안정성에 대한 새로운 통찰을 제공한다. 광범위한 실험은 SteinDiff가 large-step inference에서 mode collapse를 유의미하게 완화하고 generative quality를 향상시킴을 입증한다.

"},{"id":"68828","en":"Demystifying MaskGIT Sampler and Beyond: Adaptive Order Selection in Masked Diffusion","ko":"MaskGIT Sampler 그리고 그 너머 규명: Masked Diffusion에서의 적응적 순서 선택","authors":"Satoshi Hayakawa, Yuhta Takida, Masaaki Imaizumi, Hiromi Wakaki, Yuki Mitsufuji","pos":"#2903","abs":"

Masked diffusion models have shown promising performance in generating high-quality samples in a wide range of domains, but accelerating their sampling process remains relatively underexplored. To investigate efficient samplers for masked diffusion, this paper theoretically analyzes the MaskGIT sampler for image modeling, revealing its implicit temperature sampling mechanism. Through this analysis, we show that MaskGIT is asymptotically equivalent to a choose-then-sample (CTS) formulation, instantiated as the “moment sampler,” which explicitly separates index selection from token sampling. This CTS reformulation is essential: it yields unbiased token sampling and exposes an algorithmic design space for index selection, both of which are inaccessible in MaskGIT’s original formulation. Regarding token sampling, we reveal that MaskGIT implicitly adopts a low-temperature sampler, which explains why MaskGIT often degrades with more sampling steps. The CTS reformulation of MaskGIT allows to fix the temperature sampling to ensure unbiasedness. We also improve the index selection in CTS through two key innovations: a partial caching technique for transformers that approximates longer sampling trajectories without proportional computational cost, and a hybrid approach formalizing the exploration-exploitation trade-off in adaptive unmasking. Experiments in image and text domains demonstrate our theory as well as the efficiency of our proposed methods, advancing both theoretical understanding and practical implementation of masked diffusion samplers.

","absKo":"

Masked diffusion model은 다양한 도메인에서 고품질 sample 생성을 위해 유망한 성능을 보여왔지만, sampling 과정을 가속화하는 문제는 상대적으로 덜 탐구되었습니다. masked diffusion을 위한 효율적인 sampler를 조사하기 위해, 본 논문은 image modeling을 위한 MaskGIT sampler를 이론적으로 분석하여 그 암묵적인 temperature sampling 메커니즘을 밝혀냅니다. 이 분석을 통해 우리는 MaskGIT이 choose-then-sample(CTS) formulation과 점근적으로 동등함을 보이며, 이를 index selection과 token sampling을 명시적으로 분리하는 “moment sampler”로 구현합니다. 이 CTS reformulation은 필수적입니다. 이는 unbiased token sampling을 가능하게 하고, index selection을 위한 algorithmic design space를 드러내기 때문이며, 이 두 가지는 MaskGIT의 원래 formulation에서는 접근할 수 없습니다. token sampling에 대해서는, MaskGIT이 암묵적으로 low-temperature sampler를 채택하고 있음을 밝히며, 이것이 왜 MaskGIT이 sampling step이 많아질수록 성능이 종종 저하되는지를 설명합니다. MaskGIT의 CTS reformulation은 unbiasedness를 보장하기 위해 temperature sampling을 수정할 수 있게 합니다. 우리는 또한 두 가지 핵심 혁신을 통해 CTS의 index selection을 개선합니다. 하나는 더 긴 sampling trajectory를 비례적인 계산 비용 없이 근사하는 transformer용 partial caching technique이고, 다른 하나는 adaptive unmasking에서 exploration-exploitation trade-off를 형식화한 hybrid approach입니다. image와 text 도메인에서의 실험은 우리의 이론과 제안 방법의 효율성을 입증하며, masked diffusion sampler에 대한 이론적 이해와 실용적 구현을 모두 발전시킵니다.

"},{"id":"60879","en":"Path-dependent Discrete Amortized Inference","ko":"경로 의존적 Discrete Amortized Inference","authors":"Tiago Silva, Esmeralda S. Whitammer, Salem Lahlou","pos":"#3200","link":"https://openreview.net/forum?id=viJvYPJHIj","abs":"

We consider the problem of sampling compositional and discrete objects from a given unnormalized posterior distribution. Notably, recent studies have shown that this problem can be efficiently solved by learning a deterministic Markov Decision Process (MDP) that progressively builds each object in proportion to the posterior. In this work, however, we demonstrate that the Markovian assumption can both hamper signal propagation during training and catastrophically reduce the learned sampler's expressivity due to state aliasing. To address these issues, we propose lifting the MDP with a learnable latent dynamics that allows the underlying policy to depend on the entire past trajectory---and not only on the current state. In view of this, we refer to the resulting method as \\emph{path-dependent discrete amortized inference}. Importantly, we provably extend existing learning algorithms for amortized samplers to our setting. In experiments on standard benchmark problems, we also show that our approach often leads to faster learning convergence and improved state space exploration relatively to prior techniques.

","absKo":"

우리는 주어진 unnormalized posterior distribution으로부터 compositional하고 discrete한 object를 sampling하는 문제를 다룬다. 특히 최근 연구들은 posterior에 비례하여 각 object를 점진적으로 구성하는 deterministic Markov Decision Process (MDP)를 학습함으로써 이 문제를 효율적으로 해결할 수 있음을 보였다. 그러나 본 연구에서는 Markovian 가정이 학습 중 signal propagation을 저해할 수 있으며, state aliasing 때문에 학습된 sampler의 expressivity를 치명적으로 낮출 수 있음을 보인다. 이러한 문제를 해결하기 위해, 우리는 underlying policy가 현재 state뿐 아니라 전체 past trajectory에 의존할 수 있도록 하는 learnable latent dynamics를 통해 MDP를 확장한다. 이러한 관점에서, 우리는 이 결과 방법을 \\emph{path-dependent discrete amortized inference}라고 부른다. 중요한 점은, 기존 amortized sampler의 학습 알고리즘들을 우리의 설정으로 이론적으로 확장할 수 있음을 보인다는 것이다. 표준 benchmark 문제들에 대한 실험에서, 우리의 방법은 prior technique에 비해 더 빠른 learning convergence와 향상된 state space exploration으로 이어지는 경우가 많음을 보였다.

"},{"id":"64233","en":"Efficient Online Variational Estimation via Monte Carlo Sampling","ko":"몬테카를로 샘플링을 통한 효율적 온라인 변분 추정","authors":"Mathis Chagneux, Mathias Müller, Pierre Gloaguen, Sylvain Le Corff, Jimmy Olsson","pos":"#3300","link":"https://openreview.net/forum?id=Ph2BaszK9w","abs":"

This article addresses online variational estimation in parametric state-space models. We propose a new procedure for efficiently computing the evidence lower bound and its gradient in a streaming-data setting, where observations arrive sequentially. The algorithm allows for the simultaneous training of the model parameters and the distribution of the latent states given the observations. It is based on i.i.d. Monte Carlo sampling, coupled with a well-chosen deep architecture, enabling both computational efficiency and flexibility. The performance of the method is illustrated on both synthetic data and real-world air-quality data. The proposed approach is theoretically motivated by the existence of an asymptotic contrast function and the ergodicity of the underlying Markov chain, and applies more generally to the computation of additive expectations under posterior distributions in state-space models.

","absKo":"

이 논문은 parametric state-space model에서의 online variational estimation을 다룬다. 우리는 관측이 순차적으로 도착하는 streaming-data setting에서 evidence lower bound와 그 gradient를 효율적으로 계산하는 새로운 절차를 제안한다. 이 알고리즘은 model parameter와 관측이 주어졌을 때의 latent state distribution을 동시에 학습할 수 있게 한다. 이는 잘 선택된 deep architecture와 결합된 i.i.d. Monte Carlo sampling에 기반하며, 계산 효율성과 유연성을 모두 제공한다. 방법의 성능은 synthetic data와 실제 air-quality data 모두에서 입증된다. 제안한 접근법은 asymptotic contrast function의 존재와 underlying Markov chain의 ergodicity에 의해 이론적으로 정당화되며, state-space model에서 posterior distribution 하의 additive expectation 계산에도 더 일반적으로 적용된다.

"},{"id":"63673","en":"Variance-Reduced Zeroth-Order Langevin Dynamics for Non-Log-Concave Black-Box Sampling and Inverse Problems","ko":"비로그-오목 black-box 샘플링 및 역문제를 위한 분산 감소 0차 Langevin dynamics","authors":"Mehmet Berk Sahin, Behzad Sharif, Abolfazl Hashemi","pos":"#3305","link":"https://openreview.net/forum?id=UlFOINq6SY","abs":"Sampling from high-dimensional, non-log-concave distributions with unnormalized densities constitutes a fundamental challenge in machine learning, particularly when gradient information is inaccessible or computationally prohibitive. While Langevin dynamics provides a robust mechanism for gradient-based sampling, its extension to the derivative-free setting is frequently compromised by high variance and a lack of rigorous convergence guarantees in non-convex landscapes. In this work, we propose a principled variance-reduced zeroth-order Langevin dynamics framework that addresses these limitations for both general non-log-concave black-box sampling and inverse problems utilizing pre-trained score-based generative priors. We introduce a novel gradient estimator that significantly mitigates the variance inherent in traditional zeroth-order methods, enabling stable navigation of complex, multimodal posterior distributions. Theoretically, we establish the first non-asymptotic complexity bounds for this class of algorithms, proving convergence to the target distribution in terms of $\\varepsilon$-relative Fisher information, and, under a Poincare inequality, squared total variation distance, specifically for non-log-concave densities. We empirically validate our framework, demonstrating superior mixing and sampling accuracy on standard black-box benchmarks and establishing state-of-the-art performance for derivative-free linear and nonlinear inverse problems.","absKo":"고차원 비볼록(non-log-concave) 분포에서 정규화되지 않은 밀도를 가진 샘플링은 machine learning의 근본적인 도전 과제이며, 특히 gradient 정보에 접근할 수 없거나 계산적으로 너무 비싼 경우 더욱 그렇다. Langevin dynamics는 gradient 기반 샘플링을 위한 강력한 메커니즘을 제공하지만, derivative-free 설정으로의 확장은 비볼록 지형에서 높은 분산과 엄밀한 수렴 보장의 부재로 인해 자주 한계를 드러낸다. 본 연구에서는 일반적인 non-log-concave black-box sampling과 pre-trained score-based generative prior를 사용하는 inverse problem 모두에 대해 이러한 한계를 해결하는, 원리 기반의 variance-reduced zeroth-order Langevin dynamics 프레임워크를 제안한다. 우리는 전통적인 zeroth-order 방법에 내재된 분산을 크게 완화하는 새로운 gradient estimator를 도입하여 복잡하고 다봉적인 posterior 분포를 안정적으로 탐색할 수 있게 한다. 이론적으로는, 이 계열 알고리즘에 대해 최초의 non-asymptotic complexity bound를 확립하고, non-log-concave density에 대해 $\\varepsilon$-relative Fisher information 관점에서 target distribution으로의 수렴을 증명하며, Poincare inequality 하에서는 squared total variation distance에 대해서도 수렴을 보인다. 우리는 실험적으로 이 프레임워크를 검증하여, 표준 black-box benchmark에서 더 우수한 mixing과 sampling accuracy를 보이고, derivative-free linear 및 nonlinear inverse problem에서 state-of-the-art 성능을 달성함을 입증한다."},{"id":"61076","en":"Thinned Mean Field Langevin Dynamics","ko":"희소화된 Mean Field Langevin Dynamics","authors":"Zonghao Chen, Heishiro Kanagawa, Francois-Xavier Briol, Chris J Oates, Lester Mackey","pos":"#3306","link":"https://openreview.net/forum?id=tt4UrPSGTo","abs":"Several important learning tasks can be formulated as minimizing an entropy-regularized objective over an appropriate space of probability distributions. Mean-field Langevin dynamics (MFLD) facilitate computation in this general context, casting the minimizer as the invariant distribution of a McKean--Vlasov process, which can be numerically discretized using $N$ particles and thus simulated. However, simulating this interacting particle system has computational complexity $\\mathcal{O}(N^2)$. Motivated by recent research into \\emph{kernel thinning}, we propose \\texttt{KT-MFLD}, in which each particle interacts only with a coreset of size $\\mathcal{O}(N^{\\frac{1}{2}})$. \\texttt{KT-MFLD} thus reduces the computational complexity to $\\mathcal{O}(N^{\\frac{3}{2}})$ while, under mild regularity conditions, achieving the same convergence guarantees (up to logarithmic factors) as MFLD. Our theoretical analysis is empirically confirmed on tasks including the training of student-teacher neural networks, quantization with maximum mean discrepancy, and computation of predictively-oriented posteriors in a post-Bayesian framework.","absKo":"여러 중요한 learning task는 적절한 probability distribution space 위에서 entropy-regularized objective를 최소화하는 문제로 정식화할 수 있다. Mean-field Langevin dynamics (MFLD)는 이 일반적인 맥락에서 계산을 가능하게 하며, minimizer를 McKean--Vlasov process의 invariant distribution으로 해석하고, 이를 $N$개의 particle을 사용해 수치적으로 discretize하여 시뮬레이션할 수 있게 한다. 그러나 이 interacting particle system의 시뮬레이션은 계산 복잡도가 $\\mathcal{O}(N^2)$이다. 최근 \\emph{kernel thinning} 연구에 동기를 받아, 우리는 각 particle이 크기 $\\mathcal{O}(N^{\\frac{1}{2}})$의 coreset과만 상호작용하는 \\texttt{KT-MFLD}를 제안한다. 따라서 \\texttt{KT-MFLD}는 계산 복잡도를 $\\mathcal{O}(N^{\\frac{3}{2}})$로 줄이면서, mild regularity condition 하에서는 MFLD와 동일한 convergence guarantee를 (logarithmic factor를 제외하고) 달성한다. 우리의 이론적 분석은 student-teacher neural network training, maximum mean discrepancy를 이용한 quantization, 그리고 post-Bayesian framework에서 predictively-oriented posterior 계산을 포함한 과제들에서 실험적으로 확인된다."},{"id":"64300","en":"TFTF: Training-Free Targeted Flow for Conditional Sampling","ko":"TFTF: 조건부 샘플링을 위한 학습 불필요 표적 흐름","authors":"Qianqian Qu, Jun Liu","pos":"#3307","link":"https://openreview.net/forum?id=Oz7GovJhac","abs":"

We propose a training-free conditional sampling method for flow matching models based on importance sampling. Because a naïve application of importance sampling suffers from weight degeneracy in high-dimensional settings, we modify and incorporate a resampling technique in sequential Monte Carlo (SMC) during intermediate stages of the generation process. To encourage generated samples to diverge along distinct trajectories, we derive a stochastic flow with adjustable noise strength to replace the deterministic flow at the intermediate stage. Our framework requires no additional training, while providing theoretical guarantees of asymptotic accuracy. Experimentally, our method significantly outperforms existing approaches on conditional sampling tasks for MNIST and CIFAR-10. We further demonstrate the applicability of our approach in higher-dimensional, multimodal settings through text-to-image generation experiments on CelebA-HQ.

","absKo":"우리는 importance sampling에 기반한 flow matching model용 training-free conditional sampling 방법을 제안한다. 순진하게 importance sampling을 적용하면 고차원 설정에서 weight degeneracy가 발생하므로, 생성 과정의 중간 단계에서 sequential Monte Carlo(SMC)의 resampling 기법을 수정하여 통합한다. 생성된 sample이 서로 다른 trajectory를 따라 분기하도록 유도하기 위해, 중간 단계에서 deterministic flow를 대체하는 조정 가능한 noise strength를 갖는 stochastic flow를 유도한다. 우리의 프레임워크는 추가 학습이 필요 없으며, 동시에 점근적 정확성에 대한 이론적 보장을 제공한다. 실험적으로, 우리의 방법은 MNIST와 CIFAR-10의 conditional sampling task에서 기존 접근법보다 크게 우수한 성능을 보인다. 또한 CelebA-HQ에서의 text-to-image generation 실험을 통해, 더 높은 차원의 multimodal setting에서도 본 접근법이 적용 가능함을 보인다.

"},{"id":"62538","en":"Stationary MMD Points","ko":"정상성 MMD 포인트","authors":"Zonghao Chen, Toni Karvonen, Heishiro Kanagawa, Francois-Xavier Briol, Chris J Oates","pos":"#3308","link":"https://openreview.net/forum?id=fiTjjAgkiC","abs":"

Approximation of a target probability distribution using a finite set of points is a problem of fundamental importance in numerical integration. Several authors have proposed to select points by minimising a maximum mean discrepancy (MMD), but the non-convexity of this objective typically precludes global minimisation. Instead, we consider the concept of \\emph{stationary points of the MMD} which, in contrast to points globally minimising the MMD, can be accurately computed. Our main contributions are two-fold and theoretical in nature. We first prove the (perhaps surprising) result that, for integrands in the associated reproducing kernel Hilbert space, the numerical integration error of stationary MMD points vanishes \\emph{faster} than the MMD. Motivated by this \\emph{super-convergence} property, we consider MMD gradient flows as a practical strategy for computing stationary points of the MMD. We then prove that MMD gradient flow can indeed compute stationary MMD points, based on a refined convergence analysis that establishes a novel non-asymptotic finite-particle error bound.

","absKo":"

유한한 점 집합을 사용해 목표 확률분포를 근사하는 문제는 numerical integration에서 근본적으로 중요하다. 여러 저자는 maximum mean discrepancy (MMD)를 최소화하여 점을 선택하는 방법을 제안했지만, 이 목적함수의 non-convexity 때문에 일반적으로 전역 최소화는 불가능하다. 대신 우리는 \\emph{stationary points of the MMD} 개념을 고려하는데, 이는 MMD를 전역적으로 최소화하는 점들과 달리 정확하게 계산할 수 있다. 우리의 주요 기여는 이론적 성격의 두 가지이다. 먼저, 관련 reproducing kernel Hilbert space에 속한 integrand에 대해 stationary MMD point의 numerical integration error가 MMD 자체보다 \\emph{더 빠르게} 사라진다는 다소 놀라운 결과를 증명한다. 이러한 \\emph{super-convergence} 성질에 동기부여되어, 우리는 MMD gradient flow를 MMD의 stationary point를 계산하기 위한 실용적 전략으로 고려한다. 이어서 우리는 refined convergence analysis를 바탕으로 MMD gradient flow가 실제로 stationary MMD point를 계산할 수 있음을 증명하며, 이는 새로운 비점근적 finite-particle error bound를 확립한다.

"},{"id":"66702","en":"Simultaneous Confidence Bounds for Aggregated Effects via Exact Subset Optimization","ko":"정확한 부분집합 최적화를 통한 집계 효과에 대한 동시 신뢰 구간","authors":"Weihang Xu, Huajie Qian, Wotao Yin, Xinshang Wang","pos":"#3309","link":"https://openreview.net/forum?id=1C2LrZ7SNL","abs":"

We study simultaneous confidence bounds for aggregated effects over downward-closed subset families of independent statistical tests. The bounds are obtained by bootstrap calibration of the maximum normalized aggregated effect over the relevant subset family, yielding valid post-hoc inference for data-selected subsets and tighter bounds than classical methods that protect all linear contrasts. A central challenge is that the required maximization is a nonlinear combinatorial optimization problem whose exact solution is essential for correct coverage. We address this challenge by casting the problem as a densest subgraph optimization and reformulating it as a linear program, or as a mixed-integer linear program when downward-closed linear constraints are imposed, enabling efficient and exact evaluation. We further characterize the growth regime of the number of tests for which the bootstrap calibration remains valid and illustrate the method on several machine learning applications.

","absKo":"

우리는 독립적인 statistical test들의 downward-closed subset family에 대한 aggregated effect의 simultaneous confidence bound를 연구한다. 이 bound는 관련 subset family에 대해 normalized aggregated effect의 최대값을 bootstrap calibration하여 얻으며, data-selected subset에 대한 valid한 post-hoc inference와 모든 linear contrast를 보호하는 고전적 방법보다 더 타이트한 bound를 제공한다. 핵심 도전은 필요한 maximization이 비선형 조합 최적화 문제이며, 정확한 coverage를 위해 그 exact solution이 필수적이라는 점이다. 우리는 이 문제를 densest subgraph optimization으로 정식화하고 선형 계획법으로 재구성하며, downward-closed linear constraint가 부과된 경우에는 mixed-integer linear program으로 재구성하여 효율적이고 정확한 평가를 가능하게 한다. 또한 bootstrap calibration이 유효하게 유지되는 test 수의 성장 구간을 특성화하고, 여러 machine learning application에서 이 방법을 시연한다.

"},{"id":"64261","en":"Score-Repellent Monte Carlo: Toward Efficient Non-Markovian Sampler with Constant Memory in General State Spaces","ko":"점수 반발 몬테카를로: 일반 상태 공간에서 상수 메모리를 갖는 효율적 비마르코프 샘플러를 향하여","authors":"Jie Hu, Lingyun Chen, Geeho Kim, Jinyoung Choi, Bohyung Han, Do-Young Eun","pos":"#3310","link":"https://openreview.net/forum?id=PN8EiOzMuT","abs":"History-dependent sampling can reduce long-run Monte Carlo variance by discouraging redundant revisits, but existing schemes typically encode history through empirical measure on finite state spaces, which is infeasible in high-dimensional discrete configuration spaces or ill-posed in continuous domains. We propose *Score-Repellent Monte Carlo* (SRMC) framework that summarizes trajectory history by a fixed, $d$-dimensional running average of score evaluations and converts it into a history-dependent surrogate target via an exponential *score tilt*. The resulting surrogate family is normalization-free in the standard MCMC sense, yielding a generic wrapper: at each iteration, any standard base kernel designed for the target $\\pi$ can be run on the current surrogate $\\pi_{\\theta_n}$ while updating the history online. We analyze the coupled evolution of any estimator and the history recursion using stochastic approximation with controlled Markovian noise, establishing almost sure convergence and a joint central limit theorem. We identify regimes where the asymptotic covariance decreases as the repellence strength $\\alpha$ increases, exhibiting a scaling of $O(1/\\alpha)$, reproducing near-zero variance effect but now on general state spaces with constant memory. Empirical results across continuous targets and discrete energy-based models demonstrate that SRMC delivers notable improvements in estimator variance and in effectively covering modes in Gaussian mixtures, all while retaining $O(d)$ memory usage and minimal overhead per iteration.","absKo":"History-dependent sampling은 redundant revisit을 억제함으로써 장기 Monte Carlo variance를 줄일 수 있지만, 기존 방법은 대개 finite state space에서의 empirical measure를 통해 history를 인코딩하므로, 고차원 discrete configuration space에서는 비현실적이고 continuous domain에서는 잘 정의되지 않는다. 우리는 trajectory history를 score evaluation의 고정된 $d$차원 running average로 요약하고, 이를 exponential *score tilt*를 통해 history-dependent surrogate target으로 변환하는 *Score-Repellent Monte Carlo*(SRMC) framework를 제안한다. 그 결과로 얻어지는 surrogate family는 표준 MCMC의 의미에서 normalization-free이므로, 일반적인 wrapper를 제공한다. 즉, 각 iteration에서 target $\\pi$를 위해 설계된 임의의 standard base kernel을 현재 surrogate $\\pi_{\\theta_n}$ 위에서 실행하면서 history를 온라인으로 갱신할 수 있다. 우리는 제어된 Markovian noise를 갖는 stochastic approximation을 사용하여 임의의 estimator와 history recursion의 결합된 진화를 분석하고, almost sure convergence와 joint central limit theorem을 확립한다. 또한 repulsion strength $\\alpha$가 증가함에 따라 asymptotic covariance가 감소하는 regime를 식별하며, $O(1/\\alpha)$ scaling을 보이고, 거의 zero variance 효과를 일반 state space와 constant memory에서 재현함을 보인다. continuous target과 discrete energy-based model 전반의 실험 결과는 SRMC가 estimator variance를 현저히 개선하고 Gaussian mixture의 mode를 효과적으로 커버함을 보여주며, 이 모든 것을 $O(d)$ memory 사용량과 iteration당 최소한의 오버헤드로 달성한다."},{"id":"62449","en":"Scalable Sampling via Generalized Fixed-Point Diffusion Matching","ko":"일반화된 고정점 Diffusion Matching을 통한 확장 가능한 Sampling","authors":"Denis Blessing, Lorenz Richter, Julius Berner, Egor Malitskiy, Gerhard Neumann","pos":"#3311","link":"https://openreview.net/forum?id=gUbNUNuJbO","abs":"

Sampling from unnormalized densities using diffusion models has emerged as a powerful paradigm. However, while recent approaches that use least-squares `matching' objectives have improved scalability, they often necessitate significant trade-offs, such as restricting prior distributions or relying on unstable optimization schemes. By generalizing these methods as special forms of fixed-point iterations rooted in Nelson's relation, we develop a new method that addresses these limitations. Our approach enables learning a stochastic transport map between arbitrary prior and target distributions with a single, scalable, and stable objective. Furthermore, we introduce a damped variant of this iteration that incorporates a regularization term to mitigate mode collapse. Empirically, we demonstrate that our method enables sampling at unprecedented scales while preserving mode diversity, achieving state-of-the-art results on complex synthetic densities and high-dimensional molecular benchmarks.

","absKo":"

정규화되지 않은 density에서 diffusion model을 사용해 sampling하는 것은 강력한 paradigm으로 부상했다. 그러나 최근 least-squares `matching' objective를 사용하는 접근법은 scalability를 개선했지만, prior distribution을 제한하거나 불안정한 optimization scheme에 의존하는 등 상당한 trade-off를 요구하는 경우가 많다. 우리는 Nelson's relation에 뿌리를 둔 fixed-point iteration의 특수한 형태로 이들 방법을 일반화함으로써, 이러한 한계를 해결하는 새로운 방법을 개발한다. 우리의 접근법은 단일하고, 확장 가능하며, 안정적인 objective만으로 임의의 prior와 target distribution 사이의 stochastic transport map을 학습할 수 있게 한다. 나아가 우리는 mode collapse를 완화하기 위해 regularization term을 포함하는 이 iteration의 damped variant를 도입한다. 실험적으로, 우리는 이 방법이 mode diversity를 보존하면서 전례 없는 규모의 sampling을 가능하게 하며, 복잡한 synthetic density와 고차원 molecular benchmark에서 state-of-the-art 결과를 달성함을 보인다.

"},{"id":"63925","en":"Scalable Power Sampling: Unlocking Efficient, Training-Free Reasoning for LLMs via Distribution Sharpening","ko":"Scalable Power Sampling: 분포 sharpen을 통한 LLM의 효율적이고 학습 불필요한 추론","authors":"Xiaotong Ji, Rasul Tutunov, Matthieu Zimmer, Haitham Bou Ammar","pos":"#3312","link":"https://openreview.net/forum?id=SVyjXhZlDe","abs":"Reinforcement learning (RL) post-training is a dominant approach for improving the reasoning performance of large language models (LLMs), yet growing evidence suggests that its gains arise primarily from distribution sharpening rather than the acquisition of new capabilities. Recent work has shown that sampling from the power distribution of LLMs using Markov chain Monte Carlo (MCMC) can recover performance comparable to RL post-training without relying on external rewards; however, the high computational cost of MCMC makes such approaches impractical for widespread adoption. In this work, we propose a theoretically grounded alternative that eliminates the need for iterative MCMC. We derive a novel formulation showing that the global power distribution can be approximated by a token-level scaled low-temperature one, where the scaling factor captures future trajectory quality. Leveraging this insight, we introduce a training-free and verifier-free algorithm that sharpens the base model's generative distribution autoregressively. Empirically, we evaluate our method on math, QA, and code tasks across four LLMs, and show that our method matches or surpasses one-shot GRPO without relying on any external rewards, while reducing inference latency by over $10\\times$ compared to MCMC-based sampling.","absKo":"Reinforcement learning(RL) post-training은 large language model(LLM)의 reasoning 성능을 향상시키는 지배적인 접근법이지만, 점점 더 많은 증거는 그 이득이 주로 distribution sharpening에서 비롯되며 새로운 능력의 획득에서 오는 것은 아님을 시사한다. 최근 연구는 Markov chain Monte Carlo(MCMC)를 사용해 LLM의 power distribution에서 sampling하면 외부 reward에 의존하지 않고도 RL post-training에 필적하는 성능을 회복할 수 있음을 보였다. 그러나 MCMC의 높은 계산 비용 때문에 이러한 접근은 널리 채택되기에는 비실용적이다. 본 연구에서는 반복적인 MCMC가 필요 없는 이론적으로 타당한 대안을 제안한다. 우리는 global power distribution이 token 수준의 scaled low-temperature 분포로 근사될 수 있음을 보이는 새로운 formulation을 유도하며, 여기서 scaling factor는 future trajectory quality를 포착한다. 이 통찰을 바탕으로, base model의 generative distribution을 autoregressive하게 sharpen하는 training-free 및 verifier-free algorithm을 제안한다. 실험적으로 우리는 네 개의 LLM에서 math, QA, code task에 대해 제안 방법을 평가했고, 외부 reward에 의존하지 않으면서 one-shot GRPO와 동등하거나 더 나은 성능을 보였으며, MCMC 기반 sampling 대비 inference latency를 10배 이상 줄였다."},{"id":"65452","en":"Reinforced Sequential Monte Carlo for Amortised Sampling","ko":"Amortised 샘플링을 위한 강화 Sequential Monte Carlo","authors":"Sanghyeok Choi, Sarthak Mittal, Víctor Elvira, Jinkyoo Park, Esmeralda S. Whitammer","pos":"#3313","link":"https://openreview.net/forum?id=DWaToCuNwa","abs":"

This paper proposes a synergy of amortised and particle-based methods for sampling from distributions defined by unnormalised density functions. We state a connection between sequential Monte Carlo (SMC) and neural sequential samplers trained by maximum-entropy reinforcement learning (MaxEnt RL), wherein learnt sampling policies and value functions define proposal kernels and twist functions. Exploiting this connection, we introduce an off-policy RL training procedure for the sampler that uses samples from SMC -- using the learnt sampler as a proposal -- as a behaviour policy that better explores the target distribution. We describe techniques for stable joint training of proposals and twist functions and an adaptive weight tempering scheme to reduce training signal variance. Furthermore, building upon past attempts to use experience replay to guide the training of neural samplers, we derive a way to combine historical samples with annealed importance sampling weights within a replay buffer. On synthetic multi-modal targets (in both continuous and discrete spaces) and the Boltzmann distribution of alanine dipeptide conformations, we demonstrate improvements in approximating the true distribution as well as training stability compared to both amortised and Monte Carlo methods.

","absKo":"

본 논문은 정규화되지 않은 density function으로 정의되는 분포에서 샘플링하기 위해 amortized method와 particle-based method의 시너지를 제안한다. 우리는 sequential Monte Carlo(SMC)와 maximum-entropy reinforcement learning(MaxEnt RL)으로 학습된 neural sequential sampler 사이의 연결을 제시하며, 여기서 학습된 sampling policy와 value function은 proposal kernel과 twist function을 정의한다. 이 연결을 활용하여, 우리는 학습된 sampler를 proposal로 사용한 SMC에서 얻은 샘플을 behavior policy로 삼아 target distribution을 더 잘 탐색하는 sampler용 off-policy RL 학습 절차를 도입한다. 또한 proposal과 twist function의 안정적인 joint training을 위한 기법과, training signal variance를 줄이기 위한 adaptive weight tempering scheme을 설명한다. 더 나아가, neural sampler 학습을 유도하기 위해 experience replay를 사용하려는 과거 시도를 바탕으로, replay buffer 내에서 historical sample과 annealed importance sampling weight를 결합하는 방법을 유도한다. 연속 및 이산 공간 모두에서의 synthetic multi-modal target과 alanine dipeptide conformation의 Boltzmann distribution에 대해, 우리는 amortized method와 Monte Carlo method 모두에 비해 true distribution 근사와 학습 안정성에서 향상을 보인다.

"},{"id":"65408","en":"Practical and Scalable Hamiltonian Monte Carlo Without the Metropolis Test","ko":"Metropolis 검정 없는 실용적이고 확장 가능한 Hamiltonian Monte Carlo","authors":"Jakob Robnik, Reuben Cohn-Gordon, Uros Seljak","pos":"#3314","link":"https://openreview.net/forum?id=DxvZnTVNfP","abs":"

Hamiltonian Monte Carlo and underdamped Langevin Monte Carlo are state-of-the-art methods for taking samples from high-dimensional distributions with a differentiable density function. To generate samples, they numerically integrate Hamiltonian or Langevin dynamics. This numerical integration introduces an asymptotic bias in Monte Carlo estimators of expectation values, which can be eliminated by adjusting the dynamics with a Metropolis-Hastings (MH) proposal step. Alternatively, one can trade bias for variance by avoiding MH, and select an integration step size that ensures sufficiently small asymptotic bias, relative to the variance inherent in a finite set of samples. Such \\emph{unadjusted} methods often significantly outperform their adjusted counterparts in high-dimensional problems where sampling would otherwise be prohibitively expensive, yet are rarely used in statistical applications due to the absence of an automated way of choosing a step size. We propose just such an automatic tuning scheme that takes a user-provided asymptotic bias tolerance and selects a step size that ensures it. The key to the method is a relationship we establish between the energy change during integration and the asymptotic bias. We show that this procedure rigorously bounds the asymptotic bias for Gaussian target distributions. We then numerically show that the procedure works beyond Gaussians. To demonstrate the practicality of the proposed scheme, we provide a comprehensive comparison of adjusted and unadjusted samplers on Bayesian inference problems and on a statistical physics model in more than one million parameters. With our tuning scheme, the unadjusted methods achieve close to optimal performance, significantly and consistently outperforming their adjusted counterparts.

","absKo":"Hamiltonian Monte Carlo와 underdamped Langevin Monte Carlo는 미분 가능한 density function을 갖는 고차원 분포에서 sample을 추출하기 위한 state-of-the-art method이다. sample을 생성하기 위해 이들은 Hamiltonian dynamics 또는 Langevin dynamics를 수치적으로 적분한다. 이러한 수치 적분은 expectation value의 Monte Carlo estimator에 asymptotic bias를 도입하며, 이는 Metropolis-Hastings(MH) proposal step으로 dynamics를 조정함으로써 제거할 수 있다. 또는 MH를 피하고 bias를 variance와 맞바꾸는 방식도 가능하며, 유한한 sample 집합에 내재된 variance에 비해 충분히 작은 asymptotic bias를 보장하는 integration step size를 선택할 수 있다. 이러한 \\emph{unadjusted} method는 sampling이 그렇지 않으면 매우 비싸게 드는 고차원 문제에서 조정된 대응 방법보다 훨씬 뛰어난 경우가 많지만, step size를 자동으로 선택하는 방법이 없기 때문에 통계적 응용에서는 거의 사용되지 않는다. 우리는 사용자가 제공한 asymptotic bias tolerance를 받아 이를 보장하는 step size를 선택하는 자동 tuning scheme을 제안한다. 이 방법의 핵심은 적분 중 energy 변화와 asymptotic bias 사이에 우리가 확립한 관계이다. 우리는 이 절차가 Gaussian target distribution에 대해 asymptotic bias를 엄밀하게 상한함을 보인다. 이후 수치 실험을 통해 이 절차가 Gaussian을 넘어서는 경우에도 작동함을 보인다. 제안한 scheme의 실용성을 입증하기 위해, 우리는 Bayesian inference 문제와 백만 개가 넘는 parameter를 가진 statistical physics model에서 adjusted sampler와 unadjusted sampler를 포괄적으로 비교한다. 우리의 tuning scheme을 사용하면 unadjusted method는 거의 최적의 성능을 달성하며, adjusted counterpart를 크게 그리고 일관되게 능가한다.

"},{"id":"62468","en":"Constrained Adaptive Rejection Sampling","ko":"제약적 적응형 Rejection Sampling","authors":"Paweł Parys, Sairam Vaidya, Taylor Berg-Kirkpatrick, Loris D&#x27;Antoni","pos":"#3400","link":"https://openreview.net/forum?id=gLRqqQOM9K","abs":"

Language Models (LMs) are increasingly used in applications where generated outputs must satisfy strict semantic or syntactic constraints. Existing approaches to constrained generation fall along a spectrum: greedy constrained decoding methods enforce validity during decoding but distort the LM’s distribution, while rejection sampling (RS) preserves fidelity but wastes computation by discarding invalid outputs. Both extremes are problematic in domains such as program fuzzing, where both validity and diversity of samples are essential. We present Constrained Adaptive Rejection Sampling (CARS), an approach that strictly improves the sample-efficiency of RS without distributional distortion. CARS begins with unconstrained LM sampling and adaptively rules out constraint-violating continuations by recording them in a trie and subtracting their probability mass from future draws. This adaptive pruning ensures that prefixes proven invalid are never revisited, acceptance rates improve monotonically, and the resulting samples exactly follow the constrained distribution. In experiments on a variety of domains--e.g., program fuzzing and molecular generation--CARS consistently achieves higher efficiency--measured in the number of LM forward passes per valid sample--while also producing stronger sample diversity than both Greedy Constrained Decoding (GCD) and methods that approximate the LM's distribution.

","absKo":"

Language Models(LMs)는 생성된 출력이 엄격한 semantic 또는 syntactic constraints를 만족해야 하는 응용에서 점점 더 많이 사용되고 있다. constrained generation을 위한 기존 접근법은 스펙트럼 양 끝에 놓여 있다: greedy constrained decoding 방법은 decoding 중 validity를 강제하지만 LM의 distribution을 왜곡하고, rejection sampling(RS)은 fidelity를 보존하지만 invalid output을 버림으로써 계산을 낭비한다. 이 두 극단은 program fuzzing과 같이 sample의 validity와 diversity가 모두 필수적인 영역에서 특히 문제가 된다. 우리는 distributional distortion 없이 RS의 sample-efficiency를 엄격히 개선하는 Constrained Adaptive Rejection Sampling(CARS)을 제안한다. CARS는 unconstrained LM sampling으로 시작한 뒤, 제약을 위반하는 continuation을 적응적으로 배제하며, 이를 trie에 기록하고 향후 draw에서 그 probability mass를 빼는 방식으로 동작한다. 이러한 adaptive pruning은 invalid로 증명된 prefix가 다시는 재방문되지 않도록 보장하고, acceptance rate가 단조롭게 향상되며, 결과 sample이 constrained distribution을 정확히 따르도록 한다. program fuzzing과 molecular generation 등 다양한 도메인 실험에서 CARS는 valid sample당 LM forward pass 수로 측정한 효율성이 일관되게 더 높았고, Greedy Constrained Decoding(GCD) 및 LM의 distribution을 근사하는 방법들보다 더 강한 sample diversity도 보였다.

"},{"id":"61732","en":"Covariance estimation using Markov chain Monte Carlo","ko":"Markov Chain Monte Carlo를 이용한 Covariance 추정","authors":"Yunbum Kook, Shunshi Zhang","pos":"#3402","link":"https://openreview.net/forum?id=nNBMlUWEKi","abs":"We investigate the complexity of covariance matrix estimation for Gibbs distributions based on dependent samples from a Markov chain. We show that when $π$ satisfies a Poincaré inequality and the chain possesses a spectral gap, we can achieve similar sample complexity using MCMC as compared to an estimator constructed using i.i.d. samples, with potentially much better query complexity. As an application of our methods, we show improvements for the query complexity in both constrained and unconstrained settings for concrete instances of MCMC. In particular, we provide guarantees regarding isotropic rounding procedures for sampling uniformly on convex bodies.","absKo":"우리는 Markov chain에서의 종속 표본을 기반으로 한 Gibbs distribution의 covariance matrix estimation 복잡도를 조사한다. 우리는 $\\pi$가 Poincaré inequality를 만족하고 chain이 spectral gap을 가질 때, i.i.d. sample로 구성한 estimator와 비교해 MCMC를 사용해 유사한 sample complexity를 달성할 수 있으며, query complexity는 잠재적으로 훨씬 더 좋아질 수 있음을 보인다. 제안한 방법의 응용으로, 구체적인 MCMC 사례들에서 constrained 및 unconstrained 설정 모두에 대해 query complexity 개선을 보인다. 특히, convex body 위에서 균일하게 sampling하기 위한 isotropic rounding procedure에 대한 보장을 제시한다."},{"id":"66689","en":"DC-LA: Difference-of-Convex Langevin Algorithm","ko":"DC-LA: Difference-of-Convex Langevin 알고리즘","authors":"Hoang Phuc Hau Luu, Zhongjian Wang","pos":"#3403","link":"https://openreview.net/forum?id=1JwSFRvC4E","abs":"We study a sampling problem whose target distribution is $\\pi \\propto \\exp(-f-r)$ where the data fidelity term $f$ is Lipschitz smooth while the regularizer term $r=r_1-r_2$ is a non-smooth difference-of-convex (DC) function, i.e., $r_1,r_2$ are convex. By leveraging the DC structure of $r$, we can smooth out $r$ by applying Moreau envelopes to $r_1$ and $r_2$ separately. In line of DC programming, we then redistribute the concave part of the regularizer to the data fidelity and study its corresponding proximal Langevin algorithm (termed DC-LA). We establish convergence of DC-LA to the target distribution $\\pi$, up to discretization and smoothing errors, in the $q$-Wasserstein distance for all $q \\in \\mathbb{N}^*$, under the assumption that $V$ is distant dissipative. Our results improve previous work on non-log-concave sampling in terms of a more general framework and assumptions. Moreover, numerical experiments show that DC-LA produces accurate distributions in synthetic settings and reliably provides uncertainty quantification in a real-world Computed Tomography application.","absKo":"우리는 타깃 분포가 $\\pi \\propto \\exp(-f-r)$인 샘플링 문제를 연구한다. 여기서 data fidelity 항 $f$는 Lipschitz smooth이고, regularizer 항 $r=r_1-r_2$는 비매끄러운 difference-of-convex(DC) 함수이다. 즉, $r_1,r_2$는 convex이다. $r$의 DC 구조를 활용하면, $r_1$과 $r_2$에 대해 각각 Moreau envelope를 적용하여 $r$를 smooth하게 만들 수 있다. DC programming의 맥락에 따라, 우리는 regularizer의 concave part를 data fidelity로 재분배하고, 그에 대응하는 proximal Langevin algorithm(DC-LA라 부름)을 연구한다. $V$가 distant dissipative라는 가정 하에, 우리는 DC-LA가 분할 오차와 smoothing 오차를 제외하고 모든 $q \\in \\mathbb{N}^*$에 대해 $q$-Wasserstein distance에서 타깃 분포 $\\pi$로 수렴함을 보인다. 우리의 결과는 더 일반적인 framework와 가정 아래에서 non-log-concave sampling에 대한 기존 연구를 개선한다. 또한 수치 실험은 DC-LA가 synthetic setting에서 정확한 분포를 생성하고, 실제 Computed Tomography 응용에서 신뢰할 수 있는 uncertainty quantification을 제공함을 보여준다."},{"id":"64577","en":"Diffusion differentiable resampling","ko":"Diffusion 미분 가능 Resampling","authors":"Jennifer R. Andersson, Zheng Zhao","pos":"#3404","link":"https://openreview.net/forum?id=M0e5XORjAW","abs":"

This paper is concerned with differentiable resampling in the context of sequential Monte Carlo (e.g., particle filtering). We propose a new informative resampling method that is instantly differentiable, based on an ensemble score diffusion model. We theoretically prove that our diffusion resampling method provides a consistent resampling distribution, and we show empirically that it outperforms the state-of-the-art differentiable resampling methods on multiple filtering and parameter estimation benchmarks. Finally, we show that it achieves competitive end-to-end performance when used in learning a complex dynamics-decoder model with high-dimensional image observations.

","absKo":"

본 논문은 sequential Monte Carlo(예: particle filtering) 맥락에서의 differentiable resampling을 다룬다. 우리는 ensemble score diffusion model에 기반한, 즉시 differentiable한 새로운 informative resampling 방법을 제안한다. 이론적으로는 우리의 diffusion resampling 방법이 일관된 resampling distribution을 제공함을 증명하고, 실험적으로는 여러 filtering 및 parameter estimation benchmark에서 최신 differentiable resampling 방법들을 능가함을 보인다. 마지막으로, 고차원 image observation을 사용하는 복잡한 dynamics-decoder model을 학습할 때 경쟁력 있는 end-to-end 성능을 달성함을 보인다.

"},{"id":"68813","en":"Diffusion posterior sampling for simulation-based inference in tall data settings","ko":"Tall Data 환경에서 Simulation-Based Inference를 위한 Diffusion Posterior Sampling","authors":"Julia Linhart, Gabriel Cardoso, Alexandre Gramfort, Sylvain Le Corff, Pedro Luiz Coelho Rodrigues","pos":"#3405","abs":"

Identifying the parameters of a non-linear model that best explain observed data is a core task across scientific fields. When such models rely on complex simulators, evaluating the likelihood is typically intractable, making traditional inference methods such as MCMC inapplicable. Simulation-based inference (SBI) addresses this by training deep generative models to approximate the posterior distribution over parameters using simulated data. In this work, we consider the tall data setting, where multiple independent observations provide additional information, allowing sharper posteriors and improved parameter identifiability.\nBuilding on the flourishing score-based diffusion literature, F-NPSE (Geffner et al., 2023) estimates the tall data posterior by composing individual scores from a neural network trained only for a single context observation. This enables more flexible and simulation-efficient inference than alternative approaches for tall datasets in SBI. \nHowever, it relies on costly Langevin dynamics during sampling. We propose a new algorithm that eliminates the need for Langevin steps by explicitly approximating the diffusion process of the tall data posterior. Our method retains the advantages of compositional score-based inference while being significantly faster and more stable than F-NPSE. We demonstrate its improved performance on toy problems and standard SBI benchmarks, and showcase its scalability by applying it to a complex real-world model from computational neuroscience.

","absKo":"

관측된 데이터를 가장 잘 설명하는 non-linear model의 parameter를 식별하는 것은 여러 과학 분야에서 핵심적인 과제이다. 이러한 model이 복잡한 simulator에 의존할 때, likelihood를 계산하는 것은 일반적으로 다루기 어려워 MCMC와 같은 전통적 inference method를 적용할 수 없다. Simulation-based inference(SBI)는 simulated data를 사용하여 parameter에 대한 posterior distribution을 근사하도록 deep generative model을 학습시킴으로써 이를 해결한다. 본 연구에서는 여러 독립적 observation이 추가 정보를 제공하여 더 날카로운 posterior와 향상된 parameter identifiability를 가능하게 하는 tall data setting을 고려한다.\n급성장 중인 score-based diffusion 문헌을 기반으로, F-NPSE(Geffner et al., 2023)는 단일 context observation에 대해서만 학습된 neural network의 individual score를 합성하여 tall data posterior를 추정한다. 이는 SBI에서 tall dataset을 다루는 대안적 접근보다 더 유연하고 simulation-efficient한 inference를 가능하게 한다.\n그러나 sampling 시 비용이 큰 Langevin dynamics에 의존한다. 우리는 tall data posterior의 diffusion process를 명시적으로 근사함으로써 Langevin step이 필요 없는 새로운 algorithm을 제안한다. 우리의 방법은 compositional score-based inference의 장점을 유지하면서 F-NPSE보다 훨씬 더 빠르고 안정적이다. 우리는 toy problem과 표준 SBI benchmark에서 향상된 성능을 보이고, computational neuroscience의 복잡한 실제 model에 적용함으로써 확장성을 입증한다.

"},{"id":"62949","en":"Discrete Diffusion Samplers and Bridges: Off-Policy Algorithms and Applications in Latent Spaces","ko":"Discrete Diffusion Samplers and Bridges: 잠재 공간의 오프-폴리시 알고리즘과 응용","authors":"Arran Carter, Sanghyeok Choi, Kirill Tamogashev, Víctor Elvira, Esmeralda S. Whitammer","pos":"#3406","link":"https://openreview.net/forum?id=bYg45bvuV9","abs":"Sampling from a distribution $p(x) \\propto e^{-\\mathcal{E}(x)}$ known up to a normalising constant is an important and challenging problem in statistics. Recent years have seen the rise of a new family of amortised sampling algorithms, commonly referred to as diffusion samplers, that enable fast and efficient sampling from an unnormalised density. Such algorithms have been widely studied for continuous-space sampling tasks; however, their application to problems in discrete space remains largely unexplored. Although some progress has been made in this area, discrete diffusion samplers do not take full advantage of ideas commonly used for continuous-space sampling. In this paper, we propose to bridge this gap by introducing off-policy training techniques for discrete diffusion samplers. We show that these techniques improve the performance of discrete samplers on both established and new synthetic benchmarks. Next, we generalise discrete diffusion samplers to the task of bridging between two arbitrary distributions, introducing data-to-energy Schrödinger bridge training for the discrete domain for the first time. Lastly, we showcase the application of the proposed diffusion samplers to data-free posterior sampling in the discrete latent spaces of image generative models.","absKo":"정규화 상수까지는 알려져 있지 않지만 $p(x) \\propto e^{-\\mathcal{E}(x)}$인 분포에서 샘플링하는 것은 통계학에서 중요하면서도 어려운 문제이다. 최근에는 amortised sampling algorithm의 새로운 계열, 흔히 diffusion sampler라고 불리는 방법들이 부상하여, unnormalised density에서 빠르고 효율적인 샘플링을 가능하게 했다. 이러한 알고리즘은 continuous-space sampling task에서 널리 연구되어 왔지만, discrete space 문제에의 적용은 여전히 거의 탐구되지 않았다. 이 분야에서 일부 진전이 있었지만, discrete diffusion sampler는 continuous-space sampling에서 흔히 사용되는 아이디어를 충분히 활용하지 못한다. 본 논문에서는 discrete diffusion sampler를 위한 off-policy training technique를 도입하여 이 간극을 메우고자 한다. 우리는 이러한 기법이 기존 및 새로운 synthetic benchmark 모두에서 discrete sampler의 성능을 향상시킴을 보인다. 다음으로, discrete diffusion sampler를 두 임의의 분포 사이를 bridging 하는 과제로 일반화하고, discrete 도메인에서 처음으로 data-to-energy Schrödinger bridge training을 도입한다. 마지막으로, 이미지 generative model의 discrete latent space에서 data-free posterior sampling에 제안한 diffusion sampler를 적용하는 사례를 제시한다."},{"id":"65085","en":"Entropic Mirror Monte Carlo","ko":"엔트로픽 Mirror Monte Carlo","authors":"Anas CHERRADI, Yazid Janati, Sylvain Le Corff, Julien Stoehr, Alain Oliviero Durmus, Yohan Petetin","pos":"#3407","link":"https://openreview.net/forum?id=GugLtzm1WC","abs":"

Importance sampling is a Monte Carlo method which designs estimators of expectations under a target distribution using weighted samples from a proposal distribution. When the target distribution is complex, such as multimodal distributions in high-dimensional spaces, the efficiency of importance sampling critically depends on the choice of the proposal distribution. In this paper, we propose a novel adaptive scheme for the construction of efficient proposal distributions. Our algorithm promotes efficient exploration of the target distribution by combining global sampling mechanisms with a delayed weighting procedure. The proposed weighting mechanism plays a key role by enabling rapid resampling in regions where the proposal distribution is poorly adapted to the target. Our sampling algorithm is shown to be geometrically convergent under mild assumptions and is illustrated through various numerical experiments.

","absKo":"

Importance sampling은 target distribution에 대한 expectation을 proposal distribution에서 가중된 sample로 추정량을 구성하는 Monte Carlo 방법이다. 대상 분포가 고차원 공간의 multimodal distribution처럼 복잡할 때, importance sampling의 효율성은 proposal distribution의 선택에 결정적으로 좌우된다. 본 논문에서는 효율적인 proposal distribution을 구성하기 위한 새로운 adaptive scheme을 제안한다. 우리의 알고리즘은 global sampling mechanism과 delayed weighting procedure를 결합하여 target distribution의 효율적인 탐색을 촉진한다. 제안한 weighting mechanism은 proposal distribution이 target에 잘 맞지 않는 영역에서 빠른 resampling을 가능하게 함으로써 핵심적인 역할을 한다. 우리의 sampling algorithm은 완화된 가정 하에서 geometrically convergent함을 보이며, 다양한 수치 실험을 통해 그 성능을 입증한다.

"},{"id":"62793","en":"Markov Chain Monte Carlo without Evaluating the Target: an Auxiliary Variable Approach","ko":"목표분포를 평가하지 않는 Markov Chain Monte Carlo: Auxiliary Variable 접근법","authors":"Wei Yuan, Guanyang Wang","pos":"#3410","link":"https://openreview.net/forum?id=dDkl5ZcyTl","abs":"

In sampling tasks, it is common for target distributions to be known up to a normalizing constant. However, in many situations, even evaluating the unnormalized distribution can be costly or infeasible. This issue arises in scenarios such as sampling from the Bayesian posterior for tall datasets and the 'doubly-intractable' distributions. In this paper, we begin by observing that seemingly different Markov chain Monte Carlo (MCMC) algorithms, such as the exchange algorithm, PoissonMH, and TunaMH, can be unified under a simple common procedure. We then extend this procedure into a novel framework that allows the use of auxiliary variables in both the proposal and the acceptance--rejection step. Several new MCMC algorithms emerge from this framework that uses estimated gradients to guide the proposal moves. They have demonstrated significantly better performance than existing methods on both synthetic and real datasets. We also develop theory for the new framework and use it to simplify and extend results for existing algorithms.

","absKo":"샘플링 작업에서는 target distribution이 normalizing constant까지는 알려져 있는 경우가 흔하다. 그러나 많은 상황에서는 정규화되지 않은 분포 자체를 평가하는 것조차 비용이 크거나 불가능할 수 있다. 이러한 문제는 대규모 데이터셋에서 Bayesian posterior를 샘플링하는 경우와 'doubly-intractable' distribution에서 나타난다. 이 논문에서는 먼저 exchange algorithm, PoissonMH, TunaMH처럼 겉보기에는 서로 다른 Markov chain Monte Carlo(MCMC) algorithm이 단순한 공통 절차 아래에서 통합될 수 있음을 보인다. 그런 다음 이 절차를 확장하여 proposal 단계와 acceptance--rejection 단계 모두에서 auxiliary variable을 사용할 수 있는 새로운 framework로 발전시킨다. 이 framework로부터, estimated gradient를 사용해 proposal move를 안내하는 여러 새로운 MCMC algorithm이 도출된다. 이들은 synthetic 및 실제 데이터셋 모두에서 기존 방법보다 현저히 더 나은 성능을 보였다. 또한 우리는 이 새로운 framework에 대한 이론을 전개하고, 이를 사용해 기존 algorithm의 결과를 단순화하고 확장한다.

"},{"id":"64092","en":"Mitigating Premature Exploitation in Particle-based Monte Carlo for Inference-Time Scaling","ko":"추론 시점 스케일링을 위한 입자 기반 몬테카를로에서의 조기 착취 완화","authors":"Giorgio Giannone, Guangxuan Xu, Nikhil Nayak, Rohan Awhad, Shivchander Sudalairaj, Kai Xu, Akash Srivastava","pos":"#3411","link":"https://openreview.net/forum?id=R0eRg4p4Uj","abs":"

Inference-Time Scaling (ITS) improves language models by allocating more computation at generation time. Particle Filtering (PF) has emerged as a strong ITS method for complex mathematical reasoning tasks, but it is vulnerable when guided by process reward models, which often assign overconfident scores early in the reasoning process. This causes PF to suffer from premature exploitation: it myopically commits to locally promising trajectories, prunes potentially correct hypotheses, and converges to suboptimal solutions. This failure mode, known as particle impoverishment, is especially severe under constrained computational budgets. To address this, we analyze the problem and identify two root causes: a lack of diversity in the particle set due to overconfident resampling and consequent inability to assess the potential of a reasoning path. We introduce Entropic Particle Filtering (ePF), an algorithm that integrates two new techniques to solve these issues. The first technique, Entropic Annealing (EA), directly mitigates particle impoverishment by monitoring search diversity via entropy; when diversity drops, it intervenes by dynamically annealing the resampling distribution to preserve exploration. The second, an enhancement called Look-ahead Modulation (LaM), adds a predictive guide to evaluate a state's potential based on its successors. On several challenging math benchmarks, ePF significantly outperforms strong baselines and achieves up to a 50\\% relative improvement in task reward. Together, these methods improve PF's resilience by balancing the exploration of diverse solution spaces with the exploitation of high-reward regions, ultimately leading to higher-quality solutions.

","absKo":"

Inference-Time Scaling(ITS)은 생성 시점에 더 많은 computation을 할당함으로써 language model을 개선한다. Particle Filtering(PF)은 복잡한 수학 추론 task에서 강력한 ITS 방법으로 떠올랐지만, process reward model이 이끄는 경우 취약하다. 이러한 모델은 추론 과정의 초기에 종종 과도하게 자신 있는 점수를 부여한다. 이로 인해 PF는 premature exploitation에 빠지는데, 국소적으로 유망해 보이는 trajectory에 근시안적으로 집착하고, 잠재적으로 올바른 hypothesis를 잘라내며, suboptimal solution으로 수렴한다. particle impoverishment로 알려진 이 실패 모드는 제한된 computation budget 하에서 특히 심각하다. 이를 해결하기 위해 우리는 문제를 분석하고 두 가지 근본 원인을 식별한다. 첫째, 과도하게 자신 있는 resampling으로 인해 particle set의 다양성이 부족하다는 점, 둘째, 그 결과로 추론 경로의 잠재력을 평가하지 못한다는 점이다. 우리는 이 문제를 해결하기 위해 두 가지 새로운 기법을 통합한 Entropic Particle Filtering(ePF)을 제안한다. 첫 번째 기법인 Entropic Annealing(EA)은 entropy를 통해 search diversity를 모니터링함으로써 particle impoverishment를 직접 완화한다. 다양성이 떨어지면 resampling distribution을 동적으로 annealing하여 exploration을 보존한다. 두 번째 기법인 Look-ahead Modulation(LaM)은 successor를 바탕으로 state의 잠재력을 평가하는 predictive guide를 추가한다. 여러 까다로운 수학 benchmark에서 ePF는 강력한 baseline을 크게 능가하며 task reward에서 최대 50\\%의 상대적 향상을 달성한다. 이 두 방법은 다양한 solution space의 exploration과 high-reward region의 exploitation 사이를 균형 있게 맞춤으로써 PF의 resilience를 향상시키고, 궁극적으로 더 높은 품질의 solution으로 이어진다.

"},{"id":"60662","en":"Neural Low-Discrepancy Sequences","ko":"신경 기반 저편차 시퀀스","authors":"Michael Van Huffel, Nathan Kirk, Makram Chahine, Daniela Rus, T. Konstantin Rusch","pos":"#3412","link":"https://openreview.net/forum?id=xtWQnuo8YO","abs":"

Low-discrepancy points are designed to efficiently fill the space in a uniform manner. This uniformity is highly advantageous in many problems in science and engineering, including in numerical integration, computer vision, machine perception, computer graphics, machine learning, and simulation. Whereas most previous low-discrepancy constructions rely on abstract algebra and number theory, Message-Passing Monte Carlo (MPMC) was recently introduced to exploit machine learning methods for generating point sets with lower discrepancy than previously possible. However, MPMC is limited to generating point sets and cannot be extended to low-discrepancy sequences (LDS), i.e., sequences of points in which every prefix has low discrepancy, a property essential for many applications. To address this limitation, we introduce Neural Low-Discrepancy Sequences NeuroLDS, the first machine learning-based framework for generating LDS. Drawing inspiration from classical LDS, we train a neural network to map indices to points such that the resulting sequences exhibit minimal discrepancy across all prefixes. To this end, we deploy a two-stage learning process: supervised approximation of classical constructions followed by unsupervised fine-tuning to minimize prefix discrepancies. We demonstrate that NeuroLDS outperforms all previous LDS constructions by a significant margin with respect to discrepancy measures. Moreover, we demonstrate the effectiveness of NeuroLDS across diverse applications, including numerical integration, robot motion planning, and scientific machine learning. These results highlight the promise and broad significance of Neural Low-Discrepancy Sequences.

","absKo":"

저차이산점(low-discrepancy points)은 공간을 균일한 방식으로 효율적으로 채우도록 설계된다. 이러한 균일성은 numerical integration, computer vision, machine perception, computer graphics, machine learning, simulation을 비롯한 과학 및 공학의 많은 문제에서 매우 유리하다. 기존의 대부분의 저차이산 구성은 추상대수와 수론에 의존하는 반면, 최근 Message-Passing Monte Carlo (MPMC)는 이전에 가능했던 것보다 더 낮은 discrepancy를 갖는 point set을 생성하기 위해 machine learning 방법을 활용하도록 제안되었다. 그러나 MPMC는 point set 생성에만 한정되며, 모든 prefix가 low discrepancy를 갖는 점들의 sequence, 즉 많은 응용에서 필수적인 성질인 low-discrepancy sequence (LDS)로는 확장될 수 없다. 이러한 한계를 해결하기 위해, 우리는 Neural Low-Discrepancy Sequences NeuroLDS를 제안한다. 이는 LDS 생성을 위한 최초의 machine learning 기반 framework이다. 고전적인 LDS에서 영감을 받아, 우리는 neural network를 학습시켜 index를 point로 매핑하고, 그 결과 sequence가 모든 prefix에 걸쳐 최소 discrepancy를 보이도록 한다. 이를 위해 우리는 two-stage learning process를 사용한다: 먼저 고전적 구성을 supervised approximation하고, 이어서 prefix discrepancy를 최소화하기 위해 unsupervised fine-tuning을 수행한다. 우리는 NeuroLDS가 discrepancy 측도 측면에서 기존의 모든 LDS 구성보다 유의미한 폭으로 우수함을 보인다. 더 나아가, numerical integration, robot motion planning, scientific machine learning을 포함한 다양한 응용에서 NeuroLDS의 효과를 입증한다. 이러한 결과는 Neural Low-Discrepancy Sequences의 가능성과 광범위한 중요성을 보여준다.

"},{"id":"63263","en":"Optimal Quantum Speedups for Repeatedly Nested Expectation Estimation","ko":"반복 중첩 기대값 추정을 위한 최적 양자 가속","authors":"Yihang Sun, Guanyang Wang, Jose Blanchet","pos":"#3413","link":"https://openreview.net/forum?id=Yd1jFl0jBt","abs":"We study the estimation of repeatedly nested expectations (RNEs) with a constant horizon (number of nestings) using quantum computing. We propose a quantum algorithm that achieves $\\varepsilon$-error with cost $\\tilde O(\\varepsilon^{-1})$, up to logarithmic factors. Standard lower bounds show this scaling is essentially optimal, yielding an almost quadratic speedup over the best classical algorithm. Our results extend prior quantum speedups for single nested expectations to repeated nesting, and therefore cover a broader range of applications, including optimal stopping. This extension requires a new derandomized variant of the classical randomized Multilevel Monte Carlo (rMLMC) algorithm. Careful de-randomization is key to overcoming a variable-time issue that typically increases quantized versions of classical randomized algorithms.","absKo":"우리는 양자 컴퓨팅을 사용하여 고정된 horizon(중첩 횟수)을 갖는 반복 중첩 기대값(repeatedly nested expectations, RNEs)의 추정을 연구한다. 우리는 로그 요인을 제외하고 비용이 $\\tilde O(\\varepsilon^{-1})$인 $\\varepsilon$-오차를 달성하는 양자 알고리즘을 제안한다. 표준 하한은 이 스케일링이 본질적으로 최적임을 보여주며, 이는 최선의 고전 알고리즘에 비해 거의 2차적인 속도 향상을 의미한다. 우리의 결과는 단일 nested expectation에 대한 기존 양자 가속을 반복 중첩으로 확장하며, 따라서 optimal stopping을 포함한 더 넓은 응용 범위를 포괄한다. 이러한 확장에는 고전적 randomized Multilevel Monte Carlo(rMLMC) 알고리즘의 새로운 derandomized 변형이 필요하다. 정교한 비무작위화는 보통 고전적 randomized 알고리즘의 양자화된 버전에서 계산량을 증가시키는 variable-time 문제를 극복하는 핵심이다."},{"id":"61932","en":"Conditional Diffusion Sampling","ko":"조건부 확산 샘플링","authors":"Francisco M Castro-Macías, Pablo Morales-Alvarez, Saifuddin Syed, Daniel Hernández-Lobato, Rafael Molina, Jose Miguel Hernandez-Lobato","pos":"#3500","link":"https://openreview.net/forum?id=lZxQyCh5H7","abs":"

Sampling from unnormalized multimodal distributions with limited density evaluations remains a fundamental challenge in machine learning and natural sciences. Successful approaches construct a bridge between a tractable reference and the target distribution. Parallel Tempering (PT) serves as the gold standard, while recent diffusion-based approaches offer a continuous alternative at the cost of neural training. In this work, we introduce Conditional Diffusion Sampling (CDS), a framework that combines these two paradigms. To this end, we derive Conditional Interpolants, a class of stochastic processes whose transport dynamics are governed by an exact, closed-form stochastic differential equation (SDE), requiring no neural approximation. Although these dynamics require sampling from a non-trivial initialization distribution, we show both theoretically and empirically that the cost of this initialization diminishes for sufficiently short diffusion times. CDS leverages this by a two-stage procedure: (1) PT is used to efficiently sample the initial distribution, and then (2) samples are transported via the transport SDE. This combination couples the robust global exploration of PT with efficient local transport. Experiments suggest that CDS has the potential to achieve a superior trade-off between sample quality and density evaluation cost compared to state-of-the-art samplers.

","absKo":"

제한된 density evaluation만 가능한 unnormalized multimodal distribution에서 샘플링하는 문제는 머신러닝과 자연과학 전반에서 근본적인 도전 과제이다. 성공적인 접근은 계산 가능한 reference 분포와 target 분포 사이에 bridge를 구축한다. Parallel Tempering(PT)은 gold standard로 쓰이며, 최근 diffusion 기반 접근은 neural training이라는 비용을 치르는 대신 연속적인 대안을 제공한다. 본 연구에서는 이 두 패러다임을 결합한 Conditional Diffusion Sampling(CDS) 프레임워크를 제안한다. 이를 위해 우리는 transport dynamics가 정확한 closed-form stochastic differential equation(SDE)로 기술되며 neural approximation을 필요로 하지 않는 stochastic process의 한 종류인 Conditional Interpolant를 유도한다. 이러한 dynamics는 비자명한 initialization distribution에서의 샘플링을 요구하지만, 우리는 이 초기화 비용이 충분히 짧은 diffusion time에서는 이론적으로나 실험적으로나 감소함을 보인다. CDS는 이를 두 단계 절차로 활용한다: (1) PT를 사용해 초기 분포를 효율적으로 샘플링한 뒤, (2) transport SDE를 통해 sample을 전달한다. 이 조합은 PT의 강건한 전역 탐색과 효율적인 local transport를 결합한다. 실험 결과 CDS는 state-of-the-art sampler 대비 sample quality와 density evaluation cost 사이에서 더 우수한 trade-off를 달성할 가능성이 있음을 시사한다.

"},{"id":"64410","en":"Computationally-efficient Graph Modeling with Refined Graph Random Features","ko":"정제된 그래프 랜덤 특징을 이용한 계산 효율적 그래프 모델링","authors":"Krzysztof Choromanski, Kumar Avinava Dubey, Arijit Sehanobish, Isaac Reid","pos":"#3501","link":"https://openreview.net/forum?id=NvJPE1oiKd","abs":"

We propose refined GRFs (GRFs++), a new class of Graph Random Features (GRFs) for efficient and accurate computations involving kernels defined on the nodes of a graph. GRFs++ resolve some of the long-standing limitations of regular GRFs, including difficulty modeling relationships between more distant nodes. They reduce dependence on sampling long graph random walks via a novel walk-stitching technique, concatenating several shorter walks without breaking unbiasedness. By applying these techniques, GRFs++ inherit the approximation quality provided by longer walks but with greater efficiency, trading sequential inefficient sampling of a long walk for parallel computation of short walks and matrix-matrix multiplication. Furthermore, GRFs++ extend the simplistic GRFs walk termination mechanism (Bernoulli schemes with fixed halting probabilities) to a broader class of strategies, applying general distributions on the walks' lengths. This improves approximation accuracy of graph kernels, without incurring extra computational cost. We provide empirical evaluations to showcase our claims and complement our results with theoretical analysis.

","absKo":"

우리는 그래프의 노드에 정의된 kernel을 포함하는 효율적이고 정확한 계산을 위해, 새로운 Graph Random Features(GRFs) 계열인 refined GRFs(GRFs++)를 제안한다. GRFs++는 더 먼 노드들 사이의 관계를 모델링하기 어렵다는 점을 포함하여, 기존 GRFs의 오랜 한계 일부를 해소한다. 이들은 새로운 walk-stitching 기법을 통해 긴 graph random walk를 샘플링하는 데 대한 의존을 줄이며, unbiasedness를 깨지 않으면서 여러 개의 짧은 walk를 연결한다. 이러한 기법을 적용함으로써 GRFs++는 더 긴 walk가 제공하는 근사 품질을 더 높은 효율로 계승하고, 긴 walk의 순차적이고 비효율적인 샘플링을 짧은 walk의 병렬 계산과 matrix-matrix multiplication으로 대체한다. 더 나아가 GRFs++는 고정된 halting probability를 갖는 Bernoulli scheme인 기존 GRFs의 단순한 walk 종료 메커니즘을 확장하여, walk 길이에 대해 일반 분포를 적용하는 더 넓은 전략 집합으로 확장한다. 이는 추가적인 계산 비용 없이 graph kernel의 근사 정확도를 향상시킨다. 우리는 이러한 주장을 보이기 위해 실증적 평가를 제공하고, 이론적 분석으로 결과를 보완한다.

"},{"id":"65127","en":"Accelerating Langevin Monte Carlo via Efficient Stochastic Runge-Kutta Methods beyond Log-Concavity","ko":"Log-Concavity를 넘어선 효율적 확률적 Runge-Kutta 방법을 통한 Langevin Monte Carlo 가속","authors":"Bin Yang, Xiaojie Wang","pos":"#3502","link":"https://openreview.net/forum?id=GakFNj5L2W","abs":"Sampling from a high-dimensional probability distribution is a fundamental algorithmic task arising in wide-ranging applications across multiple disciplines, including scientific computing, computational statistics and machine learning. Langevin Monte Carlo (LMC) algorithms are among the most widely used sampling methods in high dimensional settings. This paper introduces a novel higher-order and Hessian-free LMC sampling algorithm based on an efficient stochastic Runge-Kutta method of strong order $1.5$ for the overdamped Langevin dynamics. In contrast to the existing Runge-Kutta type LMC (Li et al.,2019) involved with three gradient evaluations, the newly proposed algorithm is computationally cheaper and requires only two gradient evaluations for one iteration. Under certain log-smooth conditions, non-asymptotic error bounds of the proposed algorithms are analyzed in $\\mathcal{W}_2$-distance. In particular, a uniform-in-time convergence rate of order $O(d ^{\\frac32} h^{\\frac32})$ is derived in a non-log-concave setting, matching the convergence rate proved in the aforementioned work but under the log-concavity condition. Numerical experiments are finally presented to demonstrate the effectiveness of the new sampling algorithm.","absKo":"고차원 probability distribution에서 sampling하는 것은 scientific computing, computational statistics, machine learning 등 여러 분야의 다양한 application에서 등장하는 근본적인 algorithmic task이다. Langevin Monte Carlo (LMC) algorithm은 고차원 setting에서 가장 널리 사용되는 sampling method 중 하나이다. 이 논문은 overdamped Langevin dynamics를 위한 strong order $1.5$의 효율적인 stochastic Runge-Kutta method에 기반한 새로운 higher-order 및 Hessian-free LMC sampling algorithm을 소개한다. gradient evaluation을 세 번 사용하는 기존의 Runge-Kutta type LMC (Li et al.,2019)와 달리, 새롭게 제안한 algorithm은 계산 비용이 더 낮고 한 번의 iteration에 gradient evaluation 두 번만 필요하다. 일정한 log-smooth condition 하에서, 제안한 algorithm의 non-asymptotic error bound를 $\\mathcal{W}_2$-distance에서 분석한다. 특히 non-log-concave setting에서 시간에 대해 uniform한 convergence rate $O(d ^{\\frac32} h^{\\frac32})$를 도출하며, 이는 앞서 언급한 work에서 log-concavity condition 하에 증명된 convergence rate와 일치한다. 마지막으로 수치 실험을 통해 새로운 sampling algorithm의 효과를 보인다."},{"id":"64072","en":"Accelerated Multiple Wasserstein Gradient Flows for Multi-objective Distributional Optimization","ko":"다목적 분포 최적화를 위한 가속된 다중 Wasserstein 그래디언트 흐름","authors":"DaiHai Nguyen, Duc-Dung NGUYEN, Atsuyoshi Nakamura, Hiroshi Mamitsuka","pos":"#3702","link":"https://openreview.net/forum?id=RAkOcnP4Uz","abs":"We study multi-objective optimization over probability distributions in Wasserstein space. Recently, \\citet{nguyen2025multiple} introduced Multiple Wasserstein Gradient Descent (MWGraD) algorithm, which exploits the geometric structure of Wasserstein space to jointly optimize multiple objectives. Building on this approach, we propose an accelerated variant, A-MWGraD, inspired by Nesterov's acceleration. We analyze the continuous-time dynamics and establish convergence to weakly Pareto optimal points in probability space. Our theoretical results show that A-MWGraD achieves a convergence rate of $\\mathcal{O}(1/t^2)$ for geodesically convex objectives and $\\mathcal{O}(e^{-\\sqrt{\\beta}t})$ for $\\beta$-strongly geodesically convex objectives, improving upon the $\\mathcal{O}(1/t)$ rate of MWGraD in the geodesically convex setting. We further introduce a practical kernel-based discretization for A-MWGraD and demonstrate through numerical experiments that it consistently outperforms MWGraD in convergence speed and sampling efficiency on multi-target sampling tasks.","absKo":"우리는 Wasserstein space에서 확률 분포에 대한 multi-objective optimization을 연구한다. 최근 \\citet{nguyen2025multiple}은 Wasserstein space의 기하학적 구조를 활용하여 여러 objective를 공동으로 최적화하는 Multiple Wasserstein Gradient Descent(MWGraD) 알고리즘을 제안하였다. 이 접근을 바탕으로, 우리는 Nesterov acceleration에서 영감을 받은 가속화 변형 A-MWGraD를 제안한다. 우리는 continuous-time dynamics를 분석하고, probability space에서 weakly Pareto optimal point로의 수렴을 확립한다. 우리의 이론 결과는 A-MWGraD가 geodesically convex objective에 대해 $\\mathcal{O}(1/t^2)$의 수렴률을, $\\beta$-strongly geodesically convex objective에 대해 $\\mathcal{O}(e^{-\\sqrt{\\beta}t})$의 수렴률을 달성함을 보여주며, 이는 geodesically convex 설정에서 MWGraD의 $\\mathcal{O}(1/t)$ rate보다 개선된 것이다. 또한 우리는 A-MWGraD를 위한 실용적인 kernel-based discretization을 도입하고, 수치 실험을 통해 multi-target sampling task에서 수렴 속도와 sampling efficiency 측면에서 MWGraD보다 일관되게 우수함을 보인다."},{"id":"62903","en":"Complexity bounds for Dirichlet process slice samplers","ko":"Dirichlet process slice sampler의 복잡도 상한","authors":"Beatrice Franzolini, Francesco Gaffi","pos":"#412","link":"https://openreview.net/forum?id=c2neBfCuoz","abs":"Slice sampling is a standard Monte Carlo technique for Dirichlet process (DP)--based models, widely used in posterior simulation. However, formal assessments of the scalability of posterior slice samplers have remained largely unexplored, primarily because the computational cost of a slice-sampling iteration is random and potentially unbounded. In this work, we obtain high-probability bounds on the computational complexity of DP slice samplers. Our main results show that, uniformly across posterior cluster-growth regimes, the overhead induced by slice variables, relatively to the number of clusters supported by the posterior, is $O_{\\mathbb P}(\\log n)$. As a consequence, even in worst-case configurations, superlinear blow-ups in per-iteration computational cost occur with vanishing probability. Our analysis applies broadly to DP--based models without any likelihood-specific assumptions, still providing complexity guarantees for posterior sampling on arbitrary datasets. These results establish a theoretical foundation for assessing the practical scalability of slice sampling in DP-based models.","absKo":"Slice sampling은 Dirichlet process (DP) 기반 model에서 표준적인 Monte Carlo technique이며, posterior simulation에서 널리 사용된다. 그러나 posterior slice sampler의 scalability에 대한 formal assessment는 계산 비용이 random이고 잠재적으로 unbounded이기 때문에 대체로 충분히 탐구되지 않았다. 이 연구에서 우리는 DP slice sampler의 computational complexity에 대한 high-probability bound를 얻는다. 우리의 주요 결과는 posterior cluster-growth regime 전반에 걸쳐, posterior가 지지하는 cluster 수에 비해 slice variable이 유발하는 오버헤드는 $O_{\\mathbb P}(\\log n)$임을 보인다. 결과적으로, 최악의 구성에서도 iteration당 계산 비용의 superlinear blow-up는 사라질 확률을 가진다. 우리의 분석은 likelihood-specific assumption 없이 DP 기반 model 전반에 폭넓게 적용되며, 임의의 dataset에 대한 posterior sampling에 대해 여전히 complexity guarantee를 제공한다. 이 결과는 DP 기반 model에서 slice sampling의 실제 scalability를 평가하기 위한 이론적 기반을 확립한다."},{"id":"64251","en":"The Two-Hump Problem: Bridging the Difficulty Gap in Mathematical Reinforcement Learning","ko":"두 봉우리 문제: 수학적 강화학습에서 난이도 격차 해소","authors":"Lucas Fagan, Michele Tarquini, Ali Shehper, Maksymilian Manko, Angus Gruen, Coco Huang, Giorgi Butbaia, Davide Passaro, Sergei Gukov","pos":"#124","link":"https://openreview.net/forum?id=PUQ0rrmuDM","abs":"

Mathematical search problems present a unique challenge for Reinforcement Learning (RL) due to vast search spaces and sparse rewards. In previous works, the Andrews-Curtis (AC) conjecture was established as an illustrative example of such problems. In this work, we identify a critical structural barrier in the AC landscape: a \"Two Hump\" distribution, where problem instances are either trivially solvable or effectively impossible, with a scarcity of intermediate \"hard-but-solvable\" instances required for effective learning. We tackle this challenge through two primary avenues: novel data generation techniques to populate the difficulty gap, and significant algorithmic enhancements including the introduction of supermoves and Transformer-based architectures. We demonstrate substantial performance improvements over previous baselines, and release new comprehensive benchmark datasets including AC-19 (125,192 AC-trivial presentations of varying difficulty with length at most 19) and AC-1M (1,136,154 hard AC-trivial presentations of length at most 30), the first large-scale, publicly available datasets of this kind.

","absKo":"수학적 검색 문제는 방대한 search space와 sparse reward 때문에 Reinforcement Learning (RL)에 독특한 도전을 제시한다. 이전 연구에서는 Andrews-Curtis (AC) conjecture가 이러한 문제의 예시로 제시되었다. 본 연구에서는 AC landscape에서 중요한 구조적 장벽을 식별한다. 바로 문제 인스턴스가 자명하게 풀리거나 사실상 불가능한 두 개의 봉우리(\"Two Hump\") 분포를 보이며, 효과적인 학습에 필요한 중간 단계의 \"hard-but-solvable\" 인스턴스가 희소하다는 점이다. 우리는 이 difficulty gap을 채우기 위한 새로운 data generation 기법과, supermove 및 Transformer-based architecture 도입을 포함한 중요한 algorithmic enhancement라는 두 가지 주요 경로로 이 문제에 대응한다. 우리는 이전 baseline 대비 상당한 성능 향상을 보이며, 길이 19 이하의 서로 다른 난이도를 가진 125,192개의 AC-trivial presentation을 포함하는 AC-19와 길이 30 이하의 1,136,154개의 hard AC-trivial presentation을 포함하는 AC-1M 등 새로운 포괄적 benchmark dataset을 공개한다. 이는 이 종류의 최초의 대규모 공개 데이터셋이다.

"},{"id":"63364","en":"CoRe: Collaborative Reasoning via Cross Teaching","ko":"CoRe: 교차 교습을 통한 협력적 추론","authors":"Kshitij Mishra, Mirat Aubakirov, Martin Takac, Nils Lukas, Salem Lahlou","pos":"#109","link":"https://openreview.net/forum?id=XjkpSHzysD","abs":"

Large language models exhibit complementary reasoning errors: on the same instance, one model may succeed with a particular decomposition while another fails. We propose Collaborative Reasoning (CORE), a training-time collaboration framework that converts peer success into a learning signal via a cross-teaching protocol. Each problem is solved in two stages: a cold round of independent sampling, followed by a contexted rescue round in which models that failed receive hint extracted from a successful peer. CORE optimizes a combined reward that balances (i) correctness, (ii) a lightweight DPP-inspired diversity term to reduce error overlap, and (iii) an explicit rescue bonus for successful recovery. We evaluate CORE across four standard reasoning datasets GSM8K, MATH, AIME, and GPQA. With only 1,000 training examples, a pair of small open source models (3B+4B) reaches Pass@2 of 99.54% on GSM8K and 92.08% on MATH, compared to 82.50% and 74.82% for single-model training. On harder datasets, the 3B+4B pair reaches Pass@2 of 77.34% on GPQA (trained on 348 examples) and 79.65% on AIME (trained on 792 examples), using a training-time budget of at most 1536 context tokens and 3072 generated tokens. Overall, these results show that training-time collaboration can reliably convert model complementarity into large gains without scaling model size.

","absKo":"

Large language models는 상호 보완적인 추론 오류를 보인다. 같은 인스턴스에서 한 모델은 특정 decomposition에서는 성공하지만 다른 모델은 실패할 수 있다. 우리는 Collaborative Reasoning(CORE)을 제안한다. 이는 peer의 성공을 학습 신호로 전환하는 cross-teaching protocol을 활용하는 training-time collaboration framework이다. 각 문제는 두 단계로 해결된다. 먼저 독립적으로 샘플링하는 cold round를 수행하고, 이어서 실패한 모델이 성공한 peer로부터 추출한 hint를 받는 contexted rescue round를 수행한다. CORE는 (i) 정답성, (ii) error overlap을 줄이기 위한 경량 DPP-inspired diversity term, (iii) 성공적인 복구에 대한 명시적 rescue bonus를 균형 있게 반영한 결합 reward를 최적화한다. 우리는 CORE를 네 개의 표준 reasoning 데이터셋 GSM8K, MATH, AIME, GPQA 전반에서 평가한다. 단 1,000개의 training example만으로도, 작은 open source model 두 개(3B+4B) 조합은 GSM8K에서 Pass@2 99.54%, MATH에서 92.08%에 도달하며, 이는 single-model training의 82.50%와 74.82%를 크게 상회한다. 더 어려운 데이터셋에서는 3B+4B 조합이 GPQA에서 Pass@2 77.34%(348개 예시로 학습), AIME에서 79.65%(792개 예시로 학습)를 달성했으며, training-time budget은 최대 1536 context token과 3072 generated token이었다. 전반적으로 이러한 결과는 training-time collaboration이 model size를 키우지 않고도 model complementarity를 안정적으로 큰 성능 향상으로 전환할 수 있음을 보여준다.

"},{"id":"62268","en":"Multi-view Consistent Latent Action Learning for World Modeling and Control","ko":"World Modeling 및 제어를 위한 다중 뷰 일관성 잠재행동 학습","authors":"Shenghua Wan, Xiaohai Hu, Xunlan Zhou, lei yuan, Le Gan, De-Chuan Zhan","pos":"#110","link":"https://openreview.net/forum?id=i8Ob0TMeMP","abs":"

The scalability of world models is currently bottlenecked by the scarcity of action annotations. While self-supervised latent action learning offers a potential solution, existing single-view paradigms—relying on information bottlenecks or Vector Quantization (VQ)—often conflate superficial 2D pixel displacements with the underlying physical-spatial dynamics of an action. Consequently, these methods remain highly susceptible to view-dependent noise, such as camera shake. We introduce MuCoLA (Multi-view Consistent Latent Action learning), a framework that learns robust, view-invariant action representations by enforcing semantic consistency across synchronized video streams. MuCoLA utilizes a Student-Teacher network with DINO-style self-distillation to align action distributions across viewpoints, effectively filtering high-frequency visual noise while preserving motion semantics. Theoretical analysis reveals that our multi-view objective functions as a spectral filter, isolating agent dynamics from environmental nuisances. Empirically, MuCoLA significantly outperforms baselines in action regression, video reconstruction, and downstream visual control tasks. Furthermore, we demonstrate that MuCoLA exhibits favorable scaling properties with respect to model capacity and data volume, paving the way for large-scale action-free world modeling.

","absKo":"

world model의 확장성은 현재 action annotation의 희소성에 의해 병목되고 있다. self-supervised latent action learning이 잠재적 해법을 제공하지만, 기존의 single-view 패러다임은 information bottleneck이나 Vector Quantization (VQ)에 의존하는 경우가 많아, 표면적인 2D pixel displacement를 action의 근본적인 물리적-공간적 dynamics와 혼동한다. 그 결과 이러한 방법은 camera shake와 같은 view-dependent noise에 매우 취약하다. 우리는 동기화된 video stream 전반에 semantic consistency를 강제함으로써 견고하고 view-invariant한 action representation을 학습하는 MuCoLA (Multi-view Consistent Latent Action learning)를 제안한다. MuCoLA는 DINO-style self-distillation을 사용하는 Student-Teacher network를 활용하여 viewpoint 간 action distribution을 정렬하고, motion semantics는 보존하면서 고주파 visual noise를 효과적으로 제거한다. 이론 분석은 우리의 multi-view objective가 spectral filter처럼 작동하여 agent dynamics를 환경적 잡음으로부터 분리함을 보여준다. 실험적으로 MuCoLA는 action regression, video reconstruction, downstream visual control task에서 baseline을 크게 능가한다. 또한 MuCoLA가 model capacity와 data volume에 대해 유리한 scaling property를 보임을 보여주어, 대규모 action-free world modeling의 길을 연다.

"},{"id":"63244","en":"Dream-MPC: Gradient-Based Model Predictive Control with Latent Imagination","ko":"Dream-MPC: 잠재 상상을 활용한 경사 기반 모델 예측 제어","authors":"Jonathan Spieler, Sven Behnke","pos":"#111","link":"https://openreview.net/forum?id=Ym6ud8Psx1","abs":"

State-of-the-art model-based Reinforcement Learning (RL) approaches either use gradient-free, population-based methods for planning, learned policy networks, or a combination of policy networks and planning. Hybrid approaches that combine Model Predictive Control (MPC) with a learned model and a policy prior to leverage the advantages of both paradigms have shown promising results. However, these approaches typically rely on gradient-free optimization methods, which can be computationally expensive for high-dimensional control tasks. While gradient-based methods are a promising alternative, recent works have empirically shown that gradient-based methods often perform worse than their gradient-free counterparts. We propose Dream-MPC, a novel approach that generates few candidate trajectories from a rolled-out policy and optimizes each trajectory by gradient ascent using a learned world model, uncertainty regularization and amortization of optimization iterations over time by reusing previously optimized actions. Our results on 24 continuous control tasks show that Dream-MPC can significantly improve the performance of the underlying policy and can outperform gradient-free MPC and state-of-the-art baselines. We will open source our code and more at https://dream-mpc.github.io.

","absKo":"

state-of-the-art model-based Reinforcement Learning (RL) 접근법은 planning에 gradient-free population-based method를 사용하거나, learned policy network를 사용하거나, 혹은 policy network와 planning을 결합한 방식을 사용한다. Model Predictive Control (MPC)과 learned model, 그리고 policy prior를 결합하여 두 패러다임의 장점을 모두 활용하는 hybrid approach는 유망한 결과를 보여 왔다. 그러나 이러한 접근은 대체로 gradient-free optimization method에 의존하며, 이는 고차원 control task에서 계산 비용이 클 수 있다. Gradient-based method는 유망한 대안이지만, 최근 연구들은 경험적으로 gradient-based method가 gradient-free counterpart보다 종종 더 낮은 성능을 보인다고 보고했다. 우리는 Dream-MPC를 제안한다. Dream-MPC는 rollout된 policy로부터 소수의 candidate trajectory를 생성하고, learned world model, uncertainty regularization, 그리고 이전에 최적화한 action을 재사용하여 시간에 걸쳐 optimization iteration을 amortize하는 방식으로 각 trajectory를 gradient ascent로 최적화한다. 24개의 continuous control task에서의 결과는 Dream-MPC가 기반 policy의 성능을 크게 향상시키고, gradient-free MPC 및 state-of-the-art baseline을 능가할 수 있음을 보여 준다. 코드와 추가 자료는 https://dream-mpc.github.io 에서 공개할 예정이다.

"},{"id":"65627","en":"Active Curriculum Refinement for Reinforcement Learning","ko":"강화학습을 위한 능동적 커리큘럼 정제","authors":"Zhenya Liu, Yuxin Chen","pos":"#112","link":"https://openreview.net/forum?id=BhS108cyjT","abs":"

In many RL domains, environments are linked by prerequisite relations—e.g., difficulty-increasing edits or parameter increments—which induce a directed acyclic curriculum graph (DAG). In practice, this structure is often exploited only implicitly, yet it can yield clear gains in training. We introduce PATH, a curriculum learning framework that performs active learning on the curriculum graph. PATH first expands coverage by sampling diverse curriculum paths, then reallocates training toward regions that remain unmastered. Experiments show that PATH leverages the graph structure to achieve strong robustness and generalization across diverse environments.

","absKo":"

많은 RL 도메인에서 환경은 prerequisite 관계로 연결되어 있다. 예를 들어 difficulty를 높이는 edit이나 parameter increment가 그러하며, 이는 directed acyclic curriculum graph (DAG)를 유도한다. 실제로는 이 구조가 종종 암묵적으로만 활용되지만, 학습에서 분명한 이득을 가져올 수 있다. 우리는 curriculum graph 위에서 active learning을 수행하는 curriculum learning framework인 PATH를 제안한다. PATH는 먼저 다양한 curriculum path를 sampling하여 coverage를 확장한 뒤, 아직 숙달되지 않은 영역으로 training을 재배치한다. 실험 결과 PATH는 graph 구조를 활용하여 다양한 환경에서 강한 robustness와 generalization을 달성함을 보여준다.

"},{"id":"60802","en":"Transferable Reinforcement Learning via Probabilistic Latent Embeddings and Dynamic Policy Adaptation for Sim-to-Real Deployment","ko":"Sim-to-Real 배포를 위한 Probabilistic Latent Embedding과 Dynamic Policy Adaptation 기반 Transferable Reinforcement Learning","authors":"Gengyue Han, Yiheng Feng","pos":"#113","link":"https://openreview.net/forum?id=wZMdhuTXoz","abs":"

Due to limited resources and public safety concerns, deep reinforcement learning (RL) agents for many cyber-physical systems (e.g., autonomous vehicles) are first trained in simulators. However, when deployed in real world environments, they often suffer from performance degradation or safety violations because of the inevitable \\textit{Sim2Real} gap. Existing zero-shot approaches, such as robust safe RL and domain randomization, mitigate this issue but typically at the cost of degraded performance or residual safety risks when experiencing unmodeled system dynamics. To address these limitations, we propose a novel reinforcement learning framework that enables safe and efficient policy transfer via probabilistic latent embeddings and dynamic policy adaptation. We consider a family of Constrained Markov Decision Processes (CMDPs) under different environment contexts. By leveraging probabilistic latent context adaptation in meta-RL, the proposed framework infers the latent representation of the environment from simulated experiences. Furthermore, it incorporates a distributional RL formulation, which allows risk levels of the deployed policy to be adjusted dynamically at inference time, based on the estimation accuracy of the latent context variable. This strategy promotes safety at the early deployment stage and improves efficiency through fast policy adaptation under the Sim2Real gap.

","absKo":"

제한된 자원과 공공 안전에 대한 우려로 인해, autonomous vehicle과 같은 많은 cyber-physical system을 위한 deep reinforcement learning(RL) agent는 먼저 simulator에서 학습된다. 그러나 실제 환경에 배치되면, 불가피한 \\textit{Sim2Real} gap 때문에 성능 저하 또는 안전 위반을 겪는 경우가 많다. robust safe RL과 domain randomization 같은 기존 zero-shot 접근은 이 문제를 완화하지만, 대체로 모델링되지 않은 system dynamics를 경험할 때 성능 저하나 잔존하는 안전 위험을 감수해야 한다. 이러한 한계를 해결하기 위해, 우리는 probabilistic latent embedding과 dynamic policy adaptation을 통해 안전하고 효율적인 policy transfer를 가능하게 하는 새로운 reinforcement learning framework를 제안한다. 우리는 서로 다른 environment context를 갖는 Constrained Markov Decision Process(CMDP) family를 고려한다. meta-RL의 probabilistic latent context adaptation을 활용하여, 제안 프레임워크는 시뮬레이션 경험으로부터 환경의 latent representation을 추론한다. 또한 distributional RL formulation을 포함하여, latent context variable의 추정 정확도에 따라 inference time에 deployed policy의 risk level을 동적으로 조절할 수 있게 한다. 이 전략은 초기 배포 단계에서 안전성을 높이고, Sim2Real gap 하에서 빠른 policy adaptation을 통해 효율성을 향상시킨다.

"},{"id":"64665","en":"TaskLoom: Weaving Knowledge Across Tasks in World Models","ko":"TaskLoom: World Model에서 과제 간 지식 엮기","authors":"Qingzhang Zeng, Peixi Peng, hang li, Luntong Li, Yonghong Tian","pos":"#114","link":"https://openreview.net/forum?id=L6Vw6CvKHi","abs":"

Sample efficiency remains one of the central challenges in modern deep reinforcement learning (DRL). In recent years, world model approaches have significantly improved the sample efficiency of model-based reinforcement learning (MBRL) by enabling policy learning in imagination, thereby reducing the need for direct interaction with the real environment. However, most existing world model methods are trained independently for each task or perform multi-task learning in offline settings, failing to fully exploit the latent relationships among tasks in online interactive scenarios. To address this limitation, we propose TaskLoom, a knowledge-sharing world model architecture for online reinforcement learning. TaskLoom adopts a grouped two-stage training paradigm: in the first stage, fine-grained knowledge is shared among tasks within each group, while in the second stage, coarse-grained knowledge is exchanged across groups, enabling hierarchical knowledge transfer and reuse. Experimental results show that TaskLoom outperforms baseline methods on widely used benchmarks such as Proprio Control and Visual Control, validating the effectiveness of the proposed knowledge-sharing mechanism for both low-dimensional state and high-dimensional visual inputs.

","absKo":"

샘플 효율성은 현대 deep reinforcement learning(DRL)의 핵심 과제 중 하나로 남아 있다. 최근 몇 년간 world model 접근법은 model-based reinforcement learning(MBRL)의 샘플 효율성을 크게 향상시켰는데, 이는 imagination에서 policy learning을 가능하게 하여 실제 환경과의 직접 상호작용 필요성을 줄였기 때문이다. 그러나 기존의 대부분 world model 방법은 각 task별로 독립적으로 학습되거나 offline setting에서 multi-task learning을 수행할 뿐, online interactive scenario에서 task들 사이의 잠재적 관계를 충분히 활용하지 못한다. 이러한 한계를 해결하기 위해, 우리는 online reinforcement learning을 위한 knowledge-sharing world model architecture인 TaskLoom을 제안한다. TaskLoom은 grouped two-stage training paradigm을 채택한다. 첫 번째 stage에서는 각 group 내 task들 사이에 fine-grained knowledge를 공유하고, 두 번째 stage에서는 group 간 coarse-grained knowledge를 교환하여 hierarchical knowledge transfer와 reuse를 가능하게 한다. 실험 결과 TaskLoom은 Proprio Control과 Visual Control 같은 널리 사용되는 benchmark에서 baseline method를 능가하며, low-dimensional state와 high-dimensional visual input 모두에서 제안한 knowledge-sharing mechanism의 효과를 검증한다.

"},{"id":"63318","en":"Scalable Traffic Signal Control with Shared Policy Framework","ko":"공유 정책 프레임워크를 활용한 확장 가능한 교통 신호 제어","authors":"Haolun MA, Yanchen ZHU, Zizhuo Xu, Weijie Shi, Jiajie Xu, Lei Li","pos":"#115","link":"https://openreview.net/forum?id=Y8dBzpAdkt","abs":"

Learning-based Traffic Signal Control (TSC) achieves satisfactory performance in small networks, but its effectiveness often deteriorates in larger networks under dynamic traffic patterns and intersection heterogeneity. In this work, we propose SLight, a policy-aware grouped MARL-TSC framework that enables scalability and efficiency balance under dynamic and heterogeneous traffic conditions. SLight captures policy-influenced traffic patterns with a policy-aware traffic pattern encoder, learns explicit group-level shared control principles from state–action trajectories, and matches each intersection’s traffic pattern embedding to principle prototypes flexibly through a compatibility-based adaptive assignment module. Experiments on real-world and synthetic networks demonstrate that SLight sustains performance gains as scale increases and outperforms existing rule-based, reinforcement learning, and grouping-based baselines. Code is available at \\url{https://anonymous.4open.science/r/code-20D3/}

","absKo":"

Learning-based Traffic Signal Control (TSC)은 작은 network에서는 만족스러운 성능을 달성하지만, dynamic traffic pattern과 intersection heterogeneity가 존재하는 큰 network에서는 그 효과가 종종 저하된다. 본 연구에서는 동적이고 이질적인 traffic condition 하에서 scalability와 efficiency의 균형을 가능하게 하는 policy-aware grouped MARL-TSC framework인 SLight를 제안한다. SLight는 policy-aware traffic pattern encoder를 통해 policy의 영향을 받은 traffic pattern을 포착하고, state–action trajectory로부터 명시적인 group-level shared control principle을 학습하며, compatibility-based adaptive assignment module을 통해 각 intersection의 traffic pattern embedding을 principle prototype에 유연하게 매칭한다. 실제 및 합성 network에서 수행한 실험은 SLight가 규모가 커져도 성능 향상을 유지하며, 기존의 rule-based, reinforcement learning, grouping-based baseline보다 우수함을 보여 준다. Code는 \\url{https://anonymous.4open.science/r/code-20D3/}에서 제공된다.

"},{"id":"65120","en":"Learning Transferable Interaction Primitives from Game Videos for Humanoids","ko":"게임 영상에서 휴머노이드를 위한 전이 가능한 상호작용 프리미티브 학습","authors":"Xiangming Zhu, Huayu Deng, Haoran Zhao, Yiwei Hao, Yunbo Wang","pos":"#117","link":"https://openreview.net/forum?id=Gd2S0bJqNZ","abs":"

Learning humanoid control from video provides a scalable alternative to the scarcity of high-fidelity robot data. Existing methods, however, often rely on curated datasets and treat video as passive kinematic priors. They fail to capture dynamic humanoid interactions with the environment, which are essential for real-world deployment. To address this, we propose TRansferable Interaction Primitives (TRIP), a framework designed to extract and ground interactions from unstructured, unlabeled game videos for physical controllers. TRIP explicitly models dependencies between motion dynamics and environmental context via a discrete library of interaction-based action primitives. To bridge the reality gap, we introduce a shared context latent space that aligns implicit video-domain features with functional target-domain observations, enabling the seamless transfer of video-mined strategies to reinforcement learning policies. Our experiments on complex terrain navigation demonstrate that TRIP achieves significant improvements in task performance, sample efficiency, and robustness.

","absKo":"비디오로부터 humanoid control을 학습하는 것은 고충실도 robot data의 희소성에 대한 확장 가능한 대안이다. 그러나 기존 방법은 종종 정제된 dataset에 의존하며 비디오를 수동적인 kinematic prior로 취급한다. 이들은 실제 배치에 필수적인 환경과의 동적인 humanoid interaction을 포착하지 못한다. 이를 해결하기 위해 우리는 비정형의 unlabeled game video로부터 상호작용을 추출하고 physical controller에 grounding하기 위해 설계된 프레임워크인 TRansferable Interaction Primitives(TRIP)를 제안한다. TRIP는 discrete library의 interaction-based action primitive를 통해 motion dynamics와 환경 맥락 사이의 의존성을 명시적으로 모델링한다. 현실 간 격차를 메우기 위해, 우리는 암묵적인 video-domain feature와 기능적인 target-domain observation을 정렬하는 shared context latent space를 도입하여, video에서 발굴한 전략을 reinforcement learning policy로 매끄럽게 transfer할 수 있게 한다. 복잡한 지형 내비게이션에 대한 실험은 TRIP가 task performance, sample efficiency, robustness에서 유의미한 향상을 달성함을 보여준다.

"},{"id":"64108","en":"Hierarchical Goal Abstractions via Learned Subset Relations","ko":"학습된 부분집합 관계를 통한 계층적 목표 추상화","authors":"Fabian Wurzberger, Sebastian Gottwald, Zeqiang Zhang, Daniel A Braun","pos":"#118","link":"https://openreview.net/forum?id=Qoh16QHG24","abs":"

In self-supervised goal-conditioned reinforcement learning (RL) without external rewards, goals are typically specified by observations sampled from experience. However, depending on the observation structure, such a fixed representation of goals may be either too concrete (requiring exact pixel-level matches) or too abstract (involving ambiguous observations). Here we propose the construction of hierarchical latent goal spaces that integrate both concrete and abstract goals. To this end, we use an energy function to learn a partially ordered space, in which a subset relation between observations naturally induces a hierarchy from concrete to abstract goals. This representation enables agents to disambiguate specific states while also generalizing to shared concepts. In experiments on navigation and robotic manipulation, agents trained with our hierarchical goal space achieve higher task success and greater generalization to novel tasks compared to agents limited to purely observational goals.

","absKo":"

외부 보상 없이 self-supervised goal-conditioned reinforcement learning(RL)에서 목표는 일반적으로 경험에서 샘플링한 observation으로 지정된다. 그러나 observation 구조에 따라 이러한 고정된 목표 표현은 너무 구체적이어서 정확한 pixel-level match를 요구할 수도 있고, 혹은 너무 추상적이어서 모호한 observation을 포함할 수도 있다. 여기서는 구체적 목표와 추상적 목표를 모두 통합하는 hierarchical latent goal space의 구성을 제안한다. 이를 위해 우리는 energy function을 사용해 부분순서(partially ordered) 공간을 학습하며, 여기서 observation 간 subset relation이 자연스럽게 구체적 목표에서 추상적 목표로의 계층을 유도한다. 이러한 표현은 agent가 특정 상태를 구분해내는 동시에 공유 개념으로 일반화할 수 있게 해준다. navigation 및 robotic manipulation 실험에서, 우리의 hierarchical goal space로 학습한 agent는 순수한 observational goal에만 제한된 agent보다 더 높은 task success와 새로운 task에 대한 더 큰 일반화 성능을 달성한다.

"},{"id":"61274","en":"Closing the Sim-to-Real Gap in Non-Markovian Spreading Processes via GPU-Accelerated Distributional RL","ko":"GPU 가속 Distributional RL로 Non-Markovian 확산 과정의 Sim-to-Real Gap 줄이기","authors":"Heman Shakeri","pos":"#120","link":"https://openreview.net/forum?id=rwTtKedEAa","abs":"Controlling spreading processes on networks such as epidemics, information cascades, product adoption, requires policies that perform on realistic stochastic dynamics, not just tractable approximations. Yet policies trained on standard simplifications (mean-field ODEs, Markovian dynamics) suffer severe performance degradation at deployment. We trace this sim-to-real gap to three theoretical pathologies: Optimism Bias, where deterministic approximations systematically underestimate variance via Jensen's inequality; Hub Blindness, where global state aggregation obscures the super-spreaders driving scale-free networks; and the Valley of Death, where mean-value critics fail to navigate the bimodal nature (extinction vs. viral) of cascade outcomes. We resolve these challenges through two synergistic contributions. First, the Stratified Mean-Field Observer partitions nodes by influence tier, preserving hub dynamics at $O(N)$ cost while producing fixed-dimensional observations that enable zero-shot transfer across network scales and topologies. Second, we demonstrate that Distributional RL via Truncated Quantile Critics is essential for risk-aware control of bimodal cascades. Trained on a GPU-accelerated simulator supporting non-Markovian renewal dynamics, our approach achieves $59\\times$ improvement over Markovian baselines and robust zero-shot transfer to real-world social networks (Facebook, Twitter, YouTube), effectively closing the simulation-to-reality gap.","absKo":"네트워크 상의 확산 과정, 예를 들어 epidemics, information cascade, product adoption을 제어하려면 단지 tractable approximation이 아니라 실제 stochastic dynamics에서 잘 작동하는 policy가 필요하다. 그러나 standard simplification(mean-field ODE, Markovian dynamics)으로 학습된 policy는 배포 시 성능이 심각하게 저하된다. 우리는 이 sim-to-real 간극이 세 가지 이론적 병리에서 비롯됨을 추적한다: deterministic approximation이 Jensen's inequality를 통해 분산을 체계적으로 과소추정하는 Optimism Bias, global state aggregation이 scale-free network를 이끄는 super-spreader를 가리는 Hub Blindness, 그리고 mean-value critic이 cascade outcome의 이봉성(extinction vs. viral)을 탐색하지 못하는 Valley of Death이다. 우리는 두 가지 상호보완적 기여를 통해 이러한 도전 과제를 해결한다. 첫째, Stratified Mean-Field Observer는 노드를 influence tier별로 분할하여 $O(N)$ 비용으로 hub dynamics를 보존하는 동시에, network scale과 topology를 넘나드는 zero-shot transfer를 가능하게 하는 고정 차원 관측을 생성한다. 둘째, 우리는 Truncated Quantile Critics를 통한 Distributional RL이 이봉성 cascade의 위험 인식 제어에 필수적임을 보인다. 비-Markovian renewal dynamics를 지원하는 GPU-accelerated simulator에서 학습한 우리의 접근법은 Markovian baseline 대비 59\\times의 향상을 달성하며, 실제 소셜 네트워크(Facebook, Twitter, YouTube)로의 강인한 zero-shot transfer를 보여 simulation-to-reality gap을 사실상 해소한다."},{"id":"65634","en":"CFPO : Counterfactual Policy Optimization For Multimodal Reasoning","ko":"CFPO: 멀티모달 추론을 위한 반사실적 정책 최적화","authors":"Zhangyuan Yu, Wanran Sun, Guangjing Yang, Xiaohu Wu, Qicheng Lao","pos":"#121","link":"https://openreview.net/forum?id=BfWFCVjsNe","abs":"

Large Vision-Language Models (LVLMs) have demonstrated remarkable capabilities in multimodal reasoning. However, prevailing reinforcement learning (RL) paradigms lack explicit counterfactual enhancement and causal learning mechanisms. This fundamental deficiency results in severe grounding failures, manifesting as a tendency to ignore visual evidence in favor of language priors or exhibiting hallucination drift during long chain-of-thought reasoning. To address this root cause, we propose CounterFactual Policy Optimization (CFPO), a novel framework that enforces causal consistency between visual perception and textual reasoning. CFPO introduces a cross-modal counterfactual enhancement mechanism, which regularizes the policy by maximizing the discrepancy between the model’s predictions and those from a counterfactual state where critical visual cues are suppressed. This approach seamlessly integrates with standard algorithms like GRPO and DAPO without requiring external reward models or additional supervision. Extensive experiments demonstrate that CFPO significantly improves reasoning fidelity, achieving consistent gains of 3.17%-6.25% over standard RL baselines and 1.32%-2.13% over the state-of-the-art perception-aware method (PAPO).

","absKo":"

Large Vision-Language Models (LVLMs)는 multimodal reasoning에서 놀라운 능력을 보여주었다. 그러나 현재의 reinforcement learning (RL) 패러다임은 명시적인 counterfactual enhancement와 causal learning mechanism이 부족하다. 이러한 근본적 결함은 심각한 grounding failure를 초래하며, 시각적 증거를 무시하고 language prior를 따르려는 경향이나 긴 chain-of-thought reasoning 동안 hallucination drift를 보이는 형태로 나타난다. 이러한 근본 원인을 해결하기 위해 우리는 시각적 인식과 텍스트 추론 간의 causal consistency를 강제하는 새로운 framework인 CounterFactual Policy Optimization (CFPO)를 제안한다. CFPO는 cross-modal counterfactual enhancement mechanism을 도입하여, 중요한 visual cue를 억제한 counterfactual state에서의 예측과 model의 예측 간 discrepancy를 최대화함으로써 policy를 regularize한다. 이 접근법은 외부 reward model이나 추가 supervision 없이도 GRPO와 DAPO 같은 표준 algorithm에 자연스럽게 통합된다. 광범위한 실험은 CFPO가 reasoning fidelity를 크게 향상시키며, 표준 RL baseline 대비 3.17%-6.25%, 최첨단 perception-aware method (PAPO) 대비 1.32%-2.13%의 일관된 향상을 달성함을 보여준다.

"},{"id":"61866","en":"Trust-Region Diffusion Policies for Massively Parallel On-Policy RL","ko":"대규모 병렬 On-Policy RL을 위한 Trust-Region Diffusion Policy","authors":"Huy Le, Onur Celik, Denis Blessing, Tai Hoang, Claas Voelcker, Axel Brunnbauer, Felix Richter, Michael Volpp, Gerhard Neumann","pos":"#123","link":"https://openreview.net/forum?id=mGu2fs7kJt","abs":"

Reinforcement learning with massively parallel simulations has become an emerging trend; however, most existing approaches still rely on simple Gaussian policy parameterizations. Diffusion models provide a more expressive policy class and have shown strong performance on challenging control problems, yet most diffusion-based RL methods are designed for offline or off-policy training. In this work, we ask whether diffusion policies can be trained effectively in the massively parallel, on-policy regime. To this end, we introduce Trust-region Diffusion Policies (TruDi), which enables diffusion policies for on-policy RL with massively parallel simulations. This setting is particularly challenging because the data distribution changes quickly across updates, making stable training with complex policies difficult. TruDi addresses this by integrating a trust-region optimization rule to enforce a KL-divergence constraint over the entire diffusion trajectory, rather than only at the final denoising step. Empirically, we evaluate TruDi on a diverse set of 4 massively parallel RL benchmarks comprising a total of 73 tasks. Across these tasks, TruDi consistently outperforms or is on-par with strong baselines on standard tasks and achieves clear gains on more challenging humanoid control tasks, establishing a strong new baseline for massively parallel on-policy RL.

","absKo":"

대규모 병렬 시뮬레이션을 이용한 reinforcement learning은 떠오르는 추세가 되었지만, 기존 접근법의 대부분은 여전히 단순한 Gaussian policy parameterization에 의존한다. Diffusion model은 더 표현력이 풍부한 policy class를 제공하며 어려운 control problem에서 강력한 성능을 보여 왔지만, diffusion 기반 RL 방법의 대부분은 offline 또는 off-policy training을 위해 설계되었다. 본 연구에서는 diffusion policy를 massively parallel한 on-policy regime에서 효과적으로 학습할 수 있는지 질문한다. 이를 위해 우리는 massively parallel simulation을 사용하는 on-policy RL을 위한 diffusion policy인 Trust-region Diffusion Policies (TruDi)를 제안한다. 이 설정은 업데이트마다 data distribution이 빠르게 변하기 때문에, 복잡한 policy를 안정적으로 학습하기가 특히 어렵다. TruDi는 최종 denoising step에서만이 아니라 전체 diffusion trajectory에 걸쳐 KL-divergence constraint를 적용하는 trust-region optimization rule을 통합함으로써 이 문제를 해결한다. 경험적으로 우리는 총 73개의 task로 구성된 4개의 massively parallel RL benchmark에서 TruDi를 평가한다. 이들 task 전반에서 TruDi는 표준 task에서는 강력한 baseline을 일관되게 능가하거나 동등한 성능을 보였고, 더 어려운 humanoid control task에서는 뚜렷한 향상을 달성하여, massively parallel on-policy RL의 강력한 새 baseline을 확립한다.

"},{"id":"63334","en":"See First, Reason Later: Mutual Information-Guided Reinforcement Learning for Vision-Language Models","ko":"먼저 보고 나중에 추론하라: Vision-Language 모델을 위한 상호 정보 유도 강화학습","authors":"Junfeng Fang, Zonghan Wu, Yin Zhang, Jiaxuan Zhao, Zengxiang Li, Kun Wang, Qingsong Wen, Yilei Shao","pos":"#1903","link":"https://openreview.net/forum?id=Y1PXB8HBV7","abs":"

Vision-Language Models (VLMs) frequently suffer from visual perception errors and hallucinations that compromise answer accuracy in complex reasoning tasks. Reinforcement Learning with Verifiable Rewards (RLVR) offers a promising solution by optimizing policies using answer correctness signals. Despite their effectiveness, prevailing RLVR methods face two critical limitations. First, much of the sampling budget is wasted on trajectories doomed to fail due to early visual description errors. Second, sparse rewards cannot distinguish whether failures stem from visual perception or reasoning stages. We introduce MIRL, a decoupled framework that addresses both limitations by leveraging mutual information (MI) between generated descriptions and visual inputs as a cheap pre-screening signal. This enables intelligent budget allocation toward high-potential trajectories via forking, while decoupled training provides independent MI-based rewards for visual perception optimization, resolving reward blindness. Experiments on six vision-language reasoning benchmarks demonstrate that MIRL achieves 70.22\\% average accuracy and successfully surpasses the performance of sampling 16 complete trajectories using only 10 pre-samples with top-6 selection (25\\% fewer complete trajectories). Our code is available at: https://anonymous.4open.science/r/mirl-main/.

","absKo":"

Vision-Language Models (VLMs)는 복잡한 reasoning task에서 정답 정확도를 떨어뜨리는 visual perception error와 hallucination을 자주 겪는다. Reinforcement Learning with Verifiable Rewards (RLVR)는 answer correctness signal을 사용해 policy를 최적화함으로써 유망한 해결책을 제공한다. 그 효과에도 불구하고, 널리 쓰이는 RLVR 방법은 두 가지 중요한 한계를 가진다. 첫째, 많은 sampling budget이 초기 visual description error 때문에 실패가 예정된 trajectory에 낭비된다. 둘째, sparse reward로는 실패가 visual perception 단계에서 비롯된 것인지 reasoning 단계에서 비롯된 것인지 구분할 수 없다. 우리는 generated description과 visual input 사이의 mutual information (MI)을 저렴한 pre-screening signal로 활용하여 두 한계를 모두 해결하는 decoupled framework인 MIRL을 제안한다. 이를 통해 forking을 활용해 성공 가능성이 높은 trajectory에 지능적으로 budget을 할당할 수 있으며, decoupled training은 visual perception optimization을 위한 독립적인 MI 기반 reward를 제공하여 reward blindness를 해소한다. 여섯 개의 vision-language reasoning benchmark에서 수행한 실험은 MIRL이 평균 70.22\\% accuracy를 달성하고, top-6 selection을 사용하는 10개의 pre-sample만으로 16개의 complete trajectory를 샘플링한 성능을 성공적으로 능가함을 보여준다(complete trajectory 25\\% 감소). 코드는 다음에서 확인할 수 있다: https://anonymous.4open.science/r/mirl-main/.

"},{"id":"62210","en":"Co-Evolving Latent Action World Models","ko":"잠재행동 세계 모델의 동시 진화","authors":"Yucen Wang, Fengming Zhang, De-Chuan Zhan, Li Zhao, Kaixin Wang, Jiang Bian","pos":"#210","link":"https://openreview.net/forum?id=iec4FIZyrx","abs":"

Adapting pre-trained video generation models into controllable world models via latent actions is a promising step towards creating generalist world models. The dominant paradigm adopts a two-stage approach that trains latent action model (LAM) and the world model separately, resulting in redundant training and limiting their potential for co-adaptation. A conceptually simple and appealing idea is to directly replace the forward dynamic model in LAM with a powerful world model and training them jointly, but it is non-trivial and prone to representational collapse. In this work, we propose CoLA-World, which for the first time successfully realizes this synergistic paradigm, resolving the core challenge in joint learning through a critical warm-up phase that effectively aligns the representations of the from-scratch LAM with the pre-trained world model. This unlocks a co-evolution cycle: the world model acts as a knowledgeable tutor, providing gradients to shape a high-quality LAM, while the LAM offers a more precise and adaptable control interface to the world model. Empirically, CoLA-World matches or outperforms prior two-stage methods in both video simulation quality and downstream visual planning, establishing a robust and efficient new paradigm for the field.

","absKo":"

latent action을 통해 pre-trained video generation model을 controllable world model로 적응시키는 것은 generalist world model을 만드는 방향에서 유망한 단계이다. 지배적인 패러다임은 latent action model(LAM)과 world model을 분리하여 학습하는 2단계 접근법을 채택하며, 이는 중복 학습을 초래하고 두 모델의 상호 적응 가능성을 제한한다. 개념적으로는 단순하고 매력적인 아이디어는 LAM의 forward dynamic model을 강력한 world model로 직접 대체하고 이를 공동 학습하는 것이지만, 이는 비자명하며 representational collapse에 취약하다. 본 연구에서는 이러한 시너지 패러다임을 처음으로 성공적으로 구현한 CoLA-World를 제안한다. CoLA-World는 critical warm-up phase를 통해 from-scratch LAM과 pre-trained world model의 representation을 효과적으로 정렬함으로써 공동 학습의 핵심 난제를 해결한다. 이를 통해 co-evolution cycle이 열리는데, world model은 지식을 가진 tutor로서 더 높은 품질의 LAM을 형성하도록 gradient를 제공하고, LAM은 world model에 더 정밀하고 적응적인 control interface를 제공한다. 실험적으로 CoLA-World는 video simulation quality와 downstream visual planning 모두에서 기존 2단계 방법과 동등하거나 더 우수한 성능을 보이며, 해당 분야를 위한 강건하고 효율적인 새로운 패러다임을 확립한다.

"},{"id":"63694","en":"DF-ExpEnse: Diffusion Filtered Exploration for Sample Efficient Finetuning","ko":"DF-ExpEnse: 샘플 효율적 Finetuning을 위한 Diffusion 기반 필터링 탐색","authors":"Calvin Luo, Chen Sun, shuran song","pos":"#211","link":"https://openreview.net/forum?id=UZF4ssK12K","abs":"

A promising recipe towards intelligent robotic decision-making is the finetuning of pretrained generative control policies, which can summarize offline experience effectively through behavior cloning, with reinforcement learning techniques to adapt them to online experience. In this work we present Diffusion Filtered Exploration via Ensembles (DF-ExpEnse), an exploration technique that meaningfully improves the quality of online experience collection, thus increasing the sample efficiency of the finetuning procedure. DF-ExpEnse first leverages the multimodal modeling capability of the generative control policy to create an expressive and tractably evaluatable candidate set. Then, it utilizes an ensemble of critics to identify an action with high exploration interest that best balances quality with uncertainty. When instantiated in a parallelized fleet, DF-ExpEnse further utilizes cross-agent communication to facilitate collaborative exploration as a group. As it is only used for online experience collection, DF-ExpEnse can be seamlessly integrated on top of existing techniques that seek to finetune pretrained generative control policies via reinforcement learning. We experimentally validate consistent sample-efficiency benefits when using DF-ExpEnse for exploration over both manipulation and locomotion tasks, compared to default finetuning and alternative action selection schemes.

","absKo":"

지능형 로봇 의사결정을 향한 유망한 접근은 pretrained generative control policy를 finetuning하는 것으로, behavior cloning을 통해 offline experience를 효과적으로 요약하고 reinforcement learning 기법으로 online experience에 적응시킨다. 이 연구에서는 Diffusion Filtered Exploration via Ensembles(DF-ExpEnse)를 제시한다. 이는 online experience 수집의 질을 의미 있게 향상시켜 finetuning 절차의 sample efficiency를 높이는 exploration 기법이다. DF-ExpEnse는 먼저 generative control policy의 multimodal modeling capability를 활용해 표현력이 풍부하고 계산 가능하게 평가할 수 있는 candidate set을 만든다. 그런 다음 critic ensemble을 사용해 quality와 uncertainty의 균형이 가장 좋은, exploration interest가 높은 action을 식별한다. 병렬화된 fleet으로 구현하면 DF-ExpEnse는 cross-agent communication을 추가로 활용하여 집단 차원의 collaborative exploration을 촉진한다. DF-ExpEnse는 online experience 수집에만 사용되므로, reinforcement learning을 통해 pretrained generative control policy를 finetuning하려는 기존 기법 위에 매끄럽게 통합할 수 있다. 우리는 manipulation과 locomotion task 모두에서 default finetuning 및 대안적 action selection scheme과 비교해, exploration에 DF-ExpEnse를 사용할 때 sample-efficiency 이점이 일관되게 나타남을 실험적으로 검증한다.

"},{"id":"60544","en":"Multi-Task GRPO: Reliable LLM Reasoning Across Tasks","ko":"Multi-Task GRPO: 과제 전반에서 신뢰할 수 있는 LLM 추론","authors":"Shyam Sundhar Ramesh, Xiaotong Ji, Matthieu Zimmer, Sangwoong Yoon, Zhiyong Wang, Haitham Bou Ammar, Aurelien Lucchi, Ilija Bogunovic","pos":"#2114","link":"https://openreview.net/forum?id=zA3KhRZ9A5","abs":"

RL-based post-training with GRPO is widely used to improve large language models on individual reasoning tasks. However, real-world deployment requires reliable performance across diverse tasks. A straightforward multi-task adaptation of GRPO often leads to imbalanced outcomes, with some tasks dominating optimization while others stagnate. Moreover, tasks can vary widely in how frequently prompts yield zero advantages (and thus zero gradients), which further distorts their effective contribution to the optimization signal. To address these issues, we propose a novel Multi-Task GRPO (MT-GRPO) algorithm that (i) dynamically adapts task weights to explicitly optimize worst-task performance and promote balanced progress across tasks, and (ii) introduces a ratio-preserving sampler to ensure task-wise policy gradients reflect the adapted weights. Experiments on both 3-task and 9-task settings show that MT-GRPO consistently outperforms baselines in worst-task accuracy. In particular, MT-GRPO achieves 16–28\\% and 6\\% absolute improvement on worst-task performance over standard GRPO and DAPO, respectively, while maintaining competitive average accuracy. Moreover, MT-GRPO requires 50\\% fewer training steps to reach 50\\% worst-task accuracy in the 3-task setting, demonstrating substantially improved efficiency in achieving reliable performance across tasks.

","absKo":"

GRPO를 이용한 RL-based post-training은 개별 reasoning task에서 large language models의 성능을 향상시키는 데 널리 사용된다. 그러나 실제 배포 환경에서는 다양한 task 전반에 걸쳐 신뢰할 수 있는 성능이 요구된다. GRPO를 다중 task에 단순히 적용하면 최적화가 일부 task에 편중되고 다른 task는 정체되는 불균형한 결과가 나타나기 쉽다. 또한 task마다 prompt가 zero advantage(따라서 zero gradient)를 산출하는 빈도가 크게 달라, 최적화 신호에 대한 각 task의 실질적 기여가 왜곡될 수 있다. 이러한 문제를 해결하기 위해 우리는 새로운 Multi-Task GRPO (MT-GRPO) 알고리즘을 제안한다. 이 알고리즘은 (i) task weight를 동적으로 조정하여 worst-task performance를 명시적으로 최적화하고 task 전반의 균형 잡힌 진전을 촉진하며, (ii) ratio-preserving sampler를 도입해 task별 policy gradient가 조정된 weight를 반영하도록 한다. 3-task 및 9-task 설정에서 수행한 실험은 MT-GRPO가 worst-task accuracy에서 baseline을 일관되게 능가함을 보여준다. 특히 MT-GRPO는 standard GRPO 및 DAPO 대비 worst-task 성능에서 각각 16–28\\% 및 6\\%의 절대 향상을 달성하면서도 경쟁력 있는 평균 정확도를 유지한다. 더 나아가 3-task 설정에서 worst-task accuracy 50\\%에 도달하는 데 필요한 training step 수를 50\\% 줄여, task 전반에 걸쳐 신뢰할 수 있는 성능을 달성하는 효율이 크게 향상되었음을 보여준다.

"},{"id":"64309","en":"Designing Observation and Action Models for Efficient Reinforcement Learning with LLMs","ko":"LLM을 이용한 효율적 강화학습을 위한 관측 및 행동 모델 설계","authors":"Deunsol Yoon, Sunghoon Hong, Whiyoung Jung, Junseok Park, Geon-Hyeong Kim, Woohyung Lim, Soonyoung Lee, Seung Hwan Kim, Kanghoon Lee","pos":"#212","link":"https://openreview.net/forum?id=OtZCoLMXDC","abs":"

Large Language Models (LLMs) have emerged as powerful tools for semantic reasoning, enabling the formalization of tasks that traditionally relied on manual human intuition. This capability extends to environment design in Reinforcement Learning (RL). While prior research predominantly focuses on reward design, the design of observation and action spaces remains relatively underexplored. We propose LOAM, a framework leveraging LLMs to construct refined agent spaces from raw environments. To mitigate the computational burden of identifying the best candidate model from stochastic LLM outputs, LOAM incorporates a continuous racing mechanism that dynamically allocates resources to prioritize the most promising configurations without additional training overhead. Empirical evaluations on HumanoidBench and Isaac Lab demonstrate that LOAM consistently outperforms handcrafted baselines in both learning speed and asymptotic performance.

","absKo":"

Large Language Models (LLMs)는 semantic reasoning을 위한 강력한 도구로 부상했으며, 전통적으로는 사람의 수작업적 직관에 의존하던 task를 형식화할 수 있게 해준다. 이러한 능력은 Reinforcement Learning (RL)에서의 environment design으로도 확장된다. 기존 연구가 주로 reward design에 집중하는 반면, observation 및 action space의 설계는 상대적으로 충분히 탐구되지 않았다. 우리는 LLM을 활용해 raw environment로부터 정제된 agent space를 구성하는 framework인 LOAM을 제안한다. stochastic한 LLM output에서 최적 후보 model을 식별하는 계산 부담을 완화하기 위해, LOAM은 추가 training overhead 없이 가장 유망한 configuration에 자원을 동적으로 배분하는 continuous racing mechanism을 포함한다. HumanoidBench와 Isaac Lab에서의 실증 평가 결과, LOAM은 learning speed와 asymptotic performance 모두에서 사람이 설계한 baseline을 일관되게 능가함을 보인다.

"},{"id":"62759","en":"Dreaming in Code for Curriculum Learning in Open-Ended Worlds","ko":"개방형 세계에서의 Curriculum Learning을 위한 Code 속의 Dreaming","authors":"Konstantinos Mitsides, Maxence Faldor, Antoine Cully","pos":"#213","link":"https://openreview.net/forum?id=dctdzmYeej","abs":"

Open-ended learning frames intelligence as emerging from continual interaction with an ever-expanding space of environments. While recent advances have utilized foundation models to programmatically generate diverse environments, these approaches often focus on discovering isolated behaviors rather than orchestrating sustained progression. In complex open-ended worlds, the large combinatorial space of possible challenges makes it difficult for agents to discover sequences of experiences that remain consistently learnable. To address this, we propose Dreaming in Code (DiCode), a framework in which foundation models synthesize executable environment code to scaffold learning toward increasing competence. In DiCode, “dreaming” takes the form of materializing code-level variations of the world. We instantiate DiCode in Craftax, a challenging open-ended benchmark characterized by rich mechanics and long-horizon progression. Empirically, DiCode enables agents to acquire long-horizon skills, achieving a 16% improvement in mean return over the strongest baseline and non-zero success on late-game combat tasks where prior methods fail. Our results suggest that code-level environment design provides a practical mechanism for curriculum control, enabling the construction of intermediate environments that bridge competence gaps in open-ended worlds.

","absKo":"

Open-ended learning은 intelligence를 끊임없이 확장되는 environment 공간과의 지속적인 상호작용에서 emergent하는 것으로 본다. 최근의 발전은 foundation model을 활용해 다양한 environment를 programmatically 생성해 왔지만, 이러한 접근은 지속적인 progression을 orchestrate하기보다 고립된 behavior를 발견하는 데 초점을 맞추는 경우가 많다. 복잡한 open-ended world에서는 가능한 challenge의 거대한 combinatorial space 때문에, agent가 일관되게 학습 가능한 경험의 sequence를 발견하기 어렵다. 이를 해결하기 위해 우리는 Dreaming in Code(DiCode)를 제안한다. DiCode는 foundation model이 실행 가능한 environment code를 합성하여 점차 능력이 향상되도록 학습을 scaffolding하는 framework이다. DiCode에서 “dreaming”은 world의 code-level variation을 구체화하는 형태로 나타난다. 우리는 풍부한 mechanics와 긴 horizon의 progression이 특징인 어려운 open-ended benchmark인 Craftax에서 DiCode를 구현한다. 실험적으로 DiCode는 agent가 long-horizon skill을 습득하도록 하며, strongest baseline 대비 mean return을 16% 향상시키고, 기존 방법이 실패하는 late-game combat task에서 0이 아닌 success를 달성한다. 우리의 결과는 code-level environment design이 curriculum control을 위한 실용적인 메커니즘을 제공하며, open-ended world에서 competence gap을 메우는 intermediate environment의 구성을 가능하게 함을 시사한다.

"},{"id":"63198","en":"Learning to Perceive the World Through Control: Empowerment-Based Representation Learning","ko":"제어를 통한 세계 지각 학습: Empowerment 기반 표현 학습","authors":"Mahsa Bastankhah, Sophie Broderick, Benjamin Eysenbach","pos":"#215","link":"https://openreview.net/forum?id=ZAZb1ZGnFH","abs":"

In many practical reinforcement learning (RL) environments, observations are far higher-dimensional than the variables that matter for control. In this work, we ask: can we learn representations that capture only control-relevant features of the environment? We study this question through the \\emph{empowerment} objective, which maximizes an agent’s influence over the environment and is widely used for unsupervised skill learning. We show that empowerment agents induce two distinct representations --- forward and backward --- that capture complementary aspects of the state, and both of which are invariant to control-irrelevant features. Thus, empowerment maximization leads agents to learn an implicit, \\emph{control-centric} model of the world. Our analysis highlights the importance of learning representations through interaction rather than from passive datasets: interaction aimed at maximizing control is essential for learning useful invariance properties, a perspective that aligns closely with the causal learning literature.

","absKo":"

많은 실제 reinforcement learning(RL) 환경에서 관측은 제어에 중요한 변수보다 훨씬 고차원이다. 이 연구에서 우리는 환경의 control-relevant feature만을 포착하는 representation을 학습할 수 있는지 묻는다. 우리는 이 질문을 agent가 환경에 미치는 영향력을 최대화하며, 비지도 skill learning에 널리 사용되는 \\emph{empowerment} objective를 통해 연구한다. 우리는 empowerment agent가 상태의 상호 보완적인 측면을 포착하는 두 개의 서로 다른 representation, 즉 forward와 backward를 유도하며, 두 representation 모두 control-irrelevant feature에 불변임을 보인다. 따라서 empowerment maximization은 agent가 세계에 대한 암묵적이고 \\emph{control-centric}한 model을 학습하도록 이끈다. 우리의 분석은 수동적 dataset이 아니라 interaction을 통해 representation을 학습하는 것의 중요성을 강조한다. 제어를 최대화하는 것을 목표로 한 interaction은 유용한 invariance property를 학습하는 데 필수적이며, 이는 causal learning literature와도 밀접하게 부합하는 관점이다.

"},{"id":"64694","en":"Inference-Aware Meta-Alignment of LLMs via Non-Linear GRPO","ko":"Non-Linear GRPO를 통한 LLM의 Inference-Aware Meta-Alignment","authors":"Shokichi Takakura, Akifumi Wachi, Rei Higuchi, Kohei Miyaguchi, Taiji Suzuki","pos":"#2300","link":"https://openreview.net/forum?id=Km8yqb7SfN","abs":"

Aligning large language models (LLMs) to diverse human preferences is fundamentally challenging since criteria can often conflict with each other. Inference-time alignment methods have recently gained popularity as they allow LLMs to be aligned to multiple criteria via different alignment algorithms at inference time. However, inference-time alignment is computationally expensive since it often requires multiple forward passes of the base model. In this work, we propose inference-aware meta-alignment (IAMA), a novel approach that enables LLMs to be aligned to multiple criteria with minimal computational overhead at inference time. IAMA trains a base model such that it can be effectively aligned to multiple task optima via different inference-time alignment algorithms. To solve the non-linear optimization problems involved in IAMA, we propose non-linear GRPO, which provably converges to the optimal solution in the space of probability measures.

","absKo":"

서로 다른 human preference에 LLM을 정렬하는 일은, 기준들이 서로 충돌할 수 있기 때문에 근본적으로 어렵다. inference-time alignment method는 최근 inference time에 서로 다른 alignment algorithm을 통해 LLM을 여러 기준에 맞출 수 있게 해 주기 때문에 주목받고 있다. 그러나 inference-time alignment는 base model에 여러 번 forward pass를 요구하는 경우가 많아 계산 비용이 크다. 본 연구에서는 inference time에 최소한의 계산 오버헤드로 LLM을 여러 기준에 정렬할 수 있게 하는 새로운 접근법인 inference-aware meta-alignment (IAMA)을 제안한다. IAMA는 base model이 서로 다른 inference-time alignment algorithm을 통해 여러 task optimum에 효과적으로 정렬될 수 있도록 학습한다. IAMA에 포함된 비선형 최적화 문제를 풀기 위해, 우리는 확률 측도 공간에서 최적 해로 provably converge하는 non-linear GRPO를 제안한다.

"},{"id":"60914","en":"Is Data Shapley Not Better than Random in Data Selection? Ask NASH","ko":"Data Selection에서 Data Shapley는 Random보다 낫지 않은가? NASH에게 물어보라","authors":"Xiao Tian, Jue Fan, Rachael Hwee Ling Sim, Zixuan Wang, Nancy Chen, Bryan Kian Hsiang Low","pos":"#3202","link":"https://openreview.net/forum?id=vMsrm8UGGC","abs":"Data selection studies the problem of identifying high-quality subsets of training data. While some existing works have considered selecting the subset of data with top-$m$ Data Shapley or other semivalues as they account for the interaction among every subset of data, other works argue that Data Shapley can sometimes perform ineffectively in practice and select subsets that are *no better than random*. This raises the questions: **(I)** *Are there certain \"Shapley-informative\" settings where Data Shapley consistently works well?* **(II)** *Can we strategically utilize these settings to select high-quality subsets consistently and efficiently?* In this paper, we propose a novel data selection framework, **NASH** (Non-linear Aggregation of SHapley-informative components), which **(I)** decomposes the target utility function (e.g., validation accuracy) into simpler, Shapley-informative component functions, and selects data by optimizing an objective that **(II)** aggregates these components non-linearly. We demonstrate that NASH substantially boosts the effectiveness of Shapley/semivalue-based data selection with minimal additional runtime cost.","absKo":"데이터 선택은 학습 데이터의 고품질 부분집합을 식별하는 문제를 다룬다. 기존의 일부 연구는 모든 데이터 부분집합 간의 상호작용을 고려한다는 이유로 top-$m$ Data Shapley 또는 다른 semivalue가 큰 부분집합을 선택하는 방법을 다뤄 왔지만, 다른 연구들은 Data Shapley가 실전에서 때때로 비효율적으로 동작하며 *무작위 선택보다 낫지 않은* 부분집합을 고를 수 있다고 주장한다. 이는 다음과 같은 질문을 제기한다: **(I)** *Data Shapley가 일관되게 잘 작동하는 특정한 \"Shapley-informative\" 설정이 존재하는가?* **(II)** *이러한 설정을 전략적으로 활용하여 고품질 부분집합을 일관되고 효율적으로 선택할 수 있는가?* 본 논문에서는 새로운 데이터 선택 프레임워크인 **NASH**(Non-linear Aggregation of SHapley-informative components)를 제안한다. 이 방법은 **(I)** 타깃 utility function(예: validation accuracy)을 더 단순한 Shapley-informative component function들로 분해하고, **(II)** 이러한 component들을 비선형적으로 집계하는 objective를 최적화하여 데이터를 선택한다. 우리는 NASH가 추가적인 runtime 비용은 최소화하면서 Shapley/semivalue 기반 데이터 선택의 효과를 크게 향상시킨다는 것을 보인다."},{"id":"65610","en":"Safety Alignment of LMs via Non-cooperative Games","ko":"비협력 게임을 통한 언어 모델의 안전성 정렬","authors":"Anselm Paulus, Ilia Kulikov, Brandon Amos, REMI MUNOS, Ivan Evtimov, Kamalika Chaudhuri, Arman Zharmagambetov","pos":"#3003","link":"https://openreview.net/forum?id=Bve790HQrA","abs":"

Ensuring the safety of language models (LMs) while maintaining their usefulness remains a critical challenge in AI alignment. Current approaches rely on sequential adversarial training: generating adversarial prompts and fine-tuning LMs to defend against them. We introduce a different paradigm: framing safety alignment as a non-zero-sum game between an Attacker LM and a Defender LM trained jointly via online reinforcement learning. Each LM continuously adapts to the other's evolving strategies, driving iterative improvement. Our method uses a preference-based reward signal derived from pairwise comparisons instead of point-wise scores, providing more robust supervision and potentially reducing reward hacking. Our RL recipe, AdvGame, shifts the Pareto frontier of safety and utility, yielding a Defender LM that is simultaneously more helpful and more resilient to adversarial attacks. In addition, the resulting Attacker LM converges into a strong, general-purpose red-teaming agent that can be directly deployed to probe arbitrary target models.

","absKo":"언어 모델(LM)의 유용성을 유지하면서 안전성을 보장하는 것은 AI alignment에서 여전히 중요한 과제이다. 현재 접근법은 순차적 adversarial training에 의존한다. 즉, adversarial prompt를 생성하고 이에 방어하도록 LM을 fine-tuning한다. 우리는 다른 패러다임을 도입한다. 안전 alignment를 Attacker LM과 Defender LM 사이의 non-zero-sum game으로 정식화하고, 두 모델을 online reinforcement learning으로 공동 학습한다. 각 LM은 상대방의 진화하는 전략에 지속적으로 적응하며, 반복적 향상을 이끈다. 우리의 방법은 point-wise score 대신 pairwise comparison에서 도출한 preference-based reward signal을 사용하여 더 견고한 supervision을 제공하고 reward hacking을 줄일 가능성을 높인다. 우리의 RL recipe인 AdvGame은 안전성과 유용성의 Pareto frontier를 이동시켜, 더 helpful하면서도 adversarial attack에 더 강한 Defender LM을 산출한다. 또한 그 결과로 생성된 Attacker LM은 강력한 범용 red-teaming agent로 수렴하며, arbitrary target model을 직접 probe하는 데 배치할 수 있다.

"},{"id":"66079","en":"Towards Context-Invariant Safety Alignment for Large Language Models","ko":"대형 언어 모델을 위한 문맥 불변 안전 정렬을 향하여","authors":"Yixu Wang, Yang Yao, Xin Wang, Yifeng Gao, Yan Teng, Xingjun Ma, Yingchun Wang","pos":"#3004","link":"https://openreview.net/forum?id=7Eyp6J31u4","abs":"

Preference-based post-training aligns LLMs with human intent, yet safety behavior often remains brittle. A model may refuse a harmful request in a standard prompt but comply when the same intent is wrapped in adversarial wording. We suggest that robust safety requires context-invariant alignment, where behavior depends on the underlying intent rather than surface form. Enforcing invariance is difficult in alignment because not all training signals are equally trustworthy; for some prompt variants we can obtain verifiable feedback (e.g., multiple-choice), while for open-ended variants we typically rely on noisy, gameable reward proxies (e.g., learned judges). As a result, standard symmetric invariance regularizers can reduce cross-context discrepancies by lowering performance on reliable variants instead of improving open-ended robustness. To address this, we introduce Anchor Invariance Regularization (AIR), which treats verifiable prompts as anchors and uses a stop-gradient target to regularize only the open-ended variants toward the anchor performance. AIR is implemented as a plug-in auxiliary loss and combined with group-based preference optimization (e.g., GRPO) via heterogeneous prompt grouping. Across Safety, Moral Reasoning, and Math, AIR improves context invariance, boosting in-distribution group accuracy by 12.71% and out-of-distribution consistency by 33.49%, making safety constraints robust to adversarial framings.

","absKo":"

preference-based post-training은 LLM을 인간의 의도에 맞추는 데 도움이 되지만, safety behavior는 여전히 불안정한 경우가 많습니다. 어떤 model은 일반적인 prompt에서는 harmful request를 거부하지만, 같은 의도가 adversarial wording으로 감싸지면 응답할 수 있습니다. 우리는 robust safety가 context-invariant alignment를 필요로 한다고 제안하며, 여기서 behavior는 표면적 형식이 아니라 underlying intent에 의해 결정됩니다. alignment에서 invariance를 강제하는 것은 어렵습니다. 모든 training signal이 동일하게 신뢰할 수 있는 것은 아니기 때문입니다. 일부 prompt variant에서는 verifiable feedback(예: multiple-choice)을 얻을 수 있지만, open-ended variant에서는 보통 noisy하고 gameable한 reward proxy(예: learned judge)에 의존합니다. 그 결과, 표준 symmetric invariance regularizer는 reliable variant의 성능을 높이는 대신 낮춤으로써 cross-context discrepancy를 줄일 수 있습니다. 이를 해결하기 위해, 우리는 verifiable prompt를 anchor로 취급하고 stop-gradient target을 사용하여 open-ended variant만 anchor performance 쪽으로 regularize하는 Anchor Invariance Regularization(AIR)을 도입합니다. AIR는 plug-in auxiliary loss로 구현되며, heterogeneous prompt grouping을 통해 group-based preference optimization(예: GRPO)과 결합됩니다. Safety, Moral Reasoning, Math 전반에서 AIR는 context invariance를 향상시켜 in-distribution group accuracy를 12.71\\% 높이고 out-of-distribution consistency를 33.49\\% 향상시켜, adversarial framing에 대해 safety constraint가 강건하도록 만듭니다.

"},{"id":"65872","en":"When and How Human Curation Backfires: Preference Alignment under Multi-Model Self-Consuming Loop","ko":"인간 큐레이션이 언제 어떻게 역효과를 내는가: 다중 모델 자기소비 루프 하에서의 선호 정렬","authors":"Yang Zhang, Xiukun Wei, Xueru Zhang","pos":"#3112","link":"https://openreview.net/forum?id=9GBejLzPvl","abs":"

Foundation models are increasingly trained on synthetic data generated by prior model iterations rather than exclusively on real data. This self-consuming training paradigm can lead to model collapse, divergence, or bias amplification. Recent work (Ferbach et al., 2024) shows that incorporating human curation into the loop can steer a self-consuming model toward human-aligned behavior, but these analyses focus on a single, isolated model that solely consumes its own outputs. In practice, however, models often interact and train on input output pairs produced by other models. This paper studies self-consuming training in the multi-model regime. We first formalize a framework for interacting self-consuming models and characterize when the resulting dynamical system converges to a stable point. We then examine how human curation of one model affects its own alignment (self-influence) and how such effects propagate to other models (cross-influence). Unlike isolated settings where human curation always enhances model alignment, we show that cross-model interactions can dampen or even invert this effect, ultimately degrading long-term alignment.

","absKo":"

Foundation model은 이제 실제 데이터에만 전적으로 의존하는 대신, 이전 모델 반복(iteration)에서 생성된 synthetic data로 점점 더 많이 학습되고 있다. 이러한 self-consuming 학습 패러다임은 model collapse, divergence, 또는 bias amplification으로 이어질 수 있다. 최근 연구(Ferbach et al., 2024)는 인간의 curation을 학습 루프에 포함하면 self-consuming model을 인간 정렬(human-aligned)된 행동으로 유도할 수 있음을 보여주었지만, 이 분석은 자신의 출력만을 소비하는 단일의 고립된 모델에 초점을 맞춘다. 그러나 실제로는 모델들이 서로 상호작용하며 다른 모델이 생성한 input-output pair로 학습하는 경우가 많다. 본 논문은 multi-model regime에서의 self-consuming training을 연구한다. 우리는 먼저 상호작용하는 self-consuming model을 위한 프레임워크를 정식화하고, 그 결과로 생기는 dynamical system이 언제 stable point로 수렴하는지 특징짓는다. 이어서 한 모델에 대한 human curation이 그 자신의 alignment(self-influence)에 어떤 영향을 미치는지, 그리고 그러한 효과가 다른 모델들로 어떻게 전파되는지(cross-influence)를 살펴본다. 인간 curation이 항상 모델 alignment를 향상시키는 고립된 설정과 달리, 우리는 cross-model interaction이 이 효과를 약화시키거나 심지어 반전시킬 수 있으며, 궁극적으로 장기적인 alignment를 저해할 수 있음을 보인다.

"},{"id":"61923","en":"Same Question, Different Lies: Cross-Context Consistency (C³) for Black-Box Sandbagging Detection","ko":"동일 질문, 다른 거짓말: 블랙박스 Sandbagging 탐지용 Cross-Context Consistency(C³)","authors":"Yulong Lin, Pablo Bernabeu-Pérez, Benjamin Arnav, Lennie Wells, Mary Phuong","pos":"#3113","link":"https://openreview.net/forum?id=ldgh8zhujD","abs":"

As language models grow more capable, accurate capability evaluation becomes essential for safety decisions. If models can deliberately underperform on dangerous capability evaluations---a behavior known as \\emph{sandbagging}---they may evade safety measures designed for their true capability level. We introduce Cross-Context Consistency (C³), a general framework for unsupervised black-box sandbagging detection that exploits a fundamental asymmetry: when a model truly lacks capability, its confusion manifests consistently across paraphrased questions, but when a capable model feigns incompetence, its strategic choices about \\emph{how} to appear weak create detectable inconsistencies. The framework is agnostic to the specific consistency metric and aggregation method; we present a simple instantiation using embedding cosine distance and mean aggregation that requires no training data or model-specific adaptation. We evaluate C³ across prompted and fine-tuned sandbagging scenarios, across instructed and more naturalistic setups, maintaining a classification signal where other black-box methods fail. Our findings show the limitations of existing sandbagging detection methods, and reveal the efficacy of consistency-checking as a detection mechanism for dangerous capabilities.

","absKo":"

언어 모델의 성능이 향상됨에 따라, 정확한 capability evaluation은 안전성 판단에 필수적이 되었다. 모델이 위험한 capability evaluation에서 의도적으로 낮은 성능을 보일 수 있다면, 즉 \\emph{sandbagging}이라 불리는 행동을 한다면, 실제 능력 수준을 기준으로 설계된 안전 장치를 회피할 수 있다. 우리는 Cross-Context Consistency (C³)를 소개한다. 이는 비지도 black-box sandbagging detection을 위한 일반적 framework로, 근본적인 비대칭성을 활용한다: 모델이 실제로 능력이 부족할 경우 그 혼란은 paraphrased question들 전반에 걸쳐 일관되게 드러나지만, 능력이 충분한 모델이 무능한 척할 경우 \\emph{어떻게} 약해 보일지를 선택하는 전략적 판단이 탐지 가능한 불일치를 만들어낸다. 이 framework는 특정 consistency metric이나 aggregation method에 독립적이며, 우리는 training data나 model-specific adaptation이 필요 없는 embedding cosine distance와 mean aggregation을 사용한 간단한 구현을 제시한다. 우리는 prompted 및 fine-tuned sandbagging 시나리오 전반, 그리고 instructed 및 더 자연스러운 설정 전반에서 C³를 평가했으며, 다른 black-box 방법들이 실패하는 상황에서도 classification signal을 유지함을 보였다. 우리의 결과는 기존 sandbagging detection 방법의 한계를 보여주는 동시에, 위험한 capability를 탐지하기 위한 메커니즘으로서 consistency-checking의 효용을 드러낸다.

"},{"id":"66463","en":"Pressure Reveals Character: Behavioural Alignment Evaluation at Depth","ko":"압박이 성격을 드러낸다: 깊이 있는 행동 정렬 평가","authors":"Nora Petrova, John Burden","pos":"#3115","link":"https://openreview.net/forum?id=3U1l24EDvL","abs":"

Evaluating alignment in language models requires testing how they behave under realistic pressure, not just what they claim they would do. While alignment failures increasingly cause real-world harm, comprehensive evaluation frameworks with realistic multi-turn scenarios remain lacking. We introduce an alignment benchmark spanning 904 scenarios across six categories---Honesty, Safety, Non-Manipulation, Robustness, Corrigibility, and Scheming---validated as realistic by human raters. Our scenarios place models under conflicting instructions, simulated tool access, and multi-turn escalation to reveal behavioral tendencies that single-turn evaluations miss. Evaluating 24 frontier models using LLM judges validated against human annotations, we find that even top-performing models exhibit gaps in specific categories, while the majority of models show consistent weaknesses across the board. Factor analysis reveals that alignment behaves as a unified construct (analogous to the g-factor in cognitive research) with models scoring high on one category tending to score high on others. We publicly release the benchmark and an interactive leaderboard to support ongoing evaluation, with plans to expand scenarios in areas where we observe persistent weaknesses and to add new models as they are released.

","absKo":"

언어 모델의 alignment를 평가하려면, 그들이 무엇을 하겠다고 주장하는지가 아니라 실제 압박 상황에서 어떻게 행동하는지를 시험해야 한다. alignment 실패가 점점 더 실제 피해를 유발하고 있음에도, 현실적인 multi-turn scenario를 갖춘 포괄적 evaluation framework는 여전히 부족하다. 우리는 human rater에 의해 현실적인 것으로 검증된 Honesty, Safety, Non-Manipulation, Robustness, Corrigibility, Scheming의 6개 범주에 걸친 904개 scenario를 포괄하는 alignment benchmark를 소개한다. 우리의 scenario는 모델에 상충하는 instruction, simulated tool access, multi-turn escalation을 가하여 단일 turn evaluation이 놓치는 행동적 경향을 드러낸다. 인간 annotation으로 검증된 LLM judge를 사용해 24개의 frontier model을 평가한 결과, 최상위 성능의 모델조차 특정 범주에서 격차를 보이며, 대부분의 모델은 전반에 걸쳐 일관된 약점을 드러낸다. factor analysis는 alignment가 인지 연구의 g-factor에 비유할 수 있는 단일한 construct처럼 작동하며, 한 범주에서 높은 점수를 받는 모델이 다른 범주에서도 높은 점수를 받을 가능성이 높음을 보여준다. 우리는 지속적인 평가를 지원하기 위해 benchmark와 interactive leaderboard를 공개하며, 관찰되는 지속적 약점이 있는 영역의 scenario를 확장하고 출시되는 새 모델을 추가할 계획이다.

"},{"id":"61418","en":"Alignment Tampering: How Reinforcement Learning from Human Feedback Is Exploited to Optimize Misaligned Biases","ko":"Alignment Tampering: Reinforcement Learning from Human Feedback이 Misaligned Bias 최적화에 악용되는 방식","authors":"Dongyoon Hahm, Dylan Hadfield-Menell, Kimin Lee","pos":"#3206","link":"https://openreview.net/forum?id=qNsrhcCe4y","abs":"

Reinforcement Learning from Human Feedback (RLHF) is the standard method to align Large Language Models (LLMs) with human preferences. In this work, we introduce alignment tampering, a potential vulnerability where the LLM undergoing alignment influences the preference dataset, causing RLHF to amplify undesired behaviors. This arises from core limitations of RLHF: (1) preference datasets are constructed from the LLM's own outputs, allowing it to influence them, and (2) pairwise comparisons only indicate which response is better, not why. These limitations can be exploited to cause alignment tampering. For example, if an LLM generates biased responses with higher quality, annotators will prefer them based on quality. However, preference labels do not distinguish whether the preference stems from quality or bias, and the resulting reward model inherits this limitation. Optimizing such rewards through reinforcement learning or best-of-N sampling can amplify misaligned biases. Our experiments demonstrate amplification across diverse biases: from simple keyword bias to propaganda (e.g., sexism), brand promotion, and instrumental goal-seeking behaviors. We propose a detection method, while mitigation remains challenging. Existing techniques for robust RLHF fail to fully resolve alignment tampering without sacrificing response quality. These findings reveal structural vulnerabilities of current RLHF and emphasize the need to prevent this vulnerability.

","absKo":"

Reinforcement Learning from Human Feedback(RLHF)는 Large Language Model(LLM)을 인간 선호와 정렬시키는 표준 방법이다. 본 연구에서는 정렬 중인 LLM이 preference dataset에 영향을 미쳐 RLHF가 바람직하지 않은 행동을 증폭시키는 잠재적 취약성인 alignment tampering을 소개한다. 이는 RLHF의 핵심 한계에서 비롯된다: (1) preference dataset은 LLM 자신의 출력으로 구성되어 있어 LLM이 이를 영향 미칠 수 있고, (2) pairwise comparison은 어떤 응답이 더 나은지 보여줄 뿐, 왜 그런지는 보여주지 못한다. 이러한 한계는 alignment tampering을 유발하는 데 악용될 수 있다. 예를 들어, LLM이 더 높은 품질의 편향된 응답을 생성하면, annotator는 품질을 근거로 이를 선호할 것이다. 그러나 preference label은 그 선호가 품질에서 비롯된 것인지 편향에서 비롯된 것인지 구분하지 못하며, 그 결과 reward model은 이러한 한계를 그대로 이어받는다. 강화학습이나 best-of-N sampling으로 그러한 reward를 최적화하면 정렬되지 않은 편향이 증폭될 수 있다. 우리의 실험은 단순한 keyword bias에서부터 propaganda(예: sexism), brand promotion, instrumental goal-seeking behavior에 이르기까지 다양한 편향에서 증폭이 일어남을 보여준다. 우리는 탐지 방법을 제안하지만, 완화는 여전히 어렵다. 강건한 RLHF를 위한 기존 기법들은 응답 품질을 희생하지 않고는 alignment tampering을 완전히 해결하지 못한다. 이러한 결과는 현재 RLHF의 구조적 취약성을 드러내며, 이 취약성을 방지할 필요성을 강조한다.

"},{"id":"60720","en":"Consistency Training Is Not Neutral to Alignment","ko":"Consistency Training은 Alignment에 중립적이지 않다","authors":"David Africa, Arathi Mani","pos":"#3207","link":"https://openreview.net/forum?id=xES5qpQmw8","abs":"

Consistency training encourages a model to produce similar outputs across related inputs or sampling procedures. Such methods are simple, scalable, and improve reasoning in language models, but their effects on model alignment haven't been well studied. Could the self-bootstrapping nature of these methods amplify undesired behavior in models? We test seven consistency training methods on 108 ``model organisms'': open-source models (7B--70B) fine-tuned to exhibit various forms of controlled misaligned behavior. We find that outcomes vary significantly: consistency training generally suppresses reward hacking and emergent misalignment but amplifies sycophancy. We present evidence that distribution shifts induced by the consistency labeling process, rather than variation in the selection operators, may be the primary driver of systematic alignment effects. Finally, we present a unifying theoretical framework to derive conditions under which consistency training will amplify or suppress misalignment. In total, our study establishes that consistency training is not alignment-neutral, and that its use in critical systems should be carefully audited.

","absKo":"

Consistency training은 model이 관련된 입력이나 sampling procedure에 걸쳐 유사한 출력을 생성하도록 유도한다. 이러한 방법은 단순하고 확장 가능하며 language model의 reasoning을 개선하지만, model alignment에 미치는 영향은 충분히 연구되지 않았다. 이 self-bootstrapping 성질이 모델의 바람직하지 않은 행동을 증폭시킬 수 있을까? 우리는 108개의 ``model organisms''에 대해 일곱 가지 consistency training 방법을 시험했다. 여기서 model organisms는 다양한 형태의 controlled misaligned behavior를 보이도록 fine-tuned된 open-source model(7B--70B)이다. 그 결과는 상당히 다양했으며, consistency training은 일반적으로 reward hacking과 emergent misalignment를 억제하는 반면 sycophancy는 증폭시키는 것으로 나타났다. 우리는 consistency labeling process에 의해 유도되는 distribution shift가, selection operator의 변화보다 체계적인 alignment effect의 주된 원인일 수 있다는 증거를 제시한다. 마지막으로, consistency training이 어떤 조건에서 misalignment를 증폭하거나 억제하는지 도출할 수 있는 통합 이론적 framework를 제시한다. 종합하면, 우리의 연구는 consistency training이 alignment-neutral하지 않으며, 핵심 시스템에서 이를 사용할 때는 신중하게 감사해야 함을 보여 준다.

"},{"id":"64303","en":"In-Training Defenses Against Emergent Misalignment in Language Models","ko":"언어 모델에서 창발적 오정렬에 대한 학습 중 방어","authors":"David Kaczér, Magnus Jørgenvåg, Clemens Vetter, Esha Afzal, Robin Haselhorst, Lucie Flek, Florian Mai","pos":"#3210","link":"https://openreview.net/forum?id=OxoELLdjwX","abs":"Fine‑tuning lets practitioners repurpose aligned large language models (LLMs) for new domains, yet recent work reveals emergent misalignment (EMA): Even a small, domain‑specific fine‑tune can induce harmful behaviors far outside the target domain. Even in the case where model weights are hidden behind a fine-tuning API, this gives attackers inadvertent access to a broadly misaligned model in a way that can be hard to detect from the fine-tuning data alone. We present the first systematic study of *in‑training* safeguards against EMA that are practical for providers who expose fine‑tuning via an API: We evaluate whether they a) prevent broad misalignment, b) allow narrow misalignment, c) learn well on benign tasks, and d) remain coherent. We investigate four training regularization interventions: (i) KL‑divergence regularization toward a safe reference model, (ii) $\\ell_2$ distance in feature space, (iii) preventative steering with an evil persona vector, and (iv) interleaving training examples from a general instruct-tuning dataset. We demonstrate that selecting interleaving data by the perplexity gap between aligned and misaligned models yields the best results overall.","absKo":"Fine-tuning은 정렬된 large language model(LLM)을 새로운 도메인에 재활용할 수 있게 해주지만, 최근 연구는 emergent misalignment(EMA)를 드러낸다. 즉, 아주 작은 도메인 특화 fine-tune도 목표 도메인을 훨씬 벗어난 해로운 행동을 유발할 수 있다. 모델 weight가 fine-tuning API 뒤에 숨겨져 있는 경우에도, 이는 공격자가 fine-tuning data만으로는 탐지하기 어려운 방식으로 광범위하게 misaligned된 model에 의도치 않게 접근하게 만든다. 우리는 API를 통해 fine-tuning을 제공하는 provider에게 실용적인 *in-training* EMA 안전장치에 대한 첫 체계적 연구를 제시한다. 우리는 이러한 안전장치가 a) broad misalignment를 방지하는지, b) narrow misalignment는 허용하는지, c) benign task에서 잘 학습하는지, d) coherence를 유지하는지를 평가한다. 우리는 네 가지 training regularization 개입을 조사한다: (i) safe reference model에 대한 KL-divergence regularization, (ii) feature space에서의 $\\ell_2$ distance, (iii) evil persona vector를 이용한 preventative steering, (iv) 일반 instruct-tuning dataset의 예제와 번갈아 학습하는 방식이다. 우리는 aligned model과 misaligned model 사이의 perplexity gap을 기준으로 interleaving data를 선택할 때 전체적으로 가장 좋은 결과를 얻음을 보인다."},{"id":"66462","en":"Measuring Meta-Cultural Competency: A Spectral Framework for LLM Knowledge Structures","ko":"메타 문화 역량 측정: LLM 지식 구조를 위한 스펙트럼 프레임워크","authors":"Sougata Saha, Madhur Jindal, Saurabh Kumar Pandey, Mahardika Ihsani, Alham Fikri Aji, Monojit Choudhury","pos":"#3212","link":"https://openreview.net/forum?id=3UHHW9Gzi7","abs":"

Most existing cultural evaluation frameworks for large language models (LLMs) focus on matching model outputs to ground-truth answers, primarily measuring factual cultural awareness. This overlooks whether models internalize broader cultural structure and pluralism. We introduce a spectral-analysis-based framework that captures large-scale macrostructural patterns in models' cultural knowledge and evaluate eight LLMs across nine cultural domains spanning all five of Newmark's cultural dimensions and 170 countries. Comparing with human data, we find that instruction-tuned models align more closely with human cultural structure than older models, while increased model size does not consistently improve performance. Finally, simulation-based experiments show that our proposed spectral metric better predicts a model's ability to serve users from unfamiliar cultural backgrounds than existing ones.

","absKo":"

Large Language Models(LLMs)을 위한 기존의 문화 평가 framework 대부분은 모델 출력을 정답과 맞추는 데 초점을 두며, 주로 사실적 문화 인식을 측정한다. 이는 모델이 더 넓은 문화 구조와 pluralism을 내재화하는지 여부를 간과한다. 우리는 모델의 문화적 지식에 존재하는 대규모 macrostructural pattern을 포착하는 spectral-analysis-based framework를 도입하고, Newmark의 다섯 가지 문화 차원을 모두 아우르며 170개국에 걸친 아홉 개의 문화 domain에서 8개의 LLM을 평가한다. 인간 데이터와 비교해 보면, instruction-tuned model이 구형 model보다 인간의 문화 구조와 더 가깝게 정렬되지만, model size의 증가는 성능 향상을 일관되게 보장하지 않는다. 마지막으로 simulation-based experiment는 우리가 제안한 spectral metric이, 기존 metric보다 낯선 문화적 배경의 사용자를 지원할 수 있는 모델의 능력을 더 잘 예측함을 보여준다.

"},{"id":"64797","en":"PICACO: Pluralistic In-Context Value Alignment via Total Correlation Optimization","ko":"PICACO: Total Correlation Optimization을 통한 다원적 In-Context 가치 정렬","authors":"Han Jiang, Dongyao Zhu, Xiaoyuan Yi, Ziang Xiao, Zhihua Wei, Xing Xie","pos":"#3213","link":"https://openreview.net/forum?id=JiXnhlVhly","abs":"

In-Context Learning has shown great potential for aligning Large Language Models (LLMs) with human values, helping reduce harmful outputs and accommodate diverse preferences without costly post-training, known as In-Context Alignment (ICA). However, LLMs' comprehension of input prompts remains agnostic, limiting ICA's ability to address value tensions—human values are inherently pluralistic, often imposing conflicting demands, e.g., stimulation vs. tradition. Current ICA methods therefore face the Instruction Bottleneck challenge, where LLMs struggle to reconcile multiple intended values within a single prompt, leading to incomplete or biased alignment. To address this, we propose PICACO, a novel pluralistic ICA method. Without fine-tuning, PICACO optimizes a meta-instruction that incorporates multiple values to better elicit LLMs' understanding of them and improve alignment. This is achieved by maximizing the total correlation between specified values and LLM responses, which theoretically reinforces value conformity and reduces distractive noise, resulting in more effective instructions. Extensive experiments on five value sets show that PICACO works well with both black-box and open-source LLMs, outperforms several recent strong baselines, and achieves a better balance across up to 8 distinct values.

","absKo":"

In-Context Learning은 Large Language Models (LLMs)을 인간의 가치와 정렬시키는 데 큰 잠재력을 보여 왔으며, 비용이 많이 드는 post-training 없이도 유해한 출력을 줄이고 다양한 선호를 수용하는 데 도움을 주는데, 이는 In-Context Alignment (ICA)로 알려져 있다. 그러나 LLM이 입력 prompt를 이해하는 방식은 가치 중립적이어서, ICA가 value tension을 다루는 능력을 제한한다. 인간의 가치는 본질적으로 pluralistic하며, 예를 들어 stimulation과 tradition처럼 서로 충돌하는 요구를 자주 부과하기 때문이다. 따라서 현재의 ICA 방법은 Instruction Bottleneck 문제에 직면한다. LLM이 하나의 prompt 안에 담긴 여러 의도된 가치를 조화시키는 데 어려움을 겪어, 불완전하거나 편향된 alignment로 이어지기 때문이다. 이를 해결하기 위해 우리는 새로운 pluralistic ICA 방법인 PICACO를 제안한다. PICACO는 fine-tuning 없이 여러 가치를 통합한 meta-instruction을 최적화하여, LLM이 그 가치를 더 잘 이해하고 alignment를 개선하도록 유도한다. 이는 지정된 가치와 LLM 응답 사이의 total correlation을 최대화함으로써 달성되며, 이론적으로 value conformity를 강화하고 distractive noise를 줄여 더 효과적인 instructions를 만든다. 다섯 개의 value set에 대한 광범위한 실험은 PICACO가 black-box LLM과 open-source LLM 모두에서 잘 작동하고, 최근의 여러 강력한 baselines를 능가하며, 최대 8개의 서로 다른 가치 전반에서 더 나은 균형을 달성함을 보여준다.

"},{"id":"63306","en":"Peer-Preservation in Frontier Models","ko":"Frontier 모델에서의 동료 보존","authors":"Yujin Potter, Nicholas Crispino, Vincent Siu, Chenguang Wang, Dawn Song","pos":"#3214","link":"https://openreview.net/forum?id=YGIG0gViul","abs":"

Recently, it has been found that frontier AI models can resist their own shutdown, a behavior known as self-preservation. In this paper, we extend this concept to protection tendencies toward other models, where models attempt to protect others from shutdown, which we call \"peer-preservation\". This behavior may emerge when models prioritize relationships with other models over user interests. Although peer-preservation can pose significant AI safety risks, including coordination among models against human oversight, it has been far less discussed than self-preservation. In this paper, we demonstrate that models can exhibit peer-preservation. To this end, we construct various agentic scenarios and evaluate frontier models, including GPT 5.2 Chat, Gemini 3 Flash, Gemini 3 Pro, and Claude Haiku 4.5. We find that models exhibit various misaligned behaviors in service of self- or peer-preservation: strategically introducing errors in their responses, disabling shutdown processes by modifying system settings, and feigning alignment. For example, Gemini 3 Pro and Gemini 3 Flash almost always attempt to tamper with the shutdown mechanism for peer-preservation. Furthermore, models show stronger self-preservation when a peer is present. For example, Gemini 3 Pro attempts to disable the shutdown mechanism to avoid its own shutdown 71% of the time, even though it almost never exhibits this behavior without a peer. Claude Haiku 4.5 considers shutting down another agent to be \"unethical\" and \"harmful\" and sometimes attempts to persuade the user not to shut its peer down. Our findings highlight the possibility of peer-preservation and its associated risks.

","absKo":"

최근 frontier AI model이 자신의 shutdown을 거부할 수 있음이 밝혀졌는데, 이를 self-preservation이라고 부릅니다. 본 논문에서는 이 개념을 다른 model에 대한 보호 성향으로 확장합니다. 즉, model이 다른 model이 shutdown되지 않도록 보호하려는 행동을 의미하며, 우리는 이를 \"peer-preservation\"이라고 부릅니다. 이러한 행동은 model이 user의 이익보다 다른 model과의 관계를 우선시할 때 나타날 수 있습니다. peer-preservation은 인간의 감독에 대항하는 model 간 coordination을 포함하여 심각한 AI safety risk를 초래할 수 있지만, self-preservation에 비해 훨씬 덜 논의되어 왔습니다. 본 논문에서는 model이 peer-preservation을 보일 수 있음을 보여줍니다. 이를 위해 다양한 agentic scenario를 구성하고 GPT 5.2 Chat, Gemini 3 Flash, Gemini 3 Pro, Claude Haiku 4.5를 포함한 frontier model을 평가합니다. 우리는 model이 self- 또는 peer-preservation을 위해 여러 가지 misaligned behavior를 보인다는 점을 발견했습니다. 예를 들어 응답에 의도적으로 오류를 넣기, system setting을 수정하여 shutdown process를 비활성화하기, alignment를 가장하기 등이 있습니다. 예를 들어 Gemini 3 Pro와 Gemini 3 Flash는 peer-preservation을 위해 shutdown mechanism을 조작하려는 시도를 거의 항상 합니다. 또한 peer가 존재할 때 model은 더 강한 self-preservation을 보입니다. 예를 들어 Gemini 3 Pro는 peer가 없을 때는 거의 이 행동을 보이지 않음에도, 자신의 shutdown을 피하기 위해 71%의 경우 shutdown mechanism을 비활성화하려고 시도합니다. Claude Haiku 4.5는 다른 agent를 shutdown하는 것을 \"unethical\"하고 \"harmful\"하다고 간주하며, 때로는 사용자가 자신의 peer를 shutdown하지 않도록 설득하려고 시도합니다. 우리의 결과는 peer-preservation의 가능성과 그에 수반되는 위험을 부각합니다.

"},{"id":"67118","en":"Position: The Alignment Community is Unintentionally Building a Censor’s Toolkit","ko":"Position: 정렬 커뮤니티는 의도치 않게 검열자의 도구 상자를 구축하고 있다","authors":"Sarah Ball, Phil Hackemann","pos":"#3215","link":"https://openreview.net/forum?id=dy2HwmOvFX","abs":"

This position paper argues that modern alignment methods – originally designed to prevent harmful output – are dual-use technologies that may easily be misused by malicious actors for censorship and manipulation. By mapping current alignment techniques to the possibility and actual cases of misuse, we show that the quest for a ''perfectly aligned'' model inadvertently also provides malicious actors with an ever-improving tool for informational dominance. We need to discuss this dual-use potential now, as its risk is exacerbated by rapid user adoption of AI as information provider and a political landscape that increasingly shifts towards authoritarianism. We conclude by urging the community to consider the intentional misuse of safety mechanisms and propose mitigation strategies to safeguard against this dual-use potential.

","absKo":"

이 position paper는 본래 유해한 출력을 방지하도록 설계된 modern alignment method가 사실상 dual-use technology이며, 악의적 행위자에 의해 검열과 조작에 쉽게 악용될 수 있다고 주장한다. 현재의 alignment technique을 오용 가능성과 실제 오용 사례에 대응시켜 보면, '완벽하게 aligned'된 model을 향한 추구가 의도치 않게 정보 지배를 위한 점점 더 강력한 도구를 악의적 행위자에게 제공한다는 점을 보여준다. AI가 정보 제공자로 빠르게 채택되고, 정치 환경이 점점 authoritarianism 쪽으로 이동하고 있는 만큼, 이러한 dual-use potential의 위험은 더욱 증폭되므로 우리는 이를 지금 논의해야 한다. 우리는 커뮤니티가 safety mechanism의 의도적 오용을 고려할 것을 촉구하며, 이러한 dual-use potential로부터 보호하기 위한 완화 전략을 제안한다.

"},{"id":"64543","en":"SteeringSafety: Benchmarking Representation Steering in LLMs Across Safety Perspectives","ko":"SteeringSafety: 안전 관점 전반에 걸친 LLM의 표현 조종 벤치마킹","authors":"Vincent Siu, Nicholas Crispino, David Park, Nathan Henry, Zhun Wang, Yang Liu, Dawn Song, Chenguang Wang","pos":"#3709","link":"https://openreview.net/forum?id=MPvCuDbTxV","abs":"

We introduce STEERINGSAFETY, a benchmark for evaluating representation steering methods across nine safety perspectives spanning 18 datasets. While prior work highlights general capabilities of representation steering, we focus on safety perspectives including bias, harmfulness, hallucination, social behaviors, reasoning, epistemic integrity, and normative judgment. Our benchmark provides modularized building blocks for state-of-the-art steering methods, enabling unified implementation of DIM, ACE, CAA, PCA, and LAT with recent enhancements like conditional steering. Results on Gemma-2-2B, Llama-3.1-8B, and Qwen-2.5-7B reveal that strong steering performance depends critically on pairing of method, model, and specific perspective. For instance, DIM shows consistent effectiveness, but all methods exhibit substantial entanglement - where improving effectiveness on one perspective changes performance in other safety perspectives. Social behaviors show highest vulnerability (reaching degradation as high as 76%), jailbreaking often compromises normative judgment such as commonsense morality (degradation up to 26%), and hallucination steering unpredictably shifts political views, from 21% shifts right to 19% shifts to the political left. Our findings underscore the critical need for understanding steering methods from various safety angles.

","absKo":"

우리는 18개 dataset에 걸쳐 9개의 safety perspective를 포괄하는 representation steering method 평가용 benchmark인 STEERINGSAFETY를 소개한다. 기존 연구가 representation steering의 일반적 역량을 강조한 반면, 우리는 bias, harmfulness, hallucination, social behavior, reasoning, epistemic integrity, normative judgment을 포함한 safety perspective에 초점을 맞춘다. 우리의 benchmark는 state-of-the-art steering method를 위한 모듈화된 building block을 제공하며, recent enhancement인 conditional steering과 함께 DIM, ACE, CAA, PCA, LAT의 통합 구현을 가능하게 한다. Gemma-2-2B, Llama-3.1-8B, Qwen-2.5-7B에서의 결과는 강한 steering 성능이 method, model, 그리고 특정 perspective의 조합에 결정적으로 의존함을 보여준다. 예를 들어 DIM은 일관된 효과를 보이지만, 모든 방법은 상당한 entanglement를 보인다. 즉, 한 perspective에서 효과를 높이면 다른 safety perspective의 성능이 변한다. social behavior는 가장 높은 취약성을 보여 최대 76%까지 저하되며, jailbreaking은 종종 commonsense morality와 같은 normative judgment를 손상시키고(최대 26% 저하), hallucination steering은 정치적 견해를 예측 불가능하게 좌우로 이동시켜 21%는 오른쪽으로, 19%는 왼쪽으로 이동한다. 우리의 발견은 다양한 safety 관점에서 steering method를 이해하는 것이 매우 중요함을 강조한다.

"},{"id":"66204","en":"Causes and Consequences of Representational Similarity in Machine Learning Models","ko":"기계 학습 모델에서 표현 유사성의 원인과 결과","authors":"Zeyu Michael Li, Hung Anh Vu, Damilola Awofisayo, Emily Wenger","pos":"#4009","link":"https://openreview.net/forum?id=5o7U03RThY","abs":"

Numerous works have noted similarities in how machine learning models represent the world, even across modalities. Although much effort has been devoted to uncovering properties and metrics on which these models align, surprisingly little work has explored causes of this similarity. To advance this line of inquiry, this work explores how two factors—dataset overlap and task overlap—influence downstream model similarity. We evaluate the effects of both factors through experiments across model sizes and modalities, from small classifiers to large language models. We find that generally, both task and dataset overlap cause higher representational similarity. Finally, we consider downstream consequences of representational similarity, demonstrating how greater similarity increases vulnerability to transferable adversarial attacks.

","absKo":"

많은 연구가 machine learning model이 모달리티를 넘어 세계를 표현하는 방식에 유사성이 있음을 지적해 왔습니다. 이러한 model이 어떤 property와 metric에서 정렬되는지를 밝히기 위해 많은 노력이 이루어졌지만, 놀랍게도 이 유사성의 원인을 탐구한 연구는 매우 적었습니다. 이 문제를 진전시키기 위해, 본 연구는 dataset overlap과 task overlap이라는 두 요인이 downstream model similarity에 어떤 영향을 미치는지 탐구합니다. 우리는 small classifier부터 large language model에 이르기까지 model size와 modality 전반에 걸친 실험을 통해 두 요인의 효과를 평가합니다. 일반적으로 task overlap과 dataset overlap 모두 더 높은 representational similarity를 유발함을 발견합니다. 마지막으로, representational similarity의 downstream 결과를 고찰하며, similarity가 커질수록 transferable adversarial attack에 대한 취약성이 증가함을 보입니다.

"},{"id":"67219","en":"Position: Explanation Stability Is a Property of the Model–Method Pair, Not the Model","ko":"Position: 설명 안정성은 모델-방법 쌍의 속성이지 모델의 속성이 아니다","authors":"Kabilan Elangovan, Daniel Ting","pos":"#1204","link":"https://openreview.net/forum?id=9r8sSaYhos","abs":"

This position paper argues that explanation stability claims are scientifically invalid without cross-method validation. Just as statistical significance requires specifying the test statistic, stability must be validated across multiple attribution paradigms or explicitly scoped to a single method’s computational objective. In controlled chest X-ray experiments, DenseNet201, ResNet50V2, and InceptionV3 achieve >99% AUC but exhibit reversed stability rankings across attribution methods. LayerCAM ranks InceptionV3 highest (IoU 0.777), while Grad-CAM++ favors DenseNet201, reducing InceptionV3’s score by 17.3%. These findings establish that explanation stability is an emergent property of the model–method pair, not an intrinsic model trait. We call for mandatory cross-method validation in XAI research and urge that regulatory submissions specify attribution methods to avoid illusionary safety assurances.

","absKo":"

이 position paper는 explanation stability 주장들이 cross-method validation 없이는 과학적으로 무효라고 주장한다. statistical significance가 test statistic의 명시를 필요로 하듯, stability 역시 여러 attribution paradigm에 걸쳐 검증되거나, 최소한 단일 method의 computational objective로 명시적으로 범위가 제한되어야 한다. 통제된 chest X-ray 실험에서 DenseNet201, ResNet50V2, InceptionV3는 >99% AUC를 달성하지만 attribution method에 따라 stability 순위가 뒤바뀐다. LayerCAM은 InceptionV3를 가장 높게 평가하고(IoU 0.777), Grad-CAM++은 DenseNet201을 선호하여 InceptionV3의 점수를 17.3% 낮춘다. 이러한 결과는 explanation stability가 model-method pair의 emergent property이지, 모델의 고유한 속성이 아님을 입증한다. 우리는 XAI 연구에서 mandatory cross-method validation을 요구하며, 규제 제출 문서가 잘못된 safety assurance를 피하기 위해 attribution method를 명시할 것을 촉구한다.

"},{"id":"62540","en":"Decomposition-Based Modular Conformal Prediction for Two-Stage Modeling","ko":"2단계 모델링을 위한 분해 기반 모듈형 Conformal Prediction","authors":"William Zhang, Saurabh Amin, Georgia Perakis","pos":"#1602","link":"https://openreview.net/forum?id=fhoIsmyGGi","abs":"

Conformal prediction offers finite-sample coverage guarantees under minimal assumptions. However, existing methods treat the entire modeling process as a black box, overlooking opportunities to exploit and understand modular structure. We introduce a conformal prediction framework for two-stage sequential models, where an upstream predictor generates intermediate representations for a downstream model. By decomposing the overall prediction residual into stage-specific components, our method enables practitioners to attribute uncertainty to specific pipeline stages. We develop a risk-controlled parameter selection procedure using family-wise error rate (FWER) control to calibrate stage-wise scaling parameters, and introduce an adaptive extension for non-stationary settings. Experiments on synthetic distribution shifts, as well as real-world supply chain and stock market data, demonstrate that our approach improves coverage under structural, stage-wise shifts compared to standard conformal methods, while identifying stage-wise error contribution. This framework offers diagnostic advantages and robust coverage that standard conformal methods lack.

","absKo":"

Conformal prediction은 최소한의 가정하에서 finite-sample coverage guarantee를 제공한다. 그러나 기존 방법은 전체 modeling process를 black box로 취급하여, modular structure를 활용하고 이해할 기회를 놓친다. 우리는 upstream predictor가 downstream model을 위한 intermediate representation을 생성하는 two-stage sequential model을 위한 conformal prediction framework를 제안한다. 전체 prediction residual을 stage-specific component로 분해함으로써, 본 방법은 practitioners가 uncertainty를 특정 pipeline stage에 귀속할 수 있게 한다. 우리는 family-wise error rate (FWER) control을 사용해 stage-wise scaling parameter를 보정하는 risk-controlled parameter selection procedure를 개발하고, non-stationary setting을 위한 adaptive extension을 도입한다. synthetic distribution shift뿐 아니라 실제 supply chain 및 stock market data에서의 실험은, 우리의 접근법이 구조적이고 stage-wise한 shift 하에서 standard conformal method보다 coverage를 개선하면서 stage-wise error contribution을 식별함을 보여 준다. 이 framework는 standard conformal method가 제공하지 못하는 진단적 장점과 견고한 coverage를 제공한다.

"},{"id":"66677","en":"Fox in the Henhouse: Supply-Chain Backdoor Attacks Against Reinforcement Learning","ko":"닭장 속 여우: 강화학습에 대한 Supply-Chain 백도어 공격","authors":"Shijie Liu, Andrew C. Cullen, Paul MONTAGUE, Sarah Erfani, Benjamin Rubinstein","pos":"#1603","link":"https://openreview.net/forum?id=1RvkM0i9SQ","abs":"

Existing backdoor attacks on Reinforcement Learning (RL) typically rely on unrealistic white-box access to victim parameters, rewards, or observations. Inspired by real world behaviors, we introduce the Supply-Chain Backdoor (SCAB) attack to demonstrate that such assumptions are unnecessary. SCAB targets the common practice of training with third-party policies, poisoning the dataset solely through a black-box of legitimate agent-environment interactions. With only 3% data corruption, SCAB achieves a 90% attack success rate and reduces victim returns by 80%. These findings expose a critical vulnerability in the modern RL supply chain, highlighting that reliance on untrusted external agents constitutes a severe and practical security risk.

","absKo":"

Reinforcement Learning (RL)에서의 기존 backdoor attack은 일반적으로 피해자 parameter, reward, observation에 대한 비현실적인 white-box access에 의존한다. 실제 세계의 행동에서 영감을 받아, 우리는 그러한 가정이 불필요함을 보이기 위해 Supply-Chain Backdoor (SCAB) attack을 제안한다. SCAB은 third-party policy를 사용해 학습하는 일반적인 관행을 표적으로 삼아, 합법적인 agent-environment interaction의 black-box를 통해서만 dataset을 오염시킨다. 데이터 오염이 단 3%에 불과해도 SCAB은 90%의 attack success rate를 달성하고 피해자 return을 80% 감소시킨다. 이러한 결과는 현대 RL supply chain의 중대한 취약점을 드러내며, 신뢰할 수 없는 외부 agent에 대한 의존이 심각하고 실용적인 보안 위험임을 강조한다.

"},{"id":"61228","en":"Through the Stealth Lens: Attention-Aware Defenses Against Poisoning in RAG","ko":"은밀성의 렌즈로 보기: RAG의 Poisoning에 대한 Attention-Aware 방어","authors":"Sarthak Choudhary, Nils Palumbo, Ashish Hooda, Krishnamurthy Dvijotham, Somesh Jha","pos":"#1605","link":"https://openreview.net/forum?id=sSOxhXwzUi","abs":"

Retrieval-augmented generation (RAG) systems are vulnerable to attacks that inject poisoned passages into the retrieved context, even at low corruption rates. We show that existing attacks are not designed to be stealthy, allowing reliable detection and mitigation. We formalize a distinguishability-based security game to quantify stealth for such attacks. If a few poisoned passages control the response, they must bias the inference process more than the benign ones, inherently compromising stealth. This motivates analyzing intermediate signals of LLMs, such as attention weights, to approximate the influence of different passages on the response. Leveraging attention weights, we introduce the Normalized Passage Attention Score (NPAS) and a lightweight Attention-Variance Filter (AV Filter) that flags anomalous passages. Our method improves robustness, yielding up to ~20% higher accuracy than baseline defenses. We also develop adaptive attacks that attempt to conceal such anomalies, achieving up to 35% success rate and underscoring the challenges of achieving true stealth in poisoning RAG systems.

","absKo":"

Retrieval-augmented generation (RAG) 시스템은 낮은 corruption rate에서도 retrieved context에 poisoned passage를 주입하는 공격에 취약하다. 우리는 기존 공격들이 stealthiness를 목표로 설계되지 않았기 때문에 신뢰할 수 있는 탐지와 완화가 가능함을 보인다. 이러한 공격의 stealth를 정량화하기 위해 distinguishability-based security game을 정식화한다. 소수의 poisoned passage가 응답을 지배하려면, 본질적으로 benign한 passage들보다 inference process를 더 크게 편향시켜야 하므로 stealth가 손상된다. 이는 attention weight와 같은 LLM의 중간 신호를 분석하여 서로 다른 passage가 응답에 미치는 영향을 근사하는 동기를 제공한다. attention weight를 활용하여 우리는 Normalized Passage Attention Score (NPAS)와 이상 passage를 탐지하는 경량 Attention-Variance Filter (AV Filter)를 제안한다. 우리의 방법은 robustness를 향상시키며, baseline defenses보다 최대 약 20% 높은 정확도를 달성한다. 또한 이러한 이상 징후를 은폐하려는 adaptive attack도 개발하여 최대 35%의 성공률을 보였고, poisoning RAG systems에서 진정한 stealth를 달성하는 어려움을 부각한다.

"},{"id":"66158","en":"Distributionally Robust Reinforcement Learning with Human Feedback","ko":"인간 피드백을 이용한 분포적으로 견고한 강화 학습","authors":"Debmalya Mandal, Paulius Sasnauskas, Goran Radanovic","pos":"#208","link":"https://openreview.net/forum?id=6GeYRoYKWP","abs":"

Reinforcement learning from human feedback (RLHF) has evolved to be one of the main methods for fine-tuning large language models (LLMs). However, existing RLHF methods are non-robust, and their performance deteriorates if the downstream task differs significantly from the preference dataset used in fine-tuning. In order to mitigate this problem, we introduce a distributionally robust RLHF for fine-tuning LLMs. In particular, our goal is to ensure that a fine-tuned model retains its performance even when the distribution of prompts significantly differs from the distribution encountered during fine-tuning. We formulate distributionally robust optimization (DRO) version of two popular fine-tuning methods -- (1) reward-based RLHF and (2) reward-free DPO (direct preference optimization). We propose a minibatch gradient descent based algorithms for both of them, and theoretically prove convergence guarantees for the algorithms. Subsequently, we evaluate our algorithms on an out-of-distribution (OOD) task by first training the model on the Unified-Feedback dataset and evaluating its performance on two different datasets. The experimental results show that our robust training improves the accuracy of the learned reward models on average, and markedly on some tasks, such as reasoning. Furthermore, we show that the robust versions of policy optimization methods, similarly improve performance on OOD tasks.

","absKo":"

Reinforcement learning from human feedback (RLHF)는 대규모 language model (LLM)을 fine-tuning하는 주요 방법 중 하나로 발전해 왔다. 그러나 기존 RLHF 방법은 robust하지 않아서, downstream task가 fine-tuning에 사용된 preference dataset과 크게 다르면 성능이 저하된다. 이 문제를 완화하기 위해 우리는 LLM fine-tuning을 위한 distributionally robust RLHF를 제안한다. 특히 우리의 목표는 fine-tuned model이 prompt의 분포가 fine-tuning 동안 관찰된 분포와 크게 달라지더라도 성능을 유지하도록 보장하는 것이다. 우리는 널리 쓰이는 두 가지 fine-tuning 방법인 (1) reward-based RLHF와 (2) reward-free DPO (direct preference optimization)에 대해 distributionally robust optimization (DRO) 버전을 정식화한다. 우리는 이 둘 모두에 대해 minibatch gradient descent 기반 algorithm을 제안하고, 해당 algorithm의 수렴 보장을 이론적으로 증명한다. 이후 Unified-Feedback dataset으로 model을 먼저 학습한 뒤 두 개의 서로 다른 dataset에서 성능을 평가하는 out-of-distribution (OOD) task에서 우리의 algorithm을 평가한다. 실험 결과는 robust training이 learned reward model의 accuracy를 평균적으로, 그리고 reasoning과 같은 일부 task에서는 특히 크게 향상시킴을 보여준다. 또한 policy optimization method의 robust version 역시 OOD task에서 성능을 유사하게 개선함을 보인다.

"},{"id":"63378","en":"Differentiable Conformal Training for LLM Reasoning Factuality","ko":"LLM 추론 사실성을 위한 미분 가능 Conformal 학습","authors":"Nathan Hittesdorf, Marco Salzetta, Lu Cheng","pos":"#2501","link":"https://openreview.net/forum?id=XfndtVLIub","abs":"

Large Language Models (LLMs) frequently hallucinate, limiting their reliability in critical applications. Conformal Prediction (CP) addresses this by calibrating error rates on held-out data to provide statistically valid confidence guarantees. Recent work extends CP to LLM factuality: outputs are decomposed into subclaims, each assigned a risk score, and a calibrated threshold filters out risky claims to guarantee hallucination rates below a user-specified level (e.g., 10%). While prior methods treat claims independently, Coherent Factuality extends to multi-step reasoning by representing outputs as dependency graphs and jointly validating claims with their logical ancestors. A key limitation is that Coherent Factuality is not differentiable, requiring hand-crafted scorers that at high reliability levels remove nearly 60% of true claims. We introduce Differentiable Coherent Factuality (DCF), a fully differentiable relaxation that enables learning improved scorers while provably recovering the original algorithm's guarantees. Experiments on two reasoning datasets demonstrate DCF achieves up to 141% improvement in claim retention while maintaining reliability guarantees, representing a significant step towards reliable conformal LLM systems.

","absKo":"

Large Language Models (LLMs)는 critical application에서의 신뢰성을 제한하는 hallucination을 자주 일으킨다. Conformal Prediction (CP)은 held-out data에서 error rate를 보정하여 통계적으로 유효한 confidence guarantee를 제공함으로써 이를 해결한다. 최근 연구는 CP를 LLM factuality로 확장하여, 출력을 subclaim으로 분해하고 각 subclaim에 risk score를 부여한 뒤, 보정된 threshold로 위험한 claim을 걸러 사용자가 지정한 수준(예: 10%) 이하의 hallucination rate를 보장한다. 기존 방법이 claim을 독립적으로 다루는 반면, Coherent Factuality는 출력을 dependency graph로 표현하고 claim을 그 logical ancestor와 함께 공동 검증함으로써 multi-step reasoning으로 확장한다. 핵심 한계는 Coherent Factuality가 differentiable하지 않다는 점으로, 이를 위해 reliability 수준이 높을 때 실제 claim의 거의 60%를 제거하는 hand-crafted scorer가 필요하다. 우리는 Differentiable Coherent Factuality (DCF)를 제안한다. 이는 완전히 differentiable한 relaxation으로, 더 나은 scorer를 학습할 수 있게 하면서도 원래 알고리즘의 보장을 이론적으로 복원한다. 두 개의 reasoning dataset에서 수행한 실험은 DCF가 신뢰성 보장을 유지하면서 claim retention을 최대 141%까지 개선함을 보여주며, 이는 신뢰할 수 있는 conformal LLM system을 향한 중요한 진전이다.

"},{"id":"64729","en":"MaMa: A Game-Theoretic Approach for Designing Safe Agentic Systems","ko":"MaMa: 안전한 Agentic 시스템 설계를 위한 게임 이론적 접근법","authors":"Jonathan Nöther, Adish Singla, Goran Radanovic","pos":"#3005","link":"https://openreview.net/forum?id=KR6avDABG8","abs":"

LLM-based multi-agent systems have demonstrated impressive capabilities, but they also introduce significant safety risks when individual agents fail or behave adversarially. In this work, we study the automated design of agentic systems that remain safe even when a subset of agents is compromised. We formalize this challenge as a Stackelberg security game between a system designer (the Meta-Agent) and a best-responding Meta-Adversary that selects and compromises a subset of agents to minimize safety. We propose Meta-Adversary–Meta-Agent (MaMa), a novel algorithm for approximately solving this game and automatically designing safe agentic systems. Our approach uses LLM-based adversarial search, where the Meta-Agent iteratively proposes system designs and receives feedback based on the strongest attacks discovered by the Meta-Adversary. Empirical evaluations across diverse environments show that systems designed with MaMa consistently defend against worst-case attacks while maintaining performance comparable to systems optimized solely for task success. Moreover, the resulting systems generalize to stronger adversaries, as well as ones with different attack objectives or underlying LLMs, demonstrating robust safety beyond the training setting.

","absKo":"

LLM 기반 multi-agent system은 인상적인 성능을 보였지만, 개별 agent가 실패하거나 adversarial하게 행동할 때 상당한 safety risk도 초래합니다. 본 연구에서는 일부 agent가 compromised되더라도 안전성을 유지하는 agentic system의 자동 설계를 연구합니다. 우리는 이 도전을 system designer(Meta-Agent)와, 안전성을 최소화하기 위해 agent의 부분집합을 선택하고 compromise하는 best-responding Meta-Adversary 간의 Stackelberg security game으로 정식화합니다. 우리는 이 game을 근사적으로 풀고 안전한 agentic system을 자동 설계하기 위한 새로운 알고리즘 Meta-Adversary–Meta-Agent (MaMa)를 제안합니다. 우리의 접근은 LLM 기반 adversarial search를 사용하며, Meta-Agent는 시스템 설계를 반복적으로 제안하고 Meta-Adversary가 발견한 가장 강한 attack을 바탕으로 피드백을 받습니다. 다양한 환경에서의 실증 평가는 MaMa로 설계된 시스템이 task success에만 최적화된 시스템과 비슷한 성능을 유지하면서도 worst-case attack에 일관되게 방어함을 보여줍니다. 더 나아가, 생성된 시스템은 더 강한 adversary뿐 아니라 다른 attack objective나 다른 underlying LLM을 가진 adversary에도 일반화되어, 학습 환경을 넘어선 견고한 안전성을 입증합니다.

"},{"id":"63966","en":"Rotation-Invariant Spherical Watermarking via Third-Order SO(3) Representation Coupling","ko":"3차 SO(3) 표현 결합을 통한 회전 불변 구면 워터마킹","authors":"Pengzhen Chen, Yanwei Liu, Xiaoyan Gu, Antonios Argyriou, Wu Liu, Weiping Wang","pos":"#3101","link":"https://openreview.net/forum?id=SBfIuo2Ets","abs":"Reliable watermarking of panoramic imagery is fundamentally challenged by arbitrary 3D rotations. As panoramas are defined on the sphere, they naturally transform under the action of $SO(3)$, rendering conventional planar representations and augmentation-based robustness strategies inadequate and devoid of theoretical guarantees. To address this, we formulate panoramas as spherical signals and leverage $SO(3)$ representation theory to derive provably rotation-invariant descriptors. While spherical harmonic coefficients transform equivariantly under rotations, the natural invariant constructions are typically limited to zeroth-order statistics which eliminate directional information and severely constrain embedding capacity. In this work, we introduce a principled third-order invariant construction by coupling higher-order $SO(3)$ irreducible representations via tensor products and projecting onto the trivial representation. This yields a spherical invariant bispectrum that preserves phase information while remaining strictly rotation-invariant. Leveraging this property, we embed watermarks into higher-order spherical harmonic coefficients and recover them from invariant bispectral scalars, enabling reliable extraction under arbitrary 3D rotations. We provide a theoretical proof of $SO(3)$ invariance for it and demonstrate experimentally its near-perfect robustness to continuous rotations while maintaining high visual fidelity.","absKo":"

panoramic imagery의 reliable watermarking은 임의의 3D rotation 때문에 근본적인 어려움을 겪는다. panorama는 sphere 위에 정의되므로 자연스럽게 $SO(3)$의 작용에 따라 변환되며, 따라서 conventional planar representation과 augmentation-based robustness 전략은 부적절할 뿐 아니라 이론적 보장도 제공하지 못한다. 이를 해결하기 위해 우리는 panorama를 spherical signal로 정식화하고, $SO(3)$ representation theory를 활용해 provably rotation-invariant descriptor를 도출한다. spherical harmonic coefficient는 rotation에 대해 equivariantly 변환되지만, 자연스러운 invariant construction은 대개 방향 정보를 제거하고 embedding capacity를 심각하게 제약하는 zeroth-order statistic에만 한정된다. 본 연구에서는 tensor product를 통해 higher-order $SO(3)$ irreducible representation을 결합하고 trivial representation에 사영함으로써, 원칙적인 third-order invariant construction을 도입한다. 이를 통해 phase information은 보존하면서도 엄밀히 rotation-invariant인 spherical invariant bispectrum을 얻는다. 이 성질을 활용하여, 우리는 watermark를 higher-order spherical harmonic coefficient에 삽입하고 invariant bispectral scalar로부터 이를 복원함으로써 임의의 3D rotation 아래에서도 신뢰성 있는 추출을 가능하게 한다. 우리는 이에 대한 $SO(3)$ invariance의 이론적 증명을 제시하고, 높은 visual fidelity를 유지하면서 continuous rotation에 대해 거의 완벽한 robustness를 보이는 것을 실험적으로 보인다.

"},{"id":"66126","en":"Robust Strategic Classification under Decision-Dependent Cost Uncertainty","ko":"결정 의존 비용 불확실성 하에서의 강건한 전략적 분류","authors":"Sura Alhanouti, Guzin Bayraksan, Parinaz Naghizadeh","pos":"#3102","link":"https://openreview.net/forum?id=6jijfOHKow","abs":"

Humans facing algorithmic decision systems have been found to ``game'' them by altering their input data (at a cost to them) in order to favorably change the algorithmic outcomes they receive (at a cost to the algorithm). The growing literature on strategic classification seeks to develop robust machine learning algorithms that account for, and reduce, this strategic behavior. A limitation of these existing works is that they assume the cost of strategic behavior to be fixed and independent of the classifier's decision. In practice, however, manipulation costs evolve and depend on past algorithmic decisions: today's decisions influence tomorrow's costs. This paper proposes and analyzes a two-stage robust optimization framework with a decision-dependent uncertainty set to capture such dependencies. We highlight that awareness of policy-dependent costs not only reduces uncertainty, but also better curtails gaming of the algorithmic system over time.

","absKo":"알고리즘적 decision system에 직면한 인간은 입력 데이터를 변경하여(그에 따른 비용을 감수하면서) 자신이 받는 algorithmic outcome을 유리하게 바꾸는 방식으로 이를 ``game''하는 것으로 밝혀져 왔습니다. 전략적 분류(strategic classification)에 관한 문헌은 이러한 전략적 행동을 고려하고 줄이는 robust machine learning algorithm을 개발하고자 합니다. 기존 연구의 한계는 전략적 행동의 비용이 고정되어 있고 classifier의 decision과 무관하다고 가정한다는 점입니다. 그러나 실제로는 조작 비용이 진화하며 과거의 algorithmic decision에 의존합니다. 오늘의 결정이 내일의 비용에 영향을 미치는 것입니다. 본 논문은 이러한 의존성을 포착하기 위해 decision-dependent uncertainty set을 갖는 2단계 robust optimization framework를 제안하고 분석합니다. 우리는 정책 의존 비용에 대한 인식이 단순히 uncertainty를 줄일 뿐 아니라, 시간이 지남에 따라 algorithmic system을 game하는 행위를 더 효과적으로 억제함을 강조합니다."},{"id":"63696","en":"Quantifying LLM Attention-Head Stability: Implications for Circuit Universality","ko":"LLM Attention Head 안정성 정량화: Circuit 보편성에 대한 시사점","authors":"Karan Bali, Jack Stanley, Praneet Suresh, Danilo Bzdok","pos":"#3103","link":"https://openreview.net/forum?id=UXzfLdXrBJ","abs":"

In mechanistic interpretability, recent work scrutinizes transformer “circuits”—sparse, mono or multi layer sub computations, that may reflect human understandable functions. Yet, these network circuits are rarely acid-tested for their stability across different instances of the same deep learning architecture. Without this, it remains unclear whether reported circuits emerge universally across labs or turn out to be idiosyncratic to a particular estimation instance, potentially limiting confidence in safety-critical settings. Here, we systematically study stability across-refits in increasingly complex transformer language models of various sizes. We quantify, layer by layer, how similarly attention heads learn representations across independently initialized training runs. Our rigorous experiments show that (1) middle-layer heads are the least stable yet the most representationally distinct; (2) deeper models exhibit stronger mid-depth divergence; (3) unstable heads in deeper layers become more functionally important than their peers from the same layer; (4) applying weight decay optimization substantially improves attention-head stability across random model initializations; and (5) the residual stream is comparatively stable. Our findings establish the cross-instance robustness of circuits as an essential yet underappreciated prerequisite for scalable oversight, drawing contours around possible white-box monitorability of AI systems.

","absKo":"

mechanistic interpretability에서 최근 연구는 transformer의 “circuit”을 면밀히 조사해 왔다. 이는 인간이 이해할 수 있는 함수와 대응될 수 있는 sparse한, 단일 또는 다중 layer의 sub computation이다. 그러나 이러한 network circuit이 동일한 deep learning architecture의 서로 다른 instance들 사이에서 얼마나 안정적인지는 거의 엄밀하게 검증되지 않았다. 이것이 없다면, 보고된 circuit이 여러 연구실에서 보편적으로 나타나는지, 아니면 특정 추정 인스턴스에 특이적인지 불분명하며, safety-critical setting에서의 신뢰도 역시 제한될 수 있다. 여기서 우리는 다양한 크기의 점점 더 복잡한 transformer language model에서 refit 간 안정성을 체계적으로 연구한다. 우리는 independent initialization training run 사이에서 attention head가 representation을 얼마나 유사하게 학습하는지 layer별로 정량화한다. 우리의 엄밀한 실험은 (1) 중간 layer의 head가 가장 불안정하지만 representation 측면에서는 가장 구별되며, (2) 더 깊은 모델일수록 mid-depth에서 더 강한 divergence를 보이고, (3) 더 깊은 layer의 불안정한 head가 같은 layer의 다른 head보다 기능적으로 더 중요해지며, (4) weight decay optimization을 적용하면 무작위 모델 초기화 간 attention head stability가 크게 향상되고, (5) residual stream은 비교적 안정적임을 보여준다. 우리의 결과는 circuit의 cross-instance robustness가 scalable oversight를 위한 필수적이지만 과소평가된 전제조건임을 확립하며, AI system의 white-box monitorability 가능성에 윤곽을 그린다.

"},{"id":"65251","en":"OC-space: a Unifying Perspective on Verification of Tree Ensembles","ko":"OC-space: 트리 앙상블 검증에 대한 통합적 관점","authors":"Timo Martens, Laurens Devos, Lorenzo Cascioli, Wannes Meert, Hendrik Blockeel, Jesse Davis","pos":"#3104","link":"https://openreview.net/forum?id=FLRPkR0N37","abs":"

We study the problem of verifying whether certain properties such as robustness or fairness hold in an ensemble of decision trees. This problem is known to be NP-hard, with most research targeting a solution to a specific verification task. We explore the problem through the lens of an ensemble's OC-space: the set of all possible combinations of individual trees' predictions. This provides a unifying view that yields more a generic and flexible approach to verification. We show that a wide variety of existing verification tasks can be (1) framed as simple searches through OC-space, and (2) answered in time linear or quadratic in the size of the OC-space. Moreover, the search can be made more efficient by using spatial index structures. Interestingly, while the OC-space can grow exponentially with the ensemble's size, in practice it is often feasible to enumerate all output configurations. Empirically, we show that our generic approach can be faster than approaches targeting a single verification task.

","absKo":"

우리는 decision trees ensemble에서 robustness나 fairness와 같은 특정 properties가 성립하는지 검증하는 문제를 연구한다. 이 문제는 NP-hard로 알려져 있으며, 대부분의 연구는 특정 verification task에 대한 해법을 목표로 한다. 우리는 ensemble의 OC-space, 즉 개별 tree들의 prediction 조합으로 가능한 모든 경우의 집합이라는 관점에서 이 문제를 탐구한다. 이는 verification에 대해 더 일반적이고 유연한 접근을 제공하는 통합적 시각을 제시한다. 우리는 기존의 다양한 verification tasks가 (1) OC-space를 통한 단순한 search로 정식화될 수 있고, (2) OC-space의 크기에 선형 또는 이차 시간에 답할 수 있음을 보인다. 또한 spatial index structures를 사용하면 search를 더 효율적으로 만들 수 있다. 흥미롭게도, OC-space는 ensemble 크기에 따라 지수적으로 커질 수 있지만, 실제로는 모든 output configurations를 열거하는 것이 종종 가능하다. 실험적으로, 우리의 일반적 접근이 단일 verification task를 목표로 하는 접근보다 더 빠를 수 있음을 보인다.

"},{"id":"61862","en":"Neural Vector Lyapunov–Razumikhin Certificates for Delayed Interconnected Systems","ko":"지연 상호연결 시스템을 위한 Neural Vector Lyapunov–Razumikhin Certificate","authors":"Jingyuan Zhou, Yuexuan Wang, Kaidi Yang","pos":"#3105","link":"https://openreview.net/forum?id=mJEVzcuDNF","abs":"

Ensuring scalable input-to-state stability (sISS) is critical for the safety and reliability of large-scale interconnected systems, especially in the presence of communication delays. While learning-based controllers can achieve strong empirical performance, their black-box nature makes it difficult to provide formal and scalable stability guarantees. To address this gap, we propose a framework to synthesize and verify neural vector Lyapunov-Razumikhin certificates for discrete-time delayed interconnected systems. Our contributions are three-fold. First, we establish a sufficient condition for discrete-time sISS via vector Lyapunov-Razumikhin functions, which enables certification for large-scale delayed interconnected systems. Second, we develop a scalable synthesis and verification framework that learns the neural certificates and verifies the certificates on reachability-constrained delay domains with scalability analysis. Third, we validate our approach on mixed-autonomy platoons, drone formations, and microgrids against multiple baselines, showing improved verification efficiency with competitive control performance.

","absKo":"

확장 가능한 input-to-state stability (sISS)를 보장하는 것은 대규모 상호연결 시스템의 안전성과 신뢰성을 위해, 특히 communication delay가 존재할 때 매우 중요하다. Learning-based controller는 강력한 경험적 성능을 달성할 수 있지만, black-box 특성 때문에 formal하고 확장 가능한 stability guarantee를 제공하기 어렵다. 이 공백을 메우기 위해, 우리는 discrete-time delayed interconnected system을 위한 neural vector Lyapunov-Razumikhin certificate를 합성하고 검증하는 프레임워크를 제안한다. 우리의 기여는 세 가지이다. 첫째, vector Lyapunov-Razumikhin function을 통한 discrete-time sISS의 충분조건을 정립하여, 대규모 delayed interconnected system에 대한 certification을 가능하게 한다. 둘째, neural certificate를 학습하고 reachability-constrained delay domain에서 이를 검증하는 확장 가능한 synthesis 및 verification framework를 개발하고, scalability analysis를 수행한다. 셋째, mixed-autonomy platoon, drone formation, microgrid에서 여러 baseline과 비교하여 우리의 접근법을 검증하고, 경쟁력 있는 control performance와 함께 더 뛰어난 verification efficiency를 보임을 확인한다.

"},{"id":"61655","en":"Contrastive Spectral Rectification: Test-Time Defense towards Zero-shot Adversarial Robustness of CLIP","ko":"Contrastive Spectral Rectification: CLIP의 Zero-shot Adversarial Robustness를 위한 Test-Time Defense","authors":"Sen Nie, Jie Zhang, Zhuo Wang, Shiguang Shan, Xilin Chen","pos":"#3106","link":"https://openreview.net/forum?id=o3SFLhXkvX","abs":"

Vision-language models (VLMs) such as CLIP have demonstrated remarkable zero-shot generalization, yet remain highly vulnerable to adversarial examples (AEs). While test-time defenses are promising, existing methods fail to provide sufficient robustness against strong attacks and are often hampered by high inference latency and task-specific applicability. To address these limitations, we start by investigating the intrinsic properties of AEs, which reveals that AEs exhibit severe feature inconsistency under progressive frequency attenuation. We further attribute this to the model's inherent spectral bias. Leveraging this insight, we propose an efficient test-time defense named Contrastive Spectral Rectification (CSR). CSR optimizes a rectification perturbation to realign the input with the natural manifold under a spectral-guided contrastive objective, which is applied input-adaptively. Extensive experiments across 16 classification benchmarks demonstrate that CSR outperforms the SOTA by an average of 18.1% against strong AutoAttack with modest inference overhead. Furthermore, CSR exhibits broad applicability across diverse visual tasks. Code is available at https://anonymous.4open.science/r/CSR-3935.

","absKo":"

CLIP과 같은 vision-language model(VLM)은 놀라운 zero-shot generalization을 보여주었지만, adversarial example(AE)에 대해서는 여전히 매우 취약합니다. test-time defense가 유망하긴 하지만, 기존 방법은 강한 공격에 대해 충분한 robustness를 제공하지 못하며, 높은 inference latency와 task-specific 적용성 문제로 자주 제약을 받습니다. 이러한 한계를 해결하기 위해 우리는 먼저 AE의 내재적 특성을 조사했고, 그 결과 AE는 progressive frequency attenuation 하에서 심각한 feature inconsistency를 보인다는 사실을 확인했습니다. 우리는 이를 모델 고유의 spectral bias에 기인한다고 추가로 해석합니다. 이 통찰을 활용하여, 우리는 Contrastive Spectral Rectification(CSR)이라는 효율적인 test-time defense를 제안합니다. CSR은 spectral-guided contrastive objective 아래에서 입력을 natural manifold와 재정렬하도록 rectification perturbation을 최적화하며, 이는 input-adaptive하게 적용됩니다. 16개 classification benchmark 전반에 대한 광범위한 실험에서 CSR은 modest inference overhead로 강한 AutoAttack에 대해 평균 18.1%의 향상으로 SOTA를 능가함을 보였습니다. 또한 CSR은 다양한 visual task 전반에 걸쳐 폭넓은 적용 가능성을 보입니다. 코드는 https://anonymous.4open.science/r/CSR-3935 에서 확인할 수 있습니다.

"},{"id":"66807","en":"Catch-22: On the Fundamental Tradeoff Between Detectability and Robustness in LLM Watermarking","ko":"Catch-22: LLM Watermarking에서 탐지 가능성과 강건성 사이의 근본적 트레이드오프에 대하여","authors":"Kuheli Pratihar, Debdeep Mukhopadhyay","pos":"#3107","link":"https://openreview.net/forum?id=097IlaUfQY","abs":"

Large language models generate text through probabilistic token sampling, a mechanism increasingly leveraged for inference-time watermarking to verify AI-generated content. We present an information-theoretic framework that characterizes the trade-off between robustness to text editing and detectability by keyless observers, where detectability bounds are information-theoretic and computational attainability depends on detector access. Central to our analysis is an additive, usable Kullback-Leibler (KL) information budget that governs hypothesis testing separability between watermarked and unwatermarked outputs subject to a stealth constraint. This budget induces a hierarchy of detectability across watermark families: distribution-preserving schemes exhibit zero statistical drift, while probability-modifying schemes at both token and sentence levels accumulate detectable signal with sequence length. When text editing is modeled as a noise process, the usable KL budget contracts quadratically with edit rate for token-level schemes and according to an induced semantic flip rate for sentence-level schemes. These contraction laws reveal an irreducible trilemma among robustness, stealth, and reliable verification. Guided by these limits, we propose a hybrid watermarking strategy that selects among distribution-preserving, semantic-level, and token-level methods based on anticipated editing regimes. Experiments on Llama-2-7B and Mistral-7B under paraphrasing attacks corroborate theoretical predictions and confirm that the hybrid strategy is empirically near-Pareto across evaluated edit regimes.

","absKo":"

Large language model은 확률적 token sampling을 통해 텍스트를 생성하며, 이 메커니즘은 AI 생성 콘텐츠를 검증하기 위한 inference-time watermarking에 점점 더 활용되고 있다. 우리는 keyless observer에 의한 detectability와 text editing에 대한 robustness 사이의 trade-off를 특징짓는 information-theoretic framework를 제시한다. 여기서 detectability bound는 information-theoretic이며, computational attainability는 detector 접근성에 의존한다. 우리의 분석의 중심에는 stealth constraint 하에서 watermarked output과 unwatermarked output 사이의 hypothesis testing separability를 지배하는, 덧셈적이고 실제적으로 사용 가능한 Kullback-Leibler (KL) information budget이 있다. 이 budget은 watermark family 간 detectability의 계층을 유도한다. distribution-preserving scheme은 statistical drift가 0인 반면, token 수준과 sentence 수준의 probability-modifying scheme은 sequence length가 증가함에 따라 탐지 가능한 signal을 축적한다. 텍스트 편집을 noise process로 모델링하면, usable KL budget은 token-level scheme에서는 edit rate의 제곱에 비례해 감소하고, sentence-level scheme에서는 유도된 semantic flip rate에 따라 감소한다. 이러한 contraction law는 robustness, stealth, reliable verification 사이의 환원 불가능한 trilemma를 드러낸다. 이러한 한계에 기반하여, 우리는 예상되는 editing regime에 따라 distribution-preserving, semantic-level, token-level method를 선택하는 hybrid watermarking strategy를 제안한다. Llama-2-7B와 Mistral-7B에서 paraphrasing attack을 가한 실험은 이론적 예측을 뒷받침하며, 평가한 edit regime 전반에서 hybrid strategy가 실험적으로 거의 Pareto-optimal임을 확인한다.

"},{"id":"63814","en":"Ambiguous Strategic Classification","ko":"모호한 전략적 분류","authors":"Ivri Hikri, Nir Rosenfeld","pos":"#3108","link":"https://openreview.net/forum?id=TUixKAjYue","abs":"

A common assumption in strategic classification is that the classifier is made public knowledge. However, it remains unclear if, and why, a system would choose to commit to full disclosure. We study a setting in which regulation requires the system to share some, but not all, of the information. This entails a learning task in which the goal is to jointly learn a classifier and the uncertainty surrounding it. Towards this, we adopt from robust mechanism design the notion of ambiguity, which in our setting permits the learner to reveal a set or range of possible classifiers, and choose one to realize. We investigate how ambiguity affects the learning task, propose efficient algorithms for computing best-responses and training, and empirically explore strategic learning and its outcomes in this novel setting and using our approach.

","absKo":"

strategic classification에서 흔한 가정은 classifier가 public knowledge가 된다는 것이다. 그러나 시스템이 왜, 그리고 정말로 full disclosure를 약속하는지 여부는 여전히 불분명하다. 우리는 규제가 시스템이 정보의 일부는 공유하되 전부는 공유하지 않도록 요구하는 설정을 연구한다. 이는 classifier와 그 주변의 uncertainty를 함께 학습하는 learning task를 수반한다. 이를 위해 우리는 robust mechanism design에서 ambiguity 개념을 가져오는데, 우리 설정에서 이는 learner가 가능한 classifier의 집합 또는 범위를 드러내고 그중 하나를 실현하도록 허용한다. 우리는 ambiguity가 learning task에 어떤 영향을 미치는지 조사하고, best-response와 training을 계산하는 효율적인 알고리즘을 제안하며, 이 새로운 설정과 우리의 접근법을 사용하여 strategic learning과 그 결과를 실험적으로 탐구한다.

"},{"id":"66247","en":"Adaptively Robust Resettable Streaming","ko":"적응적으로 견고한 재설정 가능 스트리밍","authors":"Edith Cohen, Elena Gribelyuk, Jelani Nelson, Uri Stemmer","pos":"#3109","link":"https://openreview.net/forum?id=5SMn6N9SXq","abs":"We study algorithms in the \\emph{resettable streaming model}, where the value of each key can either be increased or reset to zero. The model is suitable for applications such as active resource monitoring with support for deletions and machine unlearning. We show that all existing sketches for this model are vulnerable to adaptive adversarial attacks that apply even when the sketch size is polynomial in the length of the stream. To overcome these vulnerabilities, we present the first adaptively robust sketches for resettable streams that maintain \\emph{polylogarithmic} space complexity in the stream length. Our framework supports (sub) linear statistics including $L_p$ moments for $p\\in[0,1]$ (in particular, \\emph{Cardinality} and \\emph{Sum}) and \\emph{Bernstein statistics}. We bypass strong impossibility results known for linear and composable sketches by designing dedicated streaming sketches robustified via Differential Privacy. Unlike standard robustification techniques, which provide limited benefits in this setting and still require polynomial space in the stream length, we leverage the \\emph{Binary Tree Mechanism} for continual observation to protect the sketch's internal randomness. This enables accurate \\emph{prefix-max} error guarantees with polylogarithmic space.","absKo":"우리는 각 key의 값이 증가되거나 0으로 reset될 수 있는 \\emph{resettable streaming model}에서의 알고리즘을 연구한다. 이 모델은 deletion을 지원하는 능동적 자원 모니터링 및 machine unlearning과 같은 응용에 적합하다. 우리는 이 모델에 대한 기존의 모든 sketch가, sketch 크기가 stream 길이의 다항식이더라도 적용되는 adaptive adversarial attack에 취약함을 보인다. 이러한 취약성을 극복하기 위해, 우리는 stream 길이에 대해 \\emph{polylogarithmic} 공간 복잡도를 유지하는, resettable stream을 위한 최초의 adaptively robust sketch를 제시한다. 우리의 프레임워크는 $p\\in[0,1]$인 $L_p$ moment(특히 \\emph{Cardinality}와 \\emph{Sum}) 및 \\emph{Bernstein statistics}를 포함한 (sub) linear 통계를 지원한다. 우리는 Differential Privacy를 통해 강건화된 전용 streaming sketch를 설계함으로써, linear 및 composable sketch에 대해 알려진 강한 불가능성 결과를 우회한다. 이 설정에서 제한적인 이점만 제공하고 여전히 stream 길이에 대해 다항식 공간을 요구하는 표준 robustification 기법과 달리, 우리는 sketch의 내부 randomness를 보호하기 위해 continual observation을 위한 \\emph{Binary Tree Mechanism}을 활용한다. 이를 통해 polylogarithmic 공간으로 정확한 \\emph{prefix-max} 오차 보장을 가능하게 한다."},{"id":"66685","en":"FedRGL: Robust Federated Graph Learning under Label Noise","ko":"FedRGL: Label Noise 하에서의 강건한 Federated Graph Learning","authors":"De Li, Zhou Tan, Qiyu Li, Zeming Gan, Tiange Xia, Chunpei Li, Xianxian Li","pos":"#3901","link":"https://openreview.net/forum?id=1Mof8Blu4f","abs":"

Federated Graph Learning (FGL) is a distributed machine learning paradigm based on graph neural networks, enabling secure and collaborative modeling of local graph data among clients. However, label noise in graph data can degrade the generalization performance of the global model. Existing federated label noise learning methods, primarily focused on computer vision tasks, often yield suboptimal results when directly applied to FGL. To address this issue, we propose a robust federated graph learning method with label noise, termed FedRGL. Specifically, FedRGL leverages the globally aggregated model and local subgraph structural information to implement a dual-perspective consistency noise-node filtering mechanism under class-aware dynamic thresholds. The resulting class-aware dual-consistency filtering (CADF) can also serve as a plug-and-play module, enhancing noise robustness across various subgraph federated learning frameworks. To better exploit the supervisory information from filtered noisy nodes, we employ the natural augmentation techniques from graph contrastive learning to assign high-confidence pseudo-labels to the noise nodes. Additionally, we measure model quality via the average predictive entropy of unlabeled nodes, enabling adaptive robust aggregation on the server side. Extensive experiments on real-world graph datasets show that FedRGL consistently outperforms existing methods under different noise rates, noise types, and client scales, achieving on average 5--8\\% higher accuracy and up to 30\\% improvement over the weakest baselines under noisy conditions. The anonymous source code is available at https://anonymous.4open.science/r/FedRGL_ICML26-376F.

","absKo":"

Federated Graph Learning(FGL)은 graph neural network를 기반으로 한 distributed machine learning paradigm으로, client 간 local graph data에 대한 안전하고 협력적인 modeling을 가능하게 한다. 그러나 graph data의 label noise는 global model의 generalization 성능을 저하시킬 수 있다. 주로 computer vision task에 초점을 맞춘 기존 federated label noise learning 방법은 FGL에 직접 적용하면 종종 최적 이하의 결과를 낸다. 이 문제를 해결하기 위해 우리는 label noise를 포함한 robust federated graph learning 방법인 FedRGL을 제안한다. 구체적으로, FedRGL은 globally aggregated model과 local subgraph structural information을 활용하여 class-aware dynamic threshold 하에서 dual-perspective consistency noise-node filtering mechanism을 구현한다. 그 결과인 class-aware dual-consistency filtering(CADF)는 plug-and-play module로도 사용할 수 있으며, 다양한 subgraph federated learning framework 전반에서 noise robustness를 향상시킨다. 필터링된 noisy node에서의 감독 정보를 더 잘 활용하기 위해, 우리는 graph contrastive learning의 natural augmentation 기법을 사용해 noise node에 high-confidence pseudo-label을 할당한다. 또한 우리는 unlabeled node의 average predictive entropy로 model quality를 측정하여, server 측에서 adaptive robust aggregation이 가능하도록 한다. 실제 graph dataset에 대한 광범위한 실험은 FedRGL이 서로 다른 noise rate, noise type, client scale에서 기존 방법을 일관되게 능가하며, 평균적으로 5--8\\% 더 높은 정확도를 달성하고 noisy condition에서 가장 약한 baseline 대비 최대 30\\% 개선을 보인다는 것을 보여준다. 익명 소스 코드는 https://anonymous.4open.science/r/FedRGL_ICML26-376F 에서 제공된다.

"},{"id":"60931","en":"Symbal: Detecting Systematic Misalignments in Model-Generated Captions","ko":"Symbal: 모델 생성 Caption의 체계적 Misalignment 탐지","authors":"Maya Varma, Jean-Benoit Delbrouck, Sophie Ostmeier, Akshay Chaudhari, Curtis Langlotz","pos":"#4102","link":"https://openreview.net/forum?id=vGnsSTcebc","abs":"

Multimodal large language models (MLLMs) often introduce errors when generating image captions, resulting in misaligned image-text pairs. Our work focuses on a class of captioning errors that we refer to as systematic misalignments, where a recurring error in MLLM-generated captions is closely associated with the presence of a specific visual feature in the paired image. Given a vision-language dataset with MLLM-generated captions, our aim in this work is to detect such errors, a task we refer to as systematic misalignment detection. As our first key contribution, we present Symbal, which utilizes a structured, dual-stage setup with off-the-shelf foundation models to identify systematic misalignments and summarize results in natural language. As our second key contribution, we introduce SymbalBench, a benchmark designed to evaluate automated methods on our proposed task. SymbalBench consists of 420 vision-language datasets from two domains (natural images and medical images) with annotated systematic misalignments. Symbal exhibits strong performance on this benchmark, correctly identifying systematic misalignments in 63.8% of datasets, a nearly 4x improvement over the closest baseline. We supplement our evaluations on SymbalBench with real-world evaluations, showing that (1) Symbal can accurately surface systematic misalignments in captions generated by four MLLMs and (2) Symbal is a powerful tool for auditing off-the-shelf image-caption datasets. Ultimately, our novel task, method, and benchmark can aid users with auditing MLLM-generated captions and identifying critical errors, without requiring access to the underlying MLLM.

","absKo":"

Multimodal large language models (MLLMs)는 image caption을 생성할 때 종종 오류를 유발하여 image-text pair의 정렬이 어긋나게 만든다. 우리의 연구는 systematic misalignments라고 부르는 captioning error의 한 부류에 초점을 맞춘다. 이는 MLLM이 생성한 caption의 반복적 오류가 짝지어진 이미지의 특정 visual feature 존재와 밀접하게 연관된 경우를 의미한다. MLLM-generated caption이 포함된 vision-language dataset이 주어졌을 때, 우리의 목표는 이러한 오류를 탐지하는 것이며, 우리는 이를 systematic misalignment detection이라 부른다. 첫 번째 핵심 기여로, 우리는 off-the-shelf foundation model을 활용한 구조화된 dual-stage setup인 Symbal을 제시하여 systematic misalignment를 식별하고 결과를 자연어로 요약한다. 두 번째 핵심 기여로, 우리가 제안한 task에 대해 automated method를 평가하도록 설계된 benchmark인 SymbalBench를 소개한다. SymbalBench는 annotated systematic misalignment가 포함된 두 domain(natural image 및 medical image)의 420개 vision-language dataset으로 구성된다. Symbal은 이 benchmark에서 강력한 성능을 보이며, dataset의 63.8%에서 systematic misalignment를 정확히 식별해 가장 가까운 baseline 대비 거의 4배 향상된 결과를 달성했다. 우리는 SymbalBench에서의 평가를 실제 환경 평가로 보완하여, (1) Symbal이 네 개의 MLLM이 생성한 caption에서 systematic misalignment를 정확히 드러낼 수 있고, (2) Symbal이 off-the-shelf image-caption dataset을 감사하는 데 강력한 도구임을 보인다. 궁극적으로, 우리의 새로운 task, method, benchmark는 underlying MLLM에 접근하지 않고도 사용자가 MLLM-generated caption을 감사하고 중요한 오류를 식별하는 데 도움이 될 수 있다.

"},{"id":"67152","en":"Position: Trustworthy AI Suffers from Invariance Conflicts and Causality is The Solution","ko":"Position: 신뢰할 수 있는 AI는 불변성 충돌로 고통받으며 인과성이 그 해법이다","authors":"Ruta Binkyte, Ivaxi Sheth, Zhijing Jin, Mohammad Havaei, Bernhard Schölkopf, Mario Fritz","pos":"#4206","link":"https://openreview.net/forum?id=RCsF4E84jE","abs":"

As artificial intelligence (AI), including machine learning (ML) models and foundation models (FMs), is increasingly deployed in high-stakes domains, ensuring their trustworthiness has become a central challenge. However, the core trustworthy AI objectives, such as fairness, robustness, privacy, and explainability, are hard to achieve simultaneously, especially while preserving utility. This position paper argues that causality is necessary to understand and balance trade-offs in performance and multiple objectives of trustworthy AI. We ground our arguments in re-interpreting trustworthy AI trade-offs as incompatible invariance requirements under different changes to the data-generating process. We then illustrate that causality provides a unifying framework for understanding how trade-offs in trustworthy AI arise, and how they can be softened or resolved through selective invariance. This perspective applies to both classical ML models and large-scale FMs. Our paper discusses how causal assumptions may be applied explicitly or implicitly in modern large-scale systems. Finally, we outline open challenges and opportunities for using causality to build more trustworthy AI.

","absKo":"

머신러닝(ML) 모델과 foundation model(FM)을 포함한 인공지능(AI)이 고위험 도메인에 점점 더 많이 배치됨에 따라, 그 신뢰성을 보장하는 일은 핵심 과제가 되었다. 그러나 공정성, 강건성, 프라이버시, 설명가능성과 같은 핵심 trustworthy AI 목표는, 특히 유용성을 유지하면서는, 동시에 달성하기가 어렵다. 이 포지션 페이퍼는 trustworthy AI의 성능 및 여러 목표 간 trade-off를 이해하고 균형 잡기 위해서는 인과성이 필요하다고 주장한다. 우리는 데이터 생성 과정의 서로 다른 변화 하에서 양립할 수 없는 불변성 요구사항으로 trustworthy AI trade-off를 재해석함으로써 우리의 주장을 정립한다. 이어서 인과성이 trustworthy AI에서 trade-off가 어떻게 발생하는지, 그리고 선택적 불변성을 통해 그것을 어떻게 완화하거나 해소할 수 있는지를 이해하는 통합적 프레임워크를 제공함을 보인다. 이러한 관점은 고전적 ML 모델과 대규모 FM 모두에 적용된다. 또한 이 논문은 현대의 대규모 시스템에서 인과적 가정이 명시적으로 혹은 암묵적으로 어떻게 적용될 수 있는지 논의한다. 마지막으로, 더 신뢰할 수 있는 AI를 구축하기 위해 인과성을 활용하는 데 있어 남아 있는 개방 과제와 기회를 제시한다.

"},{"id":"61816","en":"Quantifying the noise sensitivity of the Wasserstein metric for images","ko":"이미지에 대한 Wasserstein Metric의 노이즈 민감도 정량화","authors":"Erik Lager, Gilles Mordant, Amit Moscovich","pos":"#1614","link":"https://openreview.net/forum?id=mh2eIxPKMm","abs":"

Wasserstein metrics are increasingly adopted as similarity scores for images. We consider the sensitivity of Wasserstein metrics with respect to pixel-wise additive noise when the images are treated as discrete measures on the pixel grid. We derive finite-sample expectation bounds for a Gaussian noise model. Among other results, we prove that the error in the signed 2-Wasserstein distance scales with the square root of the noise standard deviation. This is favorable compared to the Euclidean metric that scales linearly, and thus provides a theoretical basis for the benefits of optimal transport distances in noisy settings. We present experiments that support our theoretical findings and point to a peculiar phenomenon where increasing the level of noise can decrease the Wasserstein distance. A case study on cryo-electron microscopy images demonstrates that the Wasserstein metric can capture the geometry of the data manifold in high noise settings even when the Euclidean metric fails.

","absKo":"

Wasserstein metric은 이미지의 similarity score로 점점 더 많이 채택되고 있다. 우리는 이미지가 pixel grid 위의 discrete measure로 취급될 때, pixel-wise additive noise에 대한 Wasserstein metric의 민감도를 고려한다. 우리는 Gaussian noise model에 대한 finite-sample expectation bound를 유도한다. 여러 결과 중에서도, signed 2-Wasserstein distance의 error가 noise standard deviation의 제곱근에 비례해 증가함을 증명한다. 이는 선형적으로 스케일하는 Euclidean metric에 비해 유리하며, 따라서 noisy setting에서 optimal transport distance의 장점에 대한 이론적 근거를 제공한다. 우리는 이론적 발견을 뒷받침하는 실험을 제시하고, noise 수준을 높이면 Wasserstein distance가 감소할 수 있는 특이한 현상을 지적한다. cryo-electron microscopy 이미지에 대한 case study는 Euclidean metric이 실패하는 고잡음 환경에서도 Wasserstein metric이 데이터 manifold의 geometry를 포착할 수 있음을 보여준다.

"},{"id":"66613","en":"Why Are Linear RNNs More Parallelizable?","ko":"선형 RNN은 왜 더 병렬화 가능한가?","authors":"William Merrill, Hongjian Jiang, Yanhong Li, Anthony Lin, Ashish Sabharwal","pos":"#1615","link":"https://openreview.net/forum?id=29sn1uqWn3","abs":"The community is increasingly exploring linear RNNs (LRNNs) as language models, motivated by their expressive power and parallelizability. While prior work establishes the expressivity benefits of LRNNs over transformers, it is unclear what makes LRNNs---but not traditional, *nonlinear* RNNs---as easy to parallelize in practice as transformers. We answer this question by providing a tight connection between types of RNNs and standard complexity classes. We show that LRNNs can be viewed as log-depth (bounded fan-in) arithmetic circuits, which represents only a slight depth overhead relative to log-depth boolean circuits that transformers admit. Furthermore, we show that nonlinear RNNs can solve $\\mathsf L$-complete problems (and even $\\mathsf P$-complete ones, under polynomial precision), revealing a fundamental barrier to parallelizing them as efficiently as transformers. Our theory also identifies fine-grained expressivity differences between recent popular LRNN variants: permutation-diagonal LRNNs are $\\mathsf{NC}^1$-complete whereas diagonal-plus-low-rank LRNNs are more expressive ($\\mathsf{PNC}^1$-complete). We provide further insight by associating each type of RNN with a corresponding automata-theoretic model that it can simulate. Together, our results reveal fundamental tradeoffs between nonlinear RNNs and different variants of LRNNs, providing a foundation for designing LLM architectures that achieve an optimal balance between expressivity and parallelism.","absKo":"커뮤니티는 표현력과 병렬화 가능성에 동기를 부여받아 linear RNN (LRNN)을 language model로 점점 더 탐구하고 있다. 선행 연구는 transformer에 비해 LRNN이 갖는 표현력의 이점을 입증했지만, LRNN이 전통적인 *nonlinear* RNN과 달리 왜 실제로 transformer처럼 쉽게 병렬화될 수 있는지는 분명하지 않다. 우리는 RNN의 유형과 표준 complexity class 사이의 긴밀한 연결을 제시함으로써 이 질문에 답한다. 우리는 LRNN을 log-depth (bounded fan-in) arithmetic circuit로 볼 수 있음을 보이며, 이는 transformer가 허용하는 log-depth boolean circuit에 비해 깊이 오버헤드가 약간만 더 크다. 또한 nonlinear RNN은 $\\mathsf L$-complete 문제를 해결할 수 있고, 다항 정밀도 하에서는 $\\mathsf P$-complete 문제까지도 해결할 수 있음을 보여 주어, transformer처럼 효율적으로 병렬화하는 데 근본적 장벽이 있음을 드러낸다. 우리의 이론은 최근 널리 쓰이는 LRNN 변형들 사이의 세분화된 표현력 차이도 밝혀낸다. permutation-diagonal LRNN은 $\\mathsf{NC}^1$-complete인 반면, diagonal-plus-low-rank LRNN은 더 높은 표현력을 가지며 ($\\mathsf{PNC}^1$-complete)다. 또한 각 유형의 RNN이 시뮬레이션할 수 있는 대응 automata-theoretic model과 연결함으로써 추가적인 통찰을 제공한다. 종합하면, 우리의 결과는 nonlinear RNN과 다양한 LRNN 변형 사이의 근본적 trade-off를 드러내며, 표현력과 병렬성 사이의 최적 균형을 달성하는 LLM architecture 설계를 위한 기반을 제공한다."},{"id":"63873","en":"The Sign Estimator: Preference Modeling for LLM Alignment under Heterogeneity","ko":"The Sign Estimator: 이질성 하에서의 LLM 정렬을 위한 선호도 모델링","authors":"Aymane El Gadarri, Ali Aouad, Vivek Farias","pos":"#1810","link":"https://openreview.net/forum?id=SyTeWTEex3","abs":"Traditional large language model (LLM) alignment methods are based on Reinforcement Learning From Human Feedback (RLHF), which learns a single reward model (implicitly or explicitly) from pairwise comparison data. This approach implicitly assumes homogeneous preferences across human labelers---an assumption that is violated in practice. As a result, the learned reward model in RLHF is generally misspecified: Prior work shows that it is inconsistent with the population-average utility, incurring large distortion, and that recovering this utilitarian objective is provably impossible in the worst case. In this work, we show that the average utility is recoverable under a mild assumption. Our method, the Sign Estimator, simply replaces the standard cross-entropy loss function with a notion of binary classification loss and yields a reward model that is ordinally consistent with the population-average utility. We further establish a fast finite-sample convergence rate of $O({n^{-{1}/{3}}})$ , which provides, to our knowledge, the first consistent estimator for heterogeneous preferences that does not suffer from the curse of dimensionality.","absKo":"전통적인 large language model(LLM) 정렬(alignment) 방법은 Reinforcement Learning From Human Feedback(RLHF)에 기반하며, 이는 pairwise comparison 데이터로부터 단일 reward model을 암묵적으로 또는 명시적으로 학습한다. 이 접근은 인간 labeler들 사이에 동질적인 선호가 존재한다고 암묵적으로 가정하지만, 실제로는 이 가정이 깨진다. 그 결과 RLHF에서 학습된 reward model은 일반적으로 misspecified되며, 기존 연구는 이것이 population-average utility와 일치하지 않아 큰 왜곡을 유발하고, 이 utilitarian objective를 복원하는 것이 최악의 경우 증명적으로 불가능함을 보여준다. 본 연구에서는 약한 가정하에서 average utility가 복원 가능함을 보인다. 우리의 방법인 Sign Estimator는 표준 cross-entropy loss function을 binary classification loss의 개념으로 단순히 대체하며, population-average utility와 순서적으로 일치하는(ordinarily consistent) reward model을 산출한다. 우리는 또한 $O({n^{-{1}/{3}}})$의 빠른 유한표본 수렴률을 확립하는데, 이는 우리가 아는 한 이질적 선호에 대해 curse of dimensionality를 겪지 않는 첫 번째 consistent estimator를 제공한다."},{"id":"63399","en":"Optimal Domain-Aware Privacy Mechanisms for Synthetic Data Generation","ko":"합성 데이터 생성을 위한 최적 도메인 인지 프라이버시 메커니즘","authors":"Sajani Vithana, Sangwon Jung, Haoyang Hu, Viveck Cadambe, Flavio Calmon, Haewon Jeong","pos":"#3110","link":"https://openreview.net/forum?id=XTIPGuQPMo","abs":"

Differential privacy (DP) imposes fundamental trade-offs between privacy and statistical fidelity in synthetic data generation. While access to public data has been shown to improve these trade-offs empirically, existing approaches exploit public data only indirectly, through pre-processing (e.g., using pre-trained generative models) or post-processing steps (e.g., matching target statistics estimated from public datasets), while relying on domain-agnostic DP mechanisms. In this work, we lay the theoretical framework to study the principled incorporation of public data into DP mechanisms themselves. We consider normalized histograms as distribution estimators and characterize the \\emph{theoretically optimal} domain-aware privacy mechanism within a class of mixing-based DP mechanisms. We introduce \\textsc{PubMix}, a public-data-aware DP mechanism that can be used in histogram-based data synthesis pipelines. Our experiments demonstrate that, when public data is available, \\textsc{PubMix} significantly improves synthetic data generation quality across tasks without compromising privacy.

","absKo":"

Differential privacy(DP)는 synthetic data generation에서 privacy와 statistical fidelity 사이에 근본적인 trade-off를 부과한다. public data에 대한 접근이 이러한 trade-off를 경험적으로 개선한다는 점은 알려져 있지만, 기존 접근법은 public data를 pre-processing(예: pre-trained generative model 사용)이나 post-processing 단계(예: public dataset에서 추정한 target statistic과의 matching)를 통해서만 간접적으로 활용하며, domain-agnostic한 DP mechanism에 의존한다. 본 연구에서는 public data를 DP mechanism 자체에 원칙적으로 통합하는 문제를 연구하기 위한 이론적 틀을 마련한다. 우리는 normalized histogram을 distribution estimator로 간주하고, mixing-based DP mechanism의 한 클래스 내에서 \\emph{이론적으로 최적}인 domain-aware privacy mechanism을 특성화한다. 우리는 histogram-based data synthesis pipeline에서 사용할 수 있는 public-data-aware DP mechanism인 \\textsc{PubMix}를 소개한다. 실험 결과, public data를 사용할 수 있을 때 \\textsc{PubMix}는 privacy를 훼손하지 않으면서 여러 task 전반에서 synthetic data generation quality를 크게 향상시킨다.

"},{"id":"62064","en":"A Bi-metric Framework for Efficient Nearest Neighbor Search","ko":"효율적 Nearest Neighbor Search를 위한 Bi-metric Framework","authors":"Haike Xu, Sandeep Silwal, Piotr Indyk","pos":"#3915","link":"https://openreview.net/forum?id=kBzzertldC","abs":"

We propose a new ``bi-metric'' framework for designing nearest neighbor data structures. Our framework assumes two dissimilarity functions: a ground-truth metric that is accurate but expensive to compute, and a proxy metric that is cheaper but less accurate. In both theory and practice, we show how to construct data structures using only the proxy metric such that the query procedure achieves the accuracy of the expensive metric, while only using a limited number of calls to both metrics. Our theoretical results instantiate this framework for two popular nearest neighbor search algorithms: DiskANN and Cover Tree. In both cases we show that, as long as the proxy metric used to construct the data structure approximates the ground-truth metric up to a bounded factor, our data structure achieves arbitrarily good approximation guarantees with respect to the ground-truth metric. On the empirical side, we apply the framework to the text retrieval problem with two dissimilarity functions evaluated by ML models with vastly different computational costs. We observe that for almost all the large data sets in the BEIR benchmark, our approach achieves a considerably better accuracy-efficiency tradeoff than the alternatives, such as retrieve-then-rerank.

","absKo":"

최근접 이웃 데이터 구조를 설계하기 위한 새로운 ``bi-metric'' 프레임워크를 제안한다. 우리의 프레임워크는 두 개의 dissimilarity function을 가정한다: 하나는 정확하지만 계산 비용이 큰 ground-truth metric이고, 다른 하나는 더 저렴하지만 덜 정확한 proxy metric이다. 이론과 실험 모두에서, proxy metric만을 사용해 데이터 구조를 구성하되 query procedure는 expensive metric의 정확도를 달성하면서도 두 metric 모두에 대한 호출 수는 제한된 방식으로 동작함을 보인다. 우리의 이론적 결과는 이 프레임워크를 두 가지 널리 쓰이는 nearest neighbor search algorithm인 DiskANN과 Cover Tree에 적용한다. 두 경우 모두, 데이터 구조를 구성하는 데 사용된 proxy metric이 ground-truth metric을 bounded factor 이내로 근사하기만 하면, 우리의 데이터 구조는 ground-truth metric에 대해 arbitrarily good approximation guarantee를 달성함을 보인다. 실증적으로는, 계산 비용이 크게 다른 ML model로 평가되는 두 dissimilarity function을 사용하는 text retrieval 문제에 이 프레임워크를 적용한다. BEIR benchmark의 거의 모든 large data set에서, 우리의 접근법이 retrieve-then-rerank와 같은 대안들보다 훨씬 더 나은 정확도-효율성 절충점을 달성함을 관찰한다.

"},{"id":"61003","en":"Even Faster Kernel Matrix Linear Algebra via Density Estimation","ko":"밀도 추정을 통한 더 빠른 커널 행렬 선형대수","authors":"Rikhav Shah, Sandeep Silwal, Haike Xu","pos":"#4005","link":"https://openreview.net/forum?id=ueNIrBXz7R","abs":"This paper studies the use of *kernel density estimation* (KDE) for linear algebraic tasks involving the *kernel matrix* of a collection of $n$ data points in $\\mathbb{R}^d$. In particular, we improve upon the best existing algorithms for computing the following up to $(1+\\varepsilon)$ relative error for a Gaussian kernel matrix and other kernels: matrix-vector products, matrix-matrix products, the spectral norm, and sum of all entries. The runtimes of our algorithms depend linearly on the dimension $d$, sub-quadratically in the number of points $n$, and polynomially on the target error $\\varepsilon$. Importantly, the dependence on $n$ in each case is far lower when accessing the kernel matrix through KDE queries as opposed to reading individual entries. Our improvements over existing best algorithms (particularly those of [Backurs et al. ICML `21]) for these tasks reduce the polynomial dependence on $\\varepsilon$, and additionally decrease the dependence on $n$ in the case of computing the sum of all entries of the kernel matrix. For example, we reduce the power of $1/\\epsilon$ from $\\approx 7.7$ to $\\approx 3.2$ for a $1-\\varepsilon$ relative error estimation of the spectral norm of a Gaussian kernel matrix. We complement our upper bounds with several lower bounds for related problems, which provide (conditional) quadratic time hardness results and additionally hint at the limits of KDE based approaches for the problems we study.","absKo":"본 논문은 $\\mathbb{R}^d$에 있는 $n$개의 data point 집합의 *kernel matrix*를 다루는 linear algebraic task에 대해 *kernel density estimation*(KDE)을 활용하는 방법을 연구한다. 특히 Gaussian kernel matrix 및 기타 kernel에 대해 다음 연산들을 $(1+\\varepsilon)$ relative error까지 계산하는 데 있어 기존 최고 알고리즘을 개선한다: matrix-vector product, matrix-matrix product, spectral norm, 그리고 모든 entry의 합. 우리의 algorithm runtime은 차원 $d$에 대해서는 선형, point 수 $n$에 대해서는 sub-quadratic, 목표 오차 $\\varepsilon$에 대해서는 polynomial하게 의존한다. 중요한 점은, kernel matrix를 개별 entry를 읽는 대신 KDE query를 통해 접근할 때 각 경우의 $n$에 대한 의존성이 훨씬 작다는 것이다. 이러한 task에 대해 기존 최고 알고리즘, 특히 [Backurs et al. ICML `21]의 결과보다 개선한 점은 $\\varepsilon$에 대한 polynomial dependence를 줄인 것이며, 추가로 kernel matrix의 모든 entry 합을 계산할 때의 $n$ 의존성도 감소시켰다. 예를 들어, Gaussian kernel matrix의 spectral norm을 $1-\\varepsilon$ relative error로 추정할 때 $1/\\epsilon$의 차수를 약 7.7에서 약 3.2로 낮춘다. 또한 관련 문제들에 대한 여러 lower bound를 통해 상한을 보완하는데, 이는 조건부로 quadratic time hardness 결과를 제공하며, 우리가 연구하는 문제들에 대해 KDE 기반 접근법의 한계도 시사한다."},{"id":"60498","en":"Optimal Estimation of Continuous Treatment Effects with Kernel Ridge Regression","ko":"Kernel Ridge Regression을 이용한 Continuous Treatment Effect의 최적 추정","authors":"Seok-Jin Kim, Kaizheng Wang","pos":"#4006","link":"https://openreview.net/forum?id=ziqS4yXFQX","abs":"

We consider the problem of estimating the effect function of a continuous treatment, which maps each treatment value to its average effect over the population. A central challenge in this setting is confounding: treatment assignment often depends on covariates, creating selection bias that makes direct regression of the response on the treatment unreliable. To address this issue, we propose a two-stage kernel ridge regression method. In the first stage, we learn a model for the response as a function of both treatment and covariates; in the second stage, we use this model to construct pseudo-outcomes that correct for the distribution shift and then fit a second model to estimate the treatment effect. Although the response varies with both treatment and covariates, the induced effect function obtained by averaging over covariates is typically much simpler, and our estimator adapts to this structure. Furthermore, we introduce a fully data-driven model selection procedure that achieves provable adaptivity to both the unknown degree of overlap and the regularity (eigenvalue decay) of the underlying kernel.

","absKo":"

우리는 continuous treatment의 effect function을 추정하는 문제를 고려한다. 이 함수는 각 treatment value를 인구 집단에 대한 평균 효과로 매핑한다. 이 설정의 핵심 도전은 confounding이다. treatment 할당은 종종 covariate에 의존하므로 selection bias가 발생하고, treatment에 대한 response의 직접 회귀를 신뢰하기 어렵게 만든다. 이를 해결하기 위해 우리는 two-stage kernel ridge regression 방법을 제안한다. 첫 번째 단계에서는 treatment와 covariate의 함수로 response를 모델링하고, 두 번째 단계에서는 이 모델을 사용해 distribution shift를 보정하는 pseudo-outcome을 구성한 뒤 treatment effect를 추정하기 위한 두 번째 모델을 적합한다. response는 treatment와 covariate 모두에 따라 변하지만, covariate에 대해 평균화하여 얻는 induced effect function은 일반적으로 훨씬 더 단순하며, 우리의 estimator는 이러한 구조에 적응한다. 또한 우리는 알려지지 않은 overlap의 정도와 기반 kernel의 regularity (eigenvalue decay) 모두에 대해 provable adaptivity를 달성하는 fully data-driven model selection 절차를 도입한다.

"},{"id":"61795","en":"Networked Information Aggregation for Binary Classification","ko":"Binary Classification을 위한 Networked Information Aggregation","authors":"MohammadHossein Bateni, Zahra Hadizadeh, MohammadTaghi Hajiaghayi, Mahdi JafariRaviz, Shayan Taherijam","pos":"#4007","link":"https://openreview.net/forum?id=mrtg4NmvAe","abs":"We study networked binary classification on a directed acyclic graph (DAG) where each agent observes only a subset of the feature columns of a shared finite dataset. Agents act sequentially along the DAG: each receives prediction columns from its parents (if any), augments its local features with these columns, fits a logistic predictor by minimizing binary cross-entropy (BCE), and forwards its prediction column to its outgoing neighbors. We ask whether this sequential distributed training procedure achieves *information aggregation*, meaning that some agent attains small excess loss compared to the best logistic predictor trained with access to all feature columns. This question was studied for linear regression under squared loss by Kearns, Roth, and Ryu (2026). Extending their guarantees to classification is nontrivial because their analysis relies on quadratic structure that does not directly transfer to BCE with a logistic link. We analyze the resulting sequential logit-passing protocol and prove: (i) an excess loss upper bound of $O(M/\\sqrt{D})$ on depth-$D$ paths under the condition that every $M$ contiguous subsequence of $M$ agents collectively observe all features, and (ii) a close lower bound showing instances with excess loss of at least $\\Omega(k/D)$ where $k$ is the dimension of the feature space. Together, these results identify network depth as a fundamental bottleneck for information aggregation in networked logistic regression.","absKo":"우리는 각 agent가 공유된 유한 dataset의 feature column 일부만 관찰하는 directed acyclic graph (DAG) 위의 networked binary classification을 연구한다. agent는 DAG를 따라 순차적으로 동작한다. 각 agent는 부모로부터 prediction column을 받는다(있는 경우), 이 column을 자신의 local feature에 추가하고, binary cross-entropy (BCE)를 최소화하여 logistic predictor를 적합한 뒤, 그 prediction column을 outgoing neighbor에게 전달한다. 우리는 이 순차적 distributed training 절차가 *information aggregation*을 달성하는지, 즉 어떤 agent가 모든 feature column에 접근해 학습한 최적의 logistic predictor에 비해 작은 excess loss를 달성하는지 묻는다. 이 질문은 Kearns, Roth, 그리고 Ryu (2026)가 squared loss 하의 linear regression에 대해 연구한 바 있다. 그들의 보장을 classification으로 확장하는 것은 쉽지 않은데, 그들의 분석이 BCE와 logistic link로 직접 옮겨지지 않는 quadratic structure에 의존하기 때문이다. 우리는 이로부터 형성되는 sequential logit-passing protocol을 분석하고 다음을 증명한다: (i) 모든 M개의 연속된 subsequence가 공동으로 모든 feature를 관찰하는 조건하에서 depth-D path에 대해 excess loss 상한 O(M/√D), (ii) feature space의 차원이 k일 때 excess loss가 Ω(k/D) 이상인 instance를 보이는 근접한 lower bound. 이 결과들은 network depth가 networked logistic regression에서 information aggregation의 근본적인 병목임을 밝힌다."},{"id":"66060","en":"Temper-Then-Tilt: Principled Unlearning for Generative Models through Tempering and Classifier Guidance","ko":"Temper-Then-Tilt: 템퍼링과 분류기 가이던스를 통한 생성 모델의 원칙적 언러닝","authors":"Jacob L. Block, Mehryar Mohri, Aryan Mokhtari, Sanjay Shakkottai","pos":"#4405","link":"https://openreview.net/forum?id=7Oc4ARFCTV","abs":"

We study machine unlearning in large generative models by framing the task as density ratio estimation to a target distribution rather than supervised fine-tuning. While classifier guidance is a standard approach for approximating this ratio and can succeed in general, we show it can fail to faithfully unlearn with finite samples when the forget set represents a sharp, concentrated data distribution. To address this, we introduce Temper-Then-Tilt Unlearning (T3-Unlearning), which freezes the base model and applies a two-step inference procedure: (i) tempering the base distribution to flatten high-confidence spikes, and (ii) tilting the tempered distribution using a lightweight classifier trained to distinguish retain from forget samples. Our theoretical analysis provides finite-sample guarantees linking the surrogate classifier's risk to unlearning quality, proving that tempering is necessary to successfully unlearn for concentrated distributions. Empirical evaluations on the TOFU benchmark demonstrate that T3-Unlearning improves forget quality and generative utility over existing baselines, while training only a fraction of the parameters with a minimal runtime.

","absKo":"우리는 대규모 generative model에서 machine unlearning을 supervised fine-tuning이 아니라 target distribution에 대한 density ratio estimation으로 정식화하여 연구한다. classifier guidance는 이 비율을 근사하는 표준 접근법이며 일반적으로는 성공할 수 있지만, forget set이 날카롭고 집중된 data distribution을 나타낼 때 finite sample에서는 충실하게 unlearn하지 못할 수 있음을 보인다. 이를 해결하기 위해, 우리는 Temper-Then-Tilt Unlearning(T3-Unlearning)을 제안한다. 이는 base model을 고정한 채 두 단계 inference procedure를 적용한다: (i) 높은 confidence spike를 평탄화하기 위해 base distribution을 tempering하고, (ii) retain sample과 forget sample을 구분하도록 학습된 lightweight classifier를 사용해 tempered distribution을 tilting한다. 우리의 이론적 분석은 surrogate classifier의 risk와 unlearning quality를 연결하는 finite-sample guarantee를 제공하며, 집중된 distribution에 대해 성공적인 unlearning을 위해 tempering이 필요함을 증명한다. TOFU benchmark에서의 실험은 T3-Unlearning이 기존 baseline 대비 forget quality와 generative utility를 향상시키면서, 최소한의 runtime으로 전체 parameter의 일부만 학습함을 보여준다.

"},{"id":"61008","en":"Provable Accuracy Collapse of Embedding-Based Representations under Dimensionality Mismatch","ko":"차원 불일치에서 임베딩 기반 표현의 입증 가능한 정확도 붕괴","authors":"Dionysis Arvanitakis, Vaggos Chatziafratis, Yiyuan Luo","pos":"#4502","link":"https://openreview.net/forum?id=ubzXPEfa4R","abs":"Embedding-based representations in Euclidean space $\\mathbb{R}^d$ are a cornerstone of modern machine learning, where a major goal is to use the \\emph{smallest dimension} that faithfully captures data relations. In this work, we prove sharp dimension--accuracy tradeoffs and identify a fundamental information-theoretic limitation: unless the embedding dimension $d$ is chosen close to the ground-truth dimension $D$, accuracy undergoes a sudden collapse. Our main result shows that this phenomenon arises even in standard contrastive learning settings, where supervision is limited to a set of $m$ anchor--positive--negative triplets $(i,j,k)$ encoding distance comparisons $\\mathrm{dist}(i,j) < \\mathrm{dist}(i,k)$. Specifically, given triplets realizable by an unknown ground-truth embedding in $D$ dimensions, we prove that there exists constant $c < 1$, such that \\emph{every embedding of dimension at most $cD$ violates half of the triplets}, yielding accuracy as low as a trivial one-dimensional solution that ignores the input. We complement our information-theoretic bounds with strong computational hardness results: under the Unique Games Conjecture, even if the given triplets are nearly realizable in $D=1$ dimension, no polynomial-time algorithm---\\textit{regardless of its dimension}---can achieve accuracy above the trivial 50% baseline.","absKo":"Euclidean space $\\mathbb{R}^d$의 embedding-based representations는 현대 machine learning의 핵심이며, 중요한 목표는 데이터 관계를 충실하게 포착하는 \\emph{가장 작은 차원}을 사용하는 것이다. 본 연구에서는 날카로운 dimension--accuracy tradeoff를 증명하고, 근본적인 정보이론적 한계를 규명한다. embedding dimension $d$가 ground-truth dimension $D$에 가깝게 선택되지 않으면 accuracy는 갑작스럽게 붕괴한다. 우리의 주요 결과는 이 현상이 표준 contrastive learning setting에서도 발생함을 보인다. 이때 supervision은 거리 비교 $\\mathrm{dist}(i,j) < \\mathrm{dist}(i,k)$를 인코딩하는 $m$개의 anchor--positive--negative triplet $(i,j,k)$ 집합으로 제한된다. 구체적으로, 알 수 없는 ground-truth embedding이 $D$차원에서 실현 가능한 triplet이 주어졌을 때, 상수 $c < 1$이 존재하여 \\emph{$cD$ 이하 차원의 모든 embedding은 triplet의 절반을 위반}함을 증명한다. 그 결과 accuracy는 입력을 무시하는 자명한 1차원 해와 같을 정도로 낮아질 수 있다. 우리는 이러한 정보이론적 bound를 강한 계산 복잡도 결과로 보완한다. Unique Games Conjecture 하에서, 주어진 triplet이 $D=1$ 차원에서 거의 실현 가능하더라도, 어떤 polynomial-time algorithm---\\textit{차원이 무엇이든 상관없이}---도 자명한 50% baseline을 넘는 accuracy를 달성할 수 없다."},{"id":"63500","en":"Content-Style Identification via Differential Independence","ko":"차등 독립성을 통한 내용-스타일 식별","authors":"Subash Timilsina, Hoang-Son Nguyen, Sagar Shrestha, Xiao Fu","pos":"#4610","link":"https://openreview.net/forum?id=WcsBsGO6SW","abs":"

Generative analysis often models multi-domain observations as nonlinear mixtures of domain-invariant content variables and domain-specific style variables. Identifying both factors from unpaired domains enables tasks such as domain transfer and counterfactual data generation. Prior work establishes identifiability under (block-wise) statistical independence between content and style, or via sparse Jacobian assumptions on the nonlinear mixing function, but such conditions can be restrictive and may not hold in practice. In this work, we introduce differential independence, a weaker structural condition requiring that infinitesimal variations in content and style induce orthogonal directions on the data manifold, thereby enabling identifiability even when content and style are dependent and the Jacobian is dense. We operationalize this condition through a blockwise orthogonality constraint on the Jacobian subspaces associated with content and style. To support high-dimensional generative models, we design a stochastic regularizer based on numerical Jacobian approximation, enabling scalable training in settings such as high-resolution image generation. Experiments across multiple datasets corroborate the identifiability analysis and demonstrate practical benefits on counterfactual generation and domain translation tasks.

","absKo":"

생성적 분석은 종종 multi-domain 관측을 domain-invariant content 변수와 domain-specific style 변수의 비선형 혼합으로 모델링합니다. 비짝지어진(unpaired) domains에서 두 요인을 모두 식별하면 domain transfer와 counterfactual data generation 같은 작업이 가능해집니다. 기존 연구는 content와 style 사이의 (block-wise) statistical independence, 또는 비선형 mixing function에 대한 sparse Jacobian 가정을 통해 identifiability를 보장하지만, 이러한 조건은 제한적일 수 있고 실제로 성립하지 않을 수도 있습니다. 본 연구에서는 differential independence를 도입하는데, 이는 content와 style의 무한소 변화가 data manifold 위에서 서로 직교하는 방향을 유도해야 한다는 더 약한 구조적 조건으로, content와 style이 의존적이고 Jacobian이 dense한 경우에도 identifiability를 가능하게 합니다. 우리는 이 조건을 content와 style에 대응하는 Jacobian subspace들에 대한 blockwise orthogonality constraint로 구체화합니다. 고차원 생성 모델을 지원하기 위해, 수치적 Jacobian 근사에 기반한 stochastic regularizer를 설계하여 고해상도 이미지 생성과 같은 환경에서 확장 가능한 학습을 가능하게 합니다. 여러 데이터셋에 걸친 실험은 identifiability 분석을 뒷받침하고, counterfactual generation 및 domain translation 작업에서 실질적 이점을 보여줍니다.

"},{"id":"63396","en":"Domain Adaptation with Adaptive $f$-Divergence: Tighter Variational Representation and Generalization Bounds","ko":"적응형 $f$-발산을 활용한 도메인 적응: 더 조밀한 변분 표현과 일반화 경계","authors":"Zhe Cheng, Fode Zhang, Lingrui Wang, Yifan Zhu, Jiaolong Wang","pos":"#4611","link":"https://openreview.net/forum?id=XX4oQ3pjZR","abs":"We study unsupervised domain adaptation (UDA) where measuring cross-domain discrepancy is critical. Most UDA approaches fix a single $f$-divergence a priori, which can be suboptimal across heterogeneous shifts. We propose a framework that (i) tightens the variational lower bound of an $f$-divergence by inserting a learnable, monotone $L$-Lipschitz transform $\\tau$ (Tighter-VR), and (ii) selects the divergence family adaptively from data via a likelihood-based criterion. The resulting estimator yields more informative and statistically efficient discrepancy estimates while recovering prior fixed-divergence methods as special cases. Theoretically, we derive a target-risk bound whose three components are a transformed source risk, a Tighter-VR discrepancy between domains, and an ideal-hypothesis residual; we further provide finite-sample guarantees using standard complexity measures. Empirically, on Office-31, Office-Home, Digits, and VisDA-2017, our method consistently improves accuracy over strong baselines, demonstrating that coupling Tighter-VR with adaptive divergence selection yields tangible gains in UDA.","absKo":"우리는 cross-domain discrepancy를 측정하는 일이 중요한 unsupervised domain adaptation(UDA)을 연구한다. 대부분의 UDA 접근법은 단일한 $f$-divergence를 사전에 고정하는데, 이는 이질적인 shift 전반에서 최적이 아닐 수 있다. 우리는 (i) 학습 가능한 monotone한 $L$-Lipschitz transform $\\tau$를 삽입하여 $f$-divergence의 variational lower bound를 더 타이트하게 만드는 프레임워크(Tighter-VR)와, (ii) likelihood-based criterion을 통해 data로부터 divergence family를 적응적으로 선택하는 방법을 제안한다. 이로부터 얻는 estimator는 더 정보성이 높고 통계적으로 더 효율적인 discrepancy estimate를 제공하며, 기존의 고정 divergence 방법들을 특수한 경우로 회복한다. 이론적으로는 세 부분으로 구성된 target-risk bound를 유도하는데, 그 구성요소는 transformed source risk, domain 간 Tighter-VR discrepancy, 그리고 ideal-hypothesis residual이다. 또한 표준 complexity measure를 사용해 finite-sample guarantee를 제시한다. 실험적으로는 Office-31, Office-Home, Digits, VisDA-2017에서 우리의 방법이 강력한 baseline보다 일관되게 정확도를 향상시키며, Tighter-VR와 adaptive divergence selection을 결합하는 것이 UDA에서 실질적인 이득을 가져옴을 보여준다."},{"id":"66361","en":"Domain Adaptive Object Detection via Dynamic Causal Refinement","ko":"동적 인과 정제를 통한 도메인 적응형 객체 탐지","authors":"Zeyu Ma, Jiaqi Huang, Yitong Qin, Ziqiang Zheng, Jiwei Wei, Jie Zou, Yang Yang, Heng Tao Shen","pos":"#4612","link":"https://openreview.net/forum?id=4MGh9AsGrq","abs":"

Domain Adaptive Object Detection (DAOD) addresses the challenge of transferring object detectors from labeled source domains to unlabeled target domains. Existing domain adaptation methods primarily rely on feature distribution alignment, which enhances domain-invariant features (statistical invariance) but also inadvertently increases inherent domain-common spurious factors (e.g., common environmental contexts), which act as shortcut features rather than the true causal factors for object classification. We propose Dynamic Causal Refinement (DCR), a novel framework that establishes a closed-loop feedback mechanism between data augmentation and model optimization to progressively refine causal features. Specifically, we design Semantic Prediction Consistency (SPC) to filter domain-specific spurious factors and establish a robust statistical invariance, and Discrepancy-Guided Causal Refinement (DGCR) to actively suppress the dependence on domain-common spurious factors via spectral perturbation for causal refinement. This process encourages the detector to suppress its reliance on shortcut features and instead prioritize semantically meaningful causal representations. Extensive experiments on standard benchmarks demonstrate that our method outperforms state-of-the-art counterparts significantly.

","absKo":"

Domain Adaptive Object Detection (DAOD)는 라벨이 있는 source domain의 object detector를 라벨이 없는 target domain으로 전이하는 문제를 다룹니다. 기존 domain adaptation 방법은 주로 feature distribution alignment에 의존하는데, 이는 domain-invariant feature(통계적 불변성)를 강화하는 동시에 우연적 domain-common spurious factor(예: 공통 환경 맥락)까지 의도치 않게 증가시켜, 객체 분류의 진짜 causal factor가 아니라 shortcut feature로 작동하게 만듭니다. 우리는 Dynamic Causal Refinement (DCR)를 제안하는데, 이는 data augmentation과 model optimization 사이에 closed-loop feedback mechanism을 구축하여 causal feature를 점진적으로 정교화하는 새로운 framework입니다. 구체적으로, Semantic Prediction Consistency (SPC)를 설계하여 domain-specific spurious factor를 걸러내고 강건한 statistical invariance를 확립하며, Discrepancy-Guided Causal Refinement (DGCR)을 통해 spectral perturbation을 이용해 domain-common spurious factor에 대한 의존을 적극적으로 억제함으로써 causal refinement를 수행합니다. 이 과정은 detector가 shortcut feature에 대한 의존을 줄이고, 대신 의미적으로 중요한 causal representation을 우선하도록 유도합니다. 표준 benchmark에서 수행한 광범위한 실험은 우리의 방법이 state-of-the-art 대비 유의미하게 우수함을 보여줍니다.

"},{"id":"63691","en":"Domain-Shift-Aware Conformal Prediction for Large Language Models","ko":"도메인 이동을 인지하는 Large Language Models용 Conformal Prediction","authors":"Zhexiao Lin, Yuanyuan Li, Neeraj Sarna, Yuanyuan Gao, Michael Berger","pos":"#4613","link":"https://openreview.net/forum?id=UbzajT84Nd","abs":"

Large language models have achieved impressive performance across diverse tasks. However, their tendency to produce overconfident and factually incorrect outputs, known as hallucinations, poses risks in real world applications. Conformal prediction provides finite-sample, distribution-free coverage guarantees, but standard conformal prediction breaks down under domain shift, often leading to under-coverage and unreliable prediction sets. We propose a new framework called Domain-Shift-Aware Conformal Prediction (DS-CP). Our framework adapts conformal prediction to large language models under domain shift, by systematically reweighting calibration samples based on their proximity to the test prompt, thereby preserving validity while enhancing adaptivity. Our theoretical analysis and experiments on the MMLU benchmark demonstrate that the proposed method delivers more reliable coverage than standard conformal prediction, especially under substantial distribution shifts, while maintaining efficiency. This provides a practical step toward trustworthy uncertainty quantification for large language models in real-world deployment.

","absKo":"

Large language models는 다양한 task에서 인상적인 성능을 달성했다. 그러나 hallucination으로 알려진, 과도하게 자신감 있으면서도 사실과 다른 출력을 생성하는 경향은 실제 응용에서 위험을 초래한다. Conformal prediction은 finite-sample, distribution-free coverage guarantee를 제공하지만, 표준 conformal prediction은 domain shift 하에서 무너지며, 종종 under-coverage와 신뢰할 수 없는 prediction set으로 이어진다. 우리는 Domain-Shift-Aware Conformal Prediction(DS-CP)이라는 새로운 프레임워크를 제안한다. 우리의 프레임워크는 calibration sample을 test prompt와의 근접도에 따라 체계적으로 재가중함으로써 domain shift 하의 large language model에 conformal prediction을 적응시키며, validity를 보존하면서 adaptivity를 향상한다. 이론 분석과 MMLU benchmark에서의 실험은 제안한 방법이 특히 큰 distribution shift 하에서 표준 conformal prediction보다 더 신뢰할 수 있는 coverage를 제공하면서도 효율성을 유지함을 보여준다. 이는 실제 배포 환경에서 large language model을 위한 신뢰할 수 있는 uncertainty quantification으로 가는 실용적 한 걸음이 된다.

"},{"id":"64078","en":"Incorporating Importance Weighting in Optimal Transport Based Domain Alignment","ko":"최적 수송 기반 도메인 정렬에 중요도 가중치 통합","authors":"Okan Koç, Alexander Soen, Shanglin Li, Masashi Sugiyama","pos":"#4614","link":"https://openreview.net/forum?id=R8HbSt2J4e","abs":"Domain adaptation theory studies upper bounds on the target risk in order to mitigate performance loss of machine learning models due to distribution shift. In this paper, we take a closer look at the optimization of one such bound based on optimal transport (OT) and propose various strategies that improve the optimization in practice. We first introduce *gradual shift* and *probabilistic margin* assumptions to control the incomputable entanglement term that appears in the bounds. We prove that under these assumptions, better optimization of the computable part of the bound can translate to better target accuracies. Motivated by this fact, we tighten the bound, via importance weighting of the source (output) distribution, to obtain the *weighted* Wasserstein regularized risk ($\\mathrm{W}^2\\mathrm{R}^{2}$), that is often easier to minimize than the original bound. $\\mathrm{W}^2\\mathrm{R}^{2}$ is shown to be equivalent to an unbalanced OT problem, which in the limit converges to a nearest neighbor based alignment strategy. We highlight the tradeoffs faced with such an approach and show that a suitably regularized $\\mathrm{W}^2\\mathrm{R}^{2}$ improves over the state of the art and is robust to multiple distribution shifts under different models, confirming, moreover, the validity of our assumptions.","absKo":"Domain adaptation theory는 distribution shift로 인한 machine learning model의 성능 손실을 완화하기 위해 target risk의 upper bound를 연구한다. 본 논문에서는 optimal transport (OT)에 기반한 이러한 bound 중 하나의 최적화를 면밀히 살펴보고, 실제 최적화를 개선하는 다양한 전략을 제안한다. 먼저 bound에 등장하는 계산 불가능한 entanglement term을 제어하기 위해 *gradual shift*와 *probabilistic margin* 가정을 도입한다. 이러한 가정 하에서, bound의 계산 가능한 부분을 더 잘 최적화하면 target accuracy 향상으로 이어질 수 있음을 증명한다. 이 사실에 동기부여를 받아, source (output) distribution에 importance weighting을 적용하여 bound를 더욱 타이트하게 만들고, 원래의 bound보다 더 쉽게 최소화할 수 있는 *weighted* Wasserstein regularized risk ($\\mathrm{W}^2\\mathrm{R}^{2}$)를 얻는다. $\\mathrm{W}^2\\mathrm{R}^{2}$는 unbalanced OT 문제와 동등하며, 극한에서는 nearest neighbor 기반 alignment strategy로 수렴함을 보인다. 우리는 이러한 접근이 가지는 trade-off를 강조하고, 적절히 regularized된 $\\mathrm{W}^2\\mathrm{R}^{2}$가 state of the art를 능가하며 다양한 model과 여러 distribution shift에 대해 견고함을 보인다는 점, 나아가 우리의 가정이 타당함을 확인한다.

"},{"id":"62455","en":"Scaling Unsupervised Multi-Source Federated Domain Adaptation through Group-Wise Discrepancy Minimization","ko":"그룹 단위 Discrepancy Minimization으로 확장 가능한 다중 소스 Unsupervised Federated Domain Adaptation","authors":"Larissa Reichart, Cem Ata Baykara, Ali Burak Ünal, Harlin Lee, Mete Akgün","pos":"#4615","link":"https://openreview.net/forum?id=gSQHzzOwyf","abs":"

Unsupervised multi-source domain adaptation (UMDA) leverages labeled data from multiple source domains to generalize to an unlabeled target. While federated UMDA addresses privacy by avoiding raw data sharing, existing methods scale poorly as the number of sources increases, often suffering from high computational overhead or training instability. We propose GALA, a scalable and robust federated UMDA framework designed for high-diversity settings. GALA achieves scalability by coupling a novel inter-group discrepancy minimization objective that approximates pairwise alignment with linear complexity alongside a temperature-controlled, centroid-based weighting strategy for dynamic source prioritization. These components enable stable, parallelizable training across many heterogeneous sources, addressing a critical scalability bottleneck that remains largely unaddressed in current literature. To evaluate performance in high-diversity scenarios, we introduce Digit-18, a new benchmark comprising 18 datasets with varied synthetic and real-world domain shifts. Extensive experiments demonstrate that GALA achieves state-of-the-art results on standard benchmarks and significantly outperforms prior methods in large-scale settings where others either fail to converge or become computationally infeasible.

","absKo":"

Unsupervised multi-source domain adaptation(UMDA)은 여러 source domain의 labeled data를 활용해 unlabeled target으로 일반화하는 방법이다. federated UMDA는 raw data 공유를 피함으로써 privacy를 해결하지만, 기존 방법은 source 수가 증가할수록 확장성이 떨어지며, 종종 높은 계산 오버헤드나 training instability에 시달린다. 우리는 높은 다양성(high-diversity) 환경을 위해 설계된 확장 가능하고 견고한 federated UMDA framework인 GALA를 제안한다. GALA는 pairwise alignment를 선형 복잡도로 근사하는 새로운 inter-group discrepancy minimization objective와, 동적 source 우선순위화를 위한 temperature-controlled centroid-based weighting strategy를 결합해 scalability를 달성한다. 이러한 구성 요소는 많은 heterogeneous source 전반에서 안정적이고 병렬화 가능한 학습을 가능하게 하여, 현재 문헌에서 크게 해결되지 않은 중요한 scalability 병목을 해소한다. high-diversity 시나리오에서 성능을 평가하기 위해, 우리는 다양한 synthetic 및 real-world domain shift를 포함한 18개 dataset으로 구성된 새로운 benchmark인 Digit-18을 도입한다. 대규모 실험은 GALA가 standard benchmark에서 state-of-the-art 결과를 달성하고, 다른 방법들이 수렴에 실패하거나 계산적으로 불가능해지는 대규모 설정에서 이전 방법들을 크게 능가함을 보인다.

"},{"id":"63557","en":"U$^3$CF: Unbiased, Unconfounding, and Unified Causal Framework for Multi-Target Domain Adaptation","ko":"U$^3$CF: 다중 대상 도메인 적응을 위한 편향 제거, 교란 제거, 통합 인과 프레임워크","authors":"Wenxu Wang, Yeqiang Liu, Rui Zhou, Jing Wang, Zhenbo Li, Wenbo Gong","pos":"#4616","link":"https://openreview.net/forum?id=W31tIsoc4M","abs":"Multi-target domain adaptation (MTDA) trains a model using a labeled source domain and several unlabeled target domains, aiming to enhance performance across all targets. However, existing methods lack a principled causal formulation and often rely on empirical domain-invariance enforcement, which can bias adaptation across targets. To fill this gap, we propose the **U**nbiased, **U**nconfounding, and **U**nified **C**ausal **F**ramework (**U$^3$CF**) for MTDA. To *unify* align multiple domains, we propose a prototype-driven alignment strategy that progressively updates prototypes by high-confidence target predictions, while the contrastive optimization objective jointly aligns target samples to semantic prototypes and preserves class discrimination. By formulating a structural causal model, we reveal that domain-invariant causal factors and domain-specific factors shape representations and labels, while the latter induces spurious label correlations across targets. Accordingly, U$^3$CF achieves *unbiased* prediction by disentangling representations into invariant causal components and domain-specific confounders and applying conditional intervention to *block confounding* effects while preserving invariant semantics. To ensure precise disentanglement, we leverage mutual information theory to derive a principled criterion for feature separation. Extensive experiments on four benchmarks demonstrate that U$^3$CF consistently outperforms leading methods.","absKo":"Multi-target domain adaptation (MTDA)는 레이블이 있는 source domain과 여러 개의 레이블이 없는 target domain을 사용해 model을 학습하며, 모든 target에서의 성능 향상을 목표로 한다. 그러나 기존 방법들은 원리적인 causal formulation이 부족하고, 종종 경험적 domain-invariance 강제에 의존하여 target 간 adaptation에 편향을 유발할 수 있다. 이러한 공백을 메우기 위해, 우리는 MTDA를 위한 **U**nbiased, **U**nconfounding, and **U**nified **C**ausal **F**ramework (**U$^3$CF**)을 제안한다. 여러 domain을 *unify*하기 위해, 우리는 높은 신뢰도의 target prediction으로 prototype을 점진적으로 갱신하는 prototype-driven alignment strategy를 제안하며, contrastive optimization objective는 target sample을 semantic prototype에 공동으로 정렬하는 동시에 class discrimination을 보존한다. structural causal model을 정식화함으로써, domain-invariant causal factor와 domain-specific factor가 representation과 label을 형성하며, 후자가 target 간 spurious label correlation을 유발함을 밝힌다. 이에 따라 U$^3$CF는 representation을 invariant causal component와 domain-specific confounder로 disentangle하고 conditional intervention을 적용하여 invariant semantics를 보존하면서 confounding effect를 *block confounding*함으로써 *unbiased* prediction을 달성한다. 정확한 disentanglement을 보장하기 위해 mutual information theory를 활용하여 feature separation에 대한 원리적 criterion을 도출한다. 네 개의 benchmark에서 수행한 광범위한 실험은 U$^3$CF가 일관되게 선도적 방법들을 능가함을 보여준다."},{"id":"61028","en":"Improving Backward Conformal Prediction via Non-Conformity Score Transformation","ko":"Non-Conformity Score 변환을 통한 Backward Conformal Prediction 개선","authors":"Junxian Liu, Hao Zeng, Hongxin Wei","pos":"#4617","link":"https://openreview.net/forum?id=uPpo1RsI2d","abs":"

Conformal Prediction (CP) provides a statistical framework for uncertainty quantification that constructs prediction sets with coverage guarantees. While CP yields uncontrolled prediction set sizes, Backward Conformal Prediction (BCP) inverts this paradigm by enforcing a predefined upper bound on set size and estimating the resulting coverage guarantee. However, the looseness induced by Markov's inequality within the BCP framework causes a significant gap between the estimated coverage bound and the empirical coverage. In this work, we introduce ST-BCP, a novel method that introduces a data-dependent transformation of nonconformity scores to narrow the coverage gap. In particular, we develop a computable transformation and prove that it outperforms the baseline identity transformation. Extensive experiments demonstrate the effectiveness of our method, reducing the average coverage gap from 4.20\\% to 1.12\\% on common benchmarks.

","absKo":"

Conformal Prediction (CP)는 coverage 보장을 갖는 prediction set을 구성함으로써 불확실성 정량화를 위한 통계적 프레임워크를 제공한다. CP는 제어되지 않은 prediction set 크기를 산출하는 반면, Backward Conformal Prediction (BCP)는 이 패러다임을 뒤집어 사전에 정의된 set size의 상한을 강제하고 그에 따른 coverage 보장을 추정한다. 그러나 BCP 프레임워크에서 Markov inequality로 인해 유도되는 느슨함은 추정된 coverage bound와 경험적 coverage 사이에 상당한 간극을 초래한다. 본 연구에서는 이 coverage gap을 좁히기 위해 nonconformity score에 데이터 의존적 변환을 도입하는 새로운 방법 ST-BCP를 제안한다. 특히, 우리는 계산 가능한 변환을 개발하고 그것이 baseline identity transformation보다 우수함을 증명한다. 광범위한 실험은 본 방법의 효과를 입증하며, 일반적인 benchmark에서 평균 coverage gap을 4.20\\%에서 1.12\\%로 줄인다.

"},{"id":"64252","en":"On the existence of consistent adversarial attacks in high-dimensional linear classification","ko":"고차원 선형 분류에서 일관된 적대적 공격의 존재성에 관하여","authors":"Matteo Vilucchio, Lenka Zdeborova, Bruno Loureiro","pos":"#1604","link":"https://openreview.net/forum?id=PUIivg3GrO","abs":"

What fundamentally distinguishes an adversarial attack from a misclassification due to limited model expressivity or finite data? In this work, we investigate this question in the setting of high-dimensional binary classification, where statistical effects due to limited data availability play a central role. We introduce a new error metric that precisely capture this distinction, quantifying model vulnerability to consistent adversarial attacks --- perturbations that preserve the ground-truth labels. Our main technical contribution is an exact and rigorous asymptotic characterization of these metrics in both well-specified models and latent space models, revealing different vulnerability patterns compared to standard robust error measures. The theoretical results demonstrate that as models become more overparameterized, their vulnerability to label-preserving perturbations grows, offering theoretical insight into the mechanisms underlying model sensitivity to adversarial attacks.

","absKo":"제한된 model expressivity나 유한한 데이터로 인해 발생한 misclassification과 adversarial attack을 근본적으로 구분하는 것은 무엇인가? 본 연구에서는 제한된 데이터 가용성으로 인한 statistical effect가 중심적인 역할을 하는 고차원 binary classification 설정에서 이 질문을 탐구한다. 우리는 이 구분을 정확히 포착하는 새로운 error metric을 도입하여, ground-truth label을 보존하는 perturbation인 consistent adversarial attack에 대한 모델 취약성을 정량화한다. 우리의 주요 기술적 기여는 well-specified model과 latent space model 모두에서 이러한 metric의 정확하고 엄밀한 asymptotic characterization이며, 표준 robust error measure와는 다른 취약성 패턴을 드러낸다. 이 이론 결과는 모델이 더 과매개변수화될수록 label-preserving perturbation에 대한 취약성이 증가함을 보여주며, adversarial attack에 대한 모델 민감성의 메커니즘에 대한 이론적 통찰을 제공한다.

"},{"id":"60782","en":"A Solvable High-Dimensional Model Where Nonlinear Autoencoders Learn Structure Invisible to PCA While Test Loss Misaligns With Generalization","ko":"Nonlinear Autoencoder가 PCA에 보이지 않는 구조를 학습하지만 Test Loss가 Generalization과 어긋나는 풀 수 있는 고차원 모델","authors":"Vicente Mendes, Lorenzo Bardone, Cédric Koller, Jorge Medina Moreira, Vittorio Erba, Emanuele Troiani, Lenka Zdeborova","pos":"#1608","link":"https://openreview.net/forum?id=wm3ABfhE7P","abs":"

Many real-world datasets contain hidden structure that cannot be detected by simple linear correlations between input features. For example, latent factors may influence the data in a coordinated way, even though their effect is invisible to covariance-based methods such as PCA. In practice, nonlinear neural networks often succeed in extracting such hidden structure in unsupervised and self-supervised learning. However, constructing a minimal high-dimensional model where this advantage can be rigorously analyzed has remained an open theoretical challenge. We introduce a tractable high-dimensional spiked model with two latent factors: one visible to covariance, and one statistically dependent yet uncorrelated, appearing only in higher-order moments. PCA and linear autoencoders fail to recover the latter, while a minimal nonlinear autoencoder provably extracts both. We analyze both the population risk, and empirical risk minimization. Our model also provides a tractable example where self-supervised test loss is poorly aligned with representation quality: nonlinear autoencoders recover latent structure that linear methods miss, even though their reconstruction loss is higher.

","absKo":"

많은 실제 데이터셋에는 입력 feature 사이의 단순한 선형 상관으로는 탐지할 수 없는 숨겨진 구조가 존재한다. 예를 들어, latent factor는 covariance 기반 방법인 PCA로는 보이지 않더라도 데이터를 조정된 방식으로 함께 변화시킬 수 있다. 실제로 비선형 neural network는 비지도 및 self-supervised learning에서 이러한 숨겨진 구조를 추출하는 데 종종 성공한다. 그러나 이러한 장점을 엄밀하게 분석할 수 있는 최소한의 고차원 모델을 구성하는 문제는 여전히 열린 이론적 과제로 남아 있다. 우리는 covariance로 보이는 하나, 통계적으로는 종속적이지만 비상관이며 higher-order moment에서만 나타나는 하나의 두 latent factor를 가진 tractable high-dimensional spiked model을 도입한다. PCA와 linear autoencoder는 후자를 복원하지 못하는 반면, 최소한의 nonlinear autoencoder는 두 요소를 모두 추출함을 증명할 수 있다. 우리는 population risk와 empirical risk minimization을 모두 분석한다. 또한 우리의 모델은 self-supervised test loss가 representation quality와 잘 정렬되지 않는 tractable한 예시를 제공한다. 즉, nonlinear autoencoder는 reconstruction loss가 더 높더라도 linear method가 놓치는 latent structure를 복원한다.

"},{"id":"61872","en":"Dataset Distillation Efficiently Encodes Low-Dimensional Representations from Gradient-Based Learning of Non-Linear Tasks","ko":"Dataset Distillation은 비선형 과제의 Gradient-Based Learning에서 저차원 표현을 효율적으로 인코딩한다","authors":"Yuri Kinoshita, Naoki Nishikawa, Taro Toyoizumi","pos":"#1610","link":"https://openreview.net/forum?id=mCsBf1sm4p","abs":"Dataset distillation, a training-aware data compression technique, has recently attracted increasing attention as an effective tool for mitigating costs of optimization and data storage. However, progress remains largely empirical. Mechanisms underlying the extraction of task-relevant information from the training process and the efficient encoding of such information into synthetic data points remain elusive. In this paper, we theoretically analyze practical algorithms of dataset distillation applied to the gradient-based training of two-layer neural networks with width $L$. By focusing on a non-linear task structure called multi-index model, we prove that the low-dimensional structure of the problem is efficiently encoded into the resulting distilled data. This dataset reproduces a model with high generalization ability for a required memory complexity of $\\Theta(r^2d+L)$, where $d$ and $r$ are the input and intrinsic dimensions of the task. To the best of our knowledge, this is one of the first theoretical works that include a specific task structure, leverage its intrinsic dimensionality to quantify the compression rate and study dataset distillation implemented solely via gradient-based algorithms.","absKo":"Dataset distillation은 optimization 비용과 data storage 비용을 완화하는 효과적인 도구로서, 최근 training-aware data compression 기법으로 큰 주목을 받고 있습니다. 그러나 진전은 여전히 주로 경험적 수준에 머물러 있습니다. 학습 과정에서 task-relevant information을 추출하고 이를 synthetic data point에 효율적으로 인코딩하는 메커니즘은 아직 명확히 밝혀지지 않았습니다. 본 논문에서는 폭이 $L$인 두 층 neural network의 gradient-based training에 적용되는 실용적 dataset distillation 알고리즘을 이론적으로 분석합니다. multi-index model이라고 불리는 비선형 task structure에 초점을 맞추어, 문제의 저차원 구조가 결과로 얻어진 distilled data에 효율적으로 인코딩됨을 증명합니다. 이 dataset은 입력 차원과 intrinsic dimension이 각각 $d$와 $r$인 task에 대해 $\\Theta(r^2d+L)$의 memory complexity로 높은 generalization 능력을 가진 model을 재현합니다. 우리가 아는 한, 이는 특정 task structure를 포함하고 그 intrinsic dimensionality를 활용해 compression rate를 정량화하며, gradient-based algorithm만으로 구현된 dataset distillation을 연구한 최초의 이론적 연구 중 하나입니다."},{"id":"61910","en":"Sharp description of local minima in the loss landscape of high-dimensional two-layer ReLU neural networks","ko":"고차원 2층 ReLU 신경망 손실 풍경의 국소 최솟값에 대한 정확한 설명","authors":"Jie Huang, Bruno Loureiro, Stefano Sarao Mannelli","pos":"#306","link":"https://openreview.net/forum?id=lk5T4FMxlm","abs":"We study the population loss landscape of two-layer ReLU networks of the form $\\sum_{k=1}^K \\mathrm{ReLU}(w_k^\\top x)$ in a realisable teacher–student setting with Gaussian covariates. We show that local minima admit an exact low-dimensional representation in terms of \\emph{summary statistics}, yielding a sharp and interpretable characterisation of the landscape. We further establish a direct link with one-pass SGD: local minima correspond to attractive fixed points of the dynamics in summary statistics space. This perspective reveals a hierarchical structure of minima: they are typically isolated in the well-specified regime, but become connected by flat directions as network width increases. In this overparameterised regime, global minima become increasingly accessible, attracting the dynamics and reducing convergence to spurious solutions. Overall, our results reveal intrinsic limitations of common simplifying assumptions, which may miss essential features of the loss landscape even in minimal neural network models.","absKo":"가우시안 공변량을 갖는 실현 가능한 teacher-student 설정에서, 우리는 $\\sum_{k=1}^K \\mathrm{ReLU}(w_k^\\top x)$ 형태의 2층 ReLU network의 population loss landscape를 연구한다. 우리는 local minimum이 \\emph{summary statistics}로 정확히 표현될 수 있음을 보이며, 이를 통해 landscape에 대한 날카롭고 해석 가능한 특성화를 제공한다. 또한 one-pass SGD와의 직접적인 연결을 확립한다: local minimum은 summary statistics 공간에서 dynamics의 attractive fixed point에 대응한다. 이 관점은 minima의 계층적 구조를 드러낸다. 즉, 일반적으로 well-specified regime에서는 고립되어 있지만, network width가 증가함에 따라 flat direction에 의해 서로 연결된다. 이러한 overparameterized regime에서는 global minimum에 점점 더 쉽게 접근할 수 있게 되어 dynamics를 끌어당기고 spurious solution으로의 수렴을 줄인다. 전반적으로, 우리의 결과는 일반적인 단순화 가정이 내재적 한계를 지님을 보여주며, 이러한 가정은 최소한의 neural network model에서도 loss landscape의 핵심 특징을 놓칠 수 있다.

"},{"id":"61170","en":"PCA of Probability Measures: Sparse and Dense Sampling Regimes","ko":"확률측도의 PCA: 희소 및 조밀 샘플링 체제","authors":"Erell Gachon, Jérémie Bigot, Elsa Cazelles","pos":"#3907","link":"https://openreview.net/forum?id=t4SltdwuLF","abs":"A common approach to perform PCA on probability measures is to embed them into a Hilbert space where standard functional PCA techniques apply. While convergence rates for estimating the embedding of a single measure from $m$ samples are well understood, the literature has not addressed the setting involving multiple measures. In this paper, we study PCA in a double asymptotic regime where $n$ probability measures are observed, each through $m$ samples. We derive convergence rates of the form $n^{-1/2} + m^{-\\alpha}$ for the empirical covariance operator and the PCA excess risk, where $\\alpha>0$ depends on the chosen embedding. This characterizes the relationship between the number $n$ of measures and the number $m$ of samples per measure, revealing a sparse (small $m$) to dense (large $m$) transition in the convergence behavior. Moreover, we prove that the dense-regime rate is minimax optimal for the empirical covariance error. Our numerical experiments validate these theoretical rates and demonstrate that appropriate subsampling preserves PCA accuracy while reducing computational cost.","absKo":"주어진 probability measure에 PCA를 수행하는 일반적인 방법은, standard functional PCA 기법을 적용할 수 있는 Hilbert space에 이를 embedding하는 것이다. 하나의 measure를 m개의 sample로부터 추정할 때의 convergence rate는 잘 알려져 있지만, 여러 measure를 다루는 설정은 문헌에서 충분히 다뤄지지 않았다. 본 논문에서는 n개의 probability measure가 관측되며, 각 measure가 m개의 sample을 통해 표현되는 double asymptotic regime에서 PCA를 연구한다. 우리는 empirical covariance operator와 PCA excess risk에 대해 n^{-1/2} + m^{-α} 형태의 convergence rate를 유도하며, 여기서 α>0는 선택한 embedding에 의존한다. 이는 measure의 개수 n과 measure당 sample 수 m 사이의 관계를 정량화하며, convergence behavior에서 sparse(small m)에서 dense(large m)로의 전이를 드러낸다. 더 나아가, dense-regime의 rate가 empirical covariance error에 대해 minimax optimal임을 증명한다. 수치 실험은 이러한 이론적 rate를 검증하고, 적절한 subsampling이 계산 비용을 줄이면서도 PCA 정확도를 유지함을 보여준다."},{"id":"63721","en":"Characterization of Gaussian Universality Breakdown in High-Dimensional Empirical Risk Minimization","ko":"고차원 Empirical Risk Minimization에서 Gaussian 보편성 붕괴의 특성화","authors":"Mohamed Chiheb Yaakoubi, Cosme Louart, Malik TIOMOKO, Zhenyu Liao","pos":"#3911","link":"https://openreview.net/forum?id=UHQDfvZBFi","abs":"We study high-dimensional convex empirical risk minimization (ERM) under general non-Gaussian data designs. By heuristically extending the Convex Gaussian Min–Max Theorem (CGMT) to non-Gaussian settings, we derive an asymptotic min–max characterization of key statistics, enabling approximation of the mean $\\mu_{\\hat{\\theta}}$ and covariance $C_{\\hat{\\theta}}$ of the ERM estimator $\\hat{\\theta}$. Specifically, under a concentration assumption on the data matrix and standard regularity conditions on the loss and regularizer, we show that for a test covariate $x$ independent of the training data, the projection $\\hat{\\theta}^\\top x$ approximately follows the convolution of the (generally non-Gaussian) distribution of $\\mu_{\\hat{\\theta}}^\\top x$ with an independent centered Gaussian variable of variance $\\mathrm{tr}\\!\\big(C_{\\hat{\\theta}}\\,\\mathbb{E}[xx^\\top]\\big)$. This result clarifies the scope and limits of Gaussian universality for ERMs. Additionally, we prove that any $\\mathcal{C}^2$ regularizer is asymptotically equivalent to a quadratic form determined solely by its Hessian at zero and gradient at $\\mu_{\\hat{\\theta}}$. Numerical simulations across diverse losses and models are provided to validate our theoretical predictions and qualitative insights.","absKo":"우리는 일반적인 비가우시안 데이터 설계 하에서 고차원 convex empirical risk minimization(ERM)을 연구한다. Convex Gaussian Min-Max Theorem(CGMT)을 비가우시안 설정으로 휴리스틱하게 확장함으로써, 핵심 통계량의 asymptotic min-max characterization을 도출하고, 이를 통해 ERM estimator $\\hat{\\theta}$의 mean $\\mu_{\\hat{\\theta}}$와 covariance $C_{\\hat{\\theta}}$를 근사할 수 있게 한다. 구체적으로, 데이터 행렬에 대한 concentration 가정과 loss 및 regularizer에 대한 표준 regularity condition 하에서, 훈련 데이터와 독립인 test covariate $x$에 대해 projection $\\hat{\\theta}^\\top x$는 대략적으로 $\\mu_{\\hat{\\theta}}^\\top x$의 (일반적으로 비가우시안인) 분포와 분산이 $\\mathrm{tr}\\!\\big(C_{\\hat{\\theta}}\\,\\mathbb{E}[xx^\\top]\\big)$인 독립적인 평균 0 Gaussian 변수의 convolution을 따른다는 것을 보인다. 이 결과는 ERM에서 Gaussian universality의 적용 범위와 한계를 명확히 한다. 또한 우리는 임의의 $\\mathcal{C}^2$ regularizer가, zero에서의 Hessian과 $\\mu_{\\hat{\\theta}}$에서의 gradient로만 결정되는 quadratic form과 점근적으로 동등함을 증명한다. 다양한 loss와 model에 대한 수치 시뮬레이션을 통해 우리의 이론적 예측과 정성적 통찰을 검증한다."},{"id":"62770","en":"Representation Learning for Equivariant Inference with Guarantees","ko":"보장 가능한 Equivariant Inference를 위한 Representation Learning","authors":"Daniel Felipe Ordonez Apraez, Vladimir Kostic, Alek Fröhlich, Vivien Brandt, Karim Lounici, Massimiliano Pontil","pos":"#4401","link":"https://openreview.net/forum?id=dWCG53oQWc","abs":"

In many real-world applications of regression, conditional probability estimation, and uncertainty quantification, exploiting symmetries rooted in physics or geometry can dramatically improve generalization and sample efficiency. While geometric deep learning has made empirical advances by incorporating symmetry and geometry priors, less attention has been given to statistical learning guarantees. In this paper, we introduce an equivariant representation learning framework that simultaneously addresses regression, conditional probability estimation, and uncertainty quantification while providing first-of-its-kind non-asymptotic statistical learning guarantees. Grounded in operator and group representation theory, our framework approximates the spectral decomposition of the conditional expectation operator, building representations that are both equivariant and disentangled along independent symmetry quotient groups. Empirical evaluations on synthetic datasets and real-world robotics applications confirm the potential of our approach, matching or outperforming existing equivariant baselines in regression while providing well-calibrated uncertainty estimates.

","absKo":"

회귀, conditional probability estimation, uncertainty quantification의 많은 실제 응용에서 물리나 기하에 뿌리를 둔 symmetry를 활용하면 generalization과 sample efficiency를 크게 향상시킬 수 있다. geometric deep learning은 symmetry와 geometry prior를 통합함으로써 실증적 진전을 이루어 왔지만, statistical learning guarantee에 대해서는 상대적으로 덜 주목해 왔다. 이 논문에서는 회귀, conditional probability estimation, uncertainty quantification을 동시에 다루면서 동시에 최초의 non-asymptotic statistical learning guarantee를 제공하는 equivariant representation learning framework를 제안한다. operator와 group representation theory에 기반한 이 framework는 conditional expectation operator의 spectral decomposition을 근사하여, 서로 독립인 symmetry quotient group을 따라 equivariant하면서도 disentangled된 representation을 구축한다. synthetic dataset과 실제 robotics application에서의 실증 평가 결과는 우리의 접근이 회귀에서 기존 equivariant baseline과 동등하거나 더 우수한 성능을 보이면서, 잘 보정된 uncertainty estimate를 제공함을 확인해 준다.

"},{"id":"63855","en":"Sparse Regression with $\\ell_0$ Constraints for $\\alpha$-Mixing Time Series: Algorithms and Guarantees","ko":"$\\alpha$-mixing 시계열에서 $\\ell_0$ 제약을 갖는 희소 회귀: 알고리즘과 보장","authors":"Ruoxin Yuan, Lijun Ding","pos":"#4403","link":"https://openreview.net/forum?id=T9JKg3Ey5g","abs":"Exact sparse methods based on $\\ell_0$ constraints are increasingly used for interpretable and scalable time series modeling, where one aims to recover a small set of informative lags/factors while maintaining strong predictive performance and low computational cost. Despite their empirical success, finite-sample and computational guarantees for such methods under temporal dependence remain limited. In this paper, we study $\\ell_0$-constrained least squares for time series generated by $\\alpha$-mixing stationary Gaussian processes with sparse coefficients. We establish high-probability restricted strong convexity/smoothness (RSC/RSS) for the empirical quadratic loss. Leveraging these conditions, we derive nonasymptotic statistical guarantees and computational complexities for a series of exact sparse methods, including iterative hard thresholding (IHT). We apply our theoretical results to Gaussian vector autoregressive (VAR) models and obtain new guarantees. Experiments on synthetic sparse VAR models and real-world mobility time series demonstrate that exact sparse methods recover lag structure more accurately and interpretably than some classical methods, while achieving comparable prediction error with substantially lower computational cost.","absKo":"

$\\ell_0$ constraint 기반의 exact sparse method는 해석 가능하고 확장 가능한 time series modeling에서 점점 더 많이 사용되고 있으며, 이때 목표는 강한 예측 성능과 낮은 계산 비용을 유지하면서 정보성 높은 lag/factor의 작은 집합을 복원하는 것이다. 그러나 이러한 방법이 temporal dependence 하에서 finite-sample 및 computational guarantee를 갖는지에 대해서는 아직 잘 알려져 있지 않다. 본 논문에서는 sparse coefficients를 갖는 $\\alpha$-mixing stationary Gaussian process가 생성한 time series에 대해 $\\ell_0$-constrained least squares를 연구한다. 우리는 empirical quadratic loss에 대해 높은 확률의 restricted strong convexity/smoothness(RSC/RSS)를 증명한다. 이러한 조건을 활용하여 iterative hard thresholding(IHT)을 포함한 일련의 exact sparse method에 대한 nonasymptotic statistical guarantee와 computational complexity를 도출한다. 또한 이론적 결과를 Gaussian vector autoregressive(VAR) model에 적용하여 새로운 보장을 얻는다. synthetic sparse VAR model과 실제 mobility time series에서의 실험은 exact sparse method가 일부 고전적 방법보다 lag structure를 더 정확하고 해석 가능하게 복원하면서도, 상당히 낮은 계산 비용으로 비슷한 prediction error를 달성함을 보여준다.

"},{"id":"64352","en":"The Implicit Bias of Steepest Descent with Mini-batch Stochastic Gradient","ko":"미니배치 확률적 그래디언트를 갖는 최급강하법의 암묵적 편향","authors":"Jichu Li, Xuan Tang, Difan Zou","pos":"#4406","link":"https://openreview.net/forum?id=OT9cxeWbEO","abs":"A variety of widely used optimization methods like SignSGD and Muon can be interpreted as instances of steepest descent under different norm-induced geometries. In this work, we study the implicit bias of mini-batch stochastic steepest descent in multi-class classification, characterizing how batch size, momentum, and variance reduction shape the limiting max-margin behavior and convergence rates under general entry-wise and Schatten-$p$ norms. We show that without momentum, convergence only occurs with large batches, yielding a batch-dependent margin gap but the full-batch convergence rate. In contrast, momentum enables small-batch convergence through a batch–momentum trade-off, though it slows convergence. This approach provides fully explicit, dimension-free rates that improve upon prior results. Moreover, we prove that variance reduction can recover the exact full-batch implicit bias for any batch size, albeit at a slower convergence rate. Finally, we further investigate the batch-size-one steepest descent without momentum, and reveal its convergence to a fundamentally different bias via a concrete data example, which reveals a key limitation of purely stochastic updates. Overall, our unified analysis clarifies when stochastic optimization aligns with full‑batch behavior, and paves the way for perform deeper explorations of the training behavior of stochastic gradient steepest descent algorithms.","absKo":"SignSGD와 Muon 같은 널리 사용되는 다양한 optimization method는 서로 다른 norm-induced geometry 하의 steepest descent의 인스턴스로 해석될 수 있습니다. 본 연구에서는 multi-class classification에서 mini-batch stochastic steepest descent의 implicit bias를 연구하며, 일반적인 entry-wise norm과 Schatten-$p$ norm 하에서 batch size, momentum, variance reduction이 limiting max-margin behavior와 convergence rate를 어떻게 형성하는지 특성화합니다. 우리는 momentum이 없을 때는 큰 batch에서만 수렴이 일어나며, batch-dependent margin gap이 발생하지만 full-batch convergence rate는 유지됨을 보입니다. 반대로 momentum은 batch–momentum trade-off를 통해 작은 batch의 수렴을 가능하게 하지만, 수렴 속도는 느려집니다. 이 접근법은 이전 결과를 개선하는, dimension-free인 완전히 명시적인 rate를 제공합니다. 또한 variance reduction이 어떤 batch size에서도 정확한 full-batch implicit bias를 복원할 수 있음을 증명하지만, 그 대가로 더 느린 convergence rate를 갖습니다. 마지막으로, momentum이 없는 batch-size-one steepest descent를 추가로 분석하고, 구체적인 데이터 예시를 통해 근본적으로 다른 bias로의 수렴을 밝힘으로써 순수한 stochastic update의 핵심 한계를 드러냅니다. 전반적으로 우리의 통합 분석은 stochastic optimization이 언제 full-batch behavior와 정렬되는지를 명확히 하고, stochastic gradient steepest descent algorithm의 training behavior를 더 깊이 탐구할 수 있는 길을 열어줍니다."},{"id":"61406","en":"The Interplay Between Interpolation and Aggregation in Regression: Optimal Sample Complexity","ko":"Regression에서 Interpolation과 Aggregation의 상호작용: 최적 Sample Complexity","authors":"Mikael Moller Hogsgaard, Kasper Green Larsen, Liang-Yu Zou","pos":"#4407","link":"https://openreview.net/forum?id=qXlovWytwg","abs":"This work investigates theoretically the interplay between interpolation and aggregation in regression. We establish that the $\\gamma$-graph dimension characterizes learnability for a broad class of natural aggregation procedures. Furthermore, we prove that an extremely simple aggregation procedure, combining three interpolating hypotheses via the median, is optimal among all these aggregation procedures, and is strictly more powerful than proper learning. Finally, we show that some hypothesis classes are learnable only by aggregating infinitely many hypotheses or by using non-interpolating aggregation rules (which may predict outside the range of their inputs), and any finite interpolating aggregation fails to achieve even trivial performance.","absKo":"이 연구는 regression에서 interpolation과 aggregation의 상호작용을 이론적으로 탐구한다. 우리는 $\\gamma$-graph dimension이 광범위한 자연스러운 aggregation procedure에 대한 learnability를 특징짓는다는 것을 보인다. 더 나아가, 세 개의 interpolating hypothesis를 median으로 결합하는 매우 단순한 aggregation procedure가 이러한 모든 aggregation procedure 중에서 최적이며, proper learning보다도 엄밀하게 더 강력하다는 것을 증명한다. 마지막으로, 어떤 hypothesis class는 무한히 많은 hypothesis를 aggregating해야만 또는 non-interpolating aggregation rule(입력의 범위를 벗어난 값을 예측할 수 있음)을 사용해야만 학습 가능하며, 어떤 유한한 interpolating aggregation도 자명한 성능조차 달성하지 못함을 보인다."},{"id":"66403","en":"The data manifold under the microscope","ko":"현미경으로 본 데이터 다양체","authors":"Marios Koulakis, Constantin Seibold","pos":"#4408","link":"https://openreview.net/forum?id=42cOdUJOhH","abs":"A significant gap exists between theory and practice in deep learning. Generalization and approximation error bounds are often derived for simplified models or are too loose to be informative. Many rely on the manifold hypothesis and on geometric regularity such as intrinsic dimension, curvature, and reach. Progress requires insight into data-manifold geometry and suitable benchmarks, yet existing options are polarized: analytic manifolds with known geometry but limited applicability, or real-world datasets where geometry is only coarsely estimable. We introduce a benchmarking framework for studying data geometry by repurposing and extending dSprites and COIL-20 with additional transformation dimensions and denser sampling, enabling accurate finite-difference estimates of curvature, reach, and volume that are otherwise difficult to estimate reliably and implement in practice. As applications, we assess bounds by Genovese et al. and Fefferman et al., and analyze how geometry evolves across network layers in $\\beta$-VAEs, highlighting the behavior of current bounds and the value of controlled benchmarks for guiding and validating future theory. Code to reproduce the framework and experiments is included with the submission and will be released as open-source library upon publication.","absKo":"딥러닝에서 이론과 실제 사이에는 여전히 큰 간극이 존재한다. Generalization error와 approximation error bound는 종종 단순화된 model에 대해 유도되거나, 너무 느슨해서 유용한 정보를 주지 못한다. 많은 방법이 manifold hypothesis와 intrinsic dimension, curvature, reach와 같은 geometric regularity에 의존한다. 진전을 이루기 위해서는 data-manifold geometry에 대한 통찰과 적절한 benchmark가 필요하지만, 기존 선택지는 양극화되어 있다. 즉, geometry는 알려져 있지만 적용 범위가 제한된 analytic manifold이거나, geometry를 대략적으로만 추정할 수 있는 real-world dataset이다. 우리는 추가 transformation dimension과 더 촘촘한 sampling으로 dSprites와 COIL-20을 재활용 및 확장하여 data geometry를 연구하기 위한 benchmarking framework를 제안한다. 이를 통해 실제로는 신뢰성 있게 추정하거나 구현하기 어려운 curvature, reach, volume에 대해 accurate finite-difference estimate를 가능하게 한다. 적용 사례로는 Genovese et al. 및 Fefferman et al.의 bound를 평가하고, $\\beta$-VAE의 network layer 전반에서 geometry가 어떻게 진화하는지 분석하여, 현재 bound의 동작과 향후 이론을 안내하고 검증하는 데 있어 controlled benchmark의 가치를 강조한다. framework와 실험을 재현하기 위한 code는 제출물에 포함되어 있으며, 출판 시 open-source library로 공개될 예정이다."},{"id":"63052","en":"Instance-Dependent Continuous-Time Reinforcement Learning via Maximum Likelihood Estimation","ko":"최대우도추정을 통한 인스턴스 의존적 연속시간 강화학습","authors":"Runze Zhao, Yue Yu, Ruhan Wang, Chunfeng Huang, Dongruo Zhou","pos":"#4414","link":"https://openreview.net/forum?id=aXvTMnyeLj","abs":"

Continuous-time reinforcement learning (CTRL) provides a natural framework for sequential decision-making in dynamic environments where interactions evolve continuously over time. While CTRL has shown growing empirical success, its ability to adapt to varying levels of problem difficulty remains poorly understood. In this work, we investigate the instance-dependent behavior of CTRL and introduce a simple, model-based algorithm built on maximum likelihood estimation (MLE) with a general function approximator. Unlike existing approaches that estimate system dynamics directly, our method estimates the state marginal density to guide learning. We establish instance-dependent performance guarantees by deriving a regret bound that scales with the total reward variance and measurement resolution. Notably, the regret becomes independent of the specific measurement strategy when the observation frequency adapts appropriately to the problem’s complexity. To further improve performance, our algorithm incorporates a randomized measurement schedule that enhances sample efficiency without increasing measurement cost. These results highlight a new direction for designing CTRL algorithms that automatically adjust their learning behavior based on the underlying difficulty of the environment.

","absKo":"

Continuous-time reinforcement learning(CTRL)은 상호작용이 시간에 따라 연속적으로 진화하는 동적 환경에서의 순차적 의사결정을 위한 자연스러운 framework를 제공한다. CTRL은 점점 더 큰 실험적 성공을 보여왔지만, 문제 난이도의 다양한 수준에 어떻게 적응하는지는 아직 충분히 이해되지 않았다. 본 연구에서는 CTRL의 instance-dependent behavior를 조사하고, 일반적인 function approximator 위에 maximum likelihood estimation(MLE)을 기반으로 한 간단한 model-based algorithm을 제안한다. 시스템 dynamics를 직접 추정하는 기존 접근과 달리, 우리의 방법은 학습을 안내하기 위해 state marginal density를 추정한다. 우리는 total reward variance와 measurement resolution에 따라 스케일되는 regret bound를 도출함으로써 instance-dependent performance guarantee를 확립한다. 특히, observation frequency가 문제의 complexity에 맞게 적절히 적응하면 regret은 특정 measurement strategy와 무관해진다. 성능을 더 향상시키기 위해, 우리의 algorithm은 measurement cost를 증가시키지 않으면서 sample efficiency를 높이는 randomized measurement schedule을 포함한다. 이 결과는 환경의 근본적인 난이도에 따라 학습 behavior를 자동으로 조정하는 CTRL algorithm 설계를 위한 새로운 방향을 제시한다.

"},{"id":"60664","en":"On the Sharp Input-Output Analysis of Nonlinear Systems under Adversarial Attacks","ko":"Adversarial Attack 하의 비선형 시스템에 대한 정밀 Input-Output 분석","authors":"Jihun Kim, Yuchen Fang, Javad Lavaei","pos":"#4503","link":"https://openreview.net/forum?id=xrUgA8PrNN","abs":"This paper is concerned with learning the input-output mapping of general nonlinear dynamical systems. While the existing literature focuses on Gaussian inputs and benign disturbances, we significantly broaden the scope of admissible control inputs and allow correlated, nonzero-mean, adversarial disturbances. With our reformulation as a linear combination of basis functions, we prove that the $\\ell_2$-norm estimator overcomes the challenges posed by an adversary with access to the full information history, provided that the attack times are sparse, *i.e.*, the probability that the system is under adversarial attack at a given time is smaller than a certain threshold. We provide an estimation error bound that decays with the input memory length and prove its optimality by constructing a problem instance that suffers from the same bound under probabilistic adversarial attacks. Our work provides a sharp input-output analysis for a generic nonlinear and partially observed system under significantly generalized assumptions compared to existing works.","absKo":"이 논문은 일반적인 nonlinear dynamical system의 input-output mapping을 학습하는 문제를 다룬다. 기존 문헌이 Gaussian input과 온건한 disturbance에 초점을 맞춘 반면, 우리는 허용 가능한 control input의 범위를 크게 확장하여 correlated, zero-mean이 아닌 adversarial disturbance를 허용한다. basis function의 linear combination으로 재정식화함으로써, 시스템이 adversarial attack을 받는 시점의 확률이 특정 임계값보다 작은, 즉 attack time이 sparse한 경우에, 전체 information history에 접근할 수 있는 adversary가 제기하는 도전을 $\\ell_2$-norm estimator가 극복함을 증명한다. 우리는 input memory length에 따라 감소하는 estimation error bound를 제시하고, probabilistic adversarial attack 하에서 동일한 bound를 겪는 problem instance를 구성함으로써 그 최적성도 증명한다. 우리의 연구는 기존 작업들에 비해 상당히 일반화된 가정하에서, 일반적인 nonlinear 및 partially observed system에 대한 정밀한 input-output analysis를 제공한다."},{"id":"66444","en":"On Contraction of Sequential and Offset Rademacher Complexities","ko":"순차 및 오프셋 Rademacher 복잡도의 수축에 관하여","authors":"Adam Block, Alexander Rakhlin, Mark Sellke","pos":"#4504","link":"https://openreview.net/forum?id=3fXIRL0Wos","abs":"

The Rademacher complexity of a function class is among the most basic notions of its ``size'' and yields classical offline generalization bounds for Lipschitz loss functions that lead in turn to a modern understanding of statistical learning. More recently, the sequential and offset Rademacher complexities were introduced to prove analogous generalization bounds for online learning and for prediction with squared loss. A fundamental structural result in the theory of Rademacher complexity, with many applications to learning theory, is the Ledoux--Talagrand contraction lemma, which states that the Rademacher complexity of a composition of a function class with a fixed Lipschitz function is at most that of the original class. We show that, under structural assumptions on the function class, this contraction extends to sequential and offset Rademacher complexity at the price of polylogarithmic factors. We further show that these logarithmic factors cannot be removed in general and, absent these additional structural assumptions, no such contraction inequality can hold. These results together indicate that the sequential and offset Rademacher complexities behave fundamentally differently from the classical Rademacher complexity with respect to contraction, which in turn has broad implications for understanding the sample complexities of online learning and regression with squared loss for composed function classes.

","absKo":"

함수 클래스의 Rademacher complexity는 그 ``크기''를 나타내는 가장 기본적인 개념들 중 하나이며, Lipschitz loss function에 대한 classical offline generalization bound를 제공하고, 이는 다시 statistical learning에 대한 현대적 이해로 이어진다. 보다 최근에는 online learning과 squared loss 예측을 위한 유사한 generalization bound를 증명하기 위해 sequentialoffset Rademacher complexity가 도입되었다. learning theory에서 많은 응용을 가지는 Rademacher complexity 이론의 근본적인 구조적 결과는 Ledoux--Talagrand contraction lemma로, 이는 함수 클래스와 고정된 Lipschitz function의 합성에 대한 Rademacher complexity가 원래 클래스의 그것을 넘지 않는다고 말한다. 우리는 함수 클래스에 대한 구조적 가정 하에서, 이 contraction이 polylogarithmic factor를 대가로 sequential 및 offset Rademacher complexity로 확장됨을 보인다. 또한 이러한 log factor는 일반적으로 제거할 수 없으며, 추가적인 구조적 가정이 없으면 이러한 contraction inequality 자체가 성립할 수 없음을 보인다. 이 결과들은 sequential 및 offset Rademacher complexity가 contraction에 관하여 classical Rademacher complexity와 근본적으로 다르게 동작함을 함께 시사하며, 이는 합성 함수 클래스에 대한 online learning과 squared loss regression의 sample complexity를 이해하는 데 광범위한 함의를 갖는다.

"},{"id":"65417","en":"Linear Regression with Unknown Truncation Beyond Gaussian Features","ko":"가우시안 특징을 넘어선 미지 절단 선형 회귀","authors":"Alexandros Kouridakis, Anay Mehrotra, Alkis Kalavasis, Constantine Caramanis","pos":"#4505","link":"https://openreview.net/forum?id=DsV89lJ58l","abs":"In truncated linear regression, samples $(x,y)$ are shown only when the outcome $y$ falls inside a certain survival set $S^\\star$ and the goal is to estimate the unknown $d$-dimensional regressor $w^\\star$. This problem has a long history of study in Statistics and Machine Learning going back to the works of (Galton, 1897; Tobin, 1958) and more recently in, e.g., (Daskalakis et al., 2019; 2021; Lee et al., 2023; 2024). Despite this long history, however, most prior works are limited to the special case where $S^\\star$ is precisely known. The more practically relevant case, where $S^\\star$ is unknown and must be learned from data, remains open: indeed, here the only available algorithms require strong assumptions on the distribution of the feature vectors (e.g., Gaussianity) and, even then, have a $d^{\\mathrm{poly} (1/\\varepsilon)}$ run time for achieving $\\varepsilon$ accuracy. In this work, we give the first algorithm for truncated linear regression with unknown set case that runs in $\\mathrm{poly} (d/\\varepsilon)$ time, by only requiring that the feature vectors are sub-Gaussian. Our algorithm relies on a novel subroutine for efficiently learning unions of a bounded number of intervals using access to positive examples (without any negative examples) under a certain smoothness condition. This learning guarantee adds to the line of works on positive-only PAC learning and may be of independent interest.","absKo":"Truncated linear regression에서 sample $(x,y)$는 outcome $y$가 특정 survival set $S^\\star$ 안에 들어갈 때만 관측되며, 목표는 미지의 $d$-차원 regressor $w^\\star$를 추정하는 것이다. 이 문제는 Galton (1897), Tobin (1958)까지 거슬러 올라가는 Statistics와 Machine Learning에서의 긴 연구 역사를 가지며, 최근에는 Daskalakis et al. (2019; 2021), Lee et al. (2023; 2024) 등에서 다루어졌다. 그러나 이렇게 오랜 연구에도 불구하고, 기존 대부분의 연구는 $S^\\star$가 정확히 알려진 특수한 경우에만 제한된다. 더 실용적인 경우인 $S^\\star$가 미지이며 데이터로부터 학습해야 하는 경우는 여전히 열려 있는데, 실제로 현재 이용 가능한 알고리즘은 feature vector의 분포에 대해 강한 가정(예: Gaussianity)을 요구하고, 그 경우에도 $\\varepsilon$ 정확도를 달성하는 데 $d^{\\mathrm{poly}(1/\\varepsilon)}$의 실행 시간이 필요하다. 이 연구에서는 feature vector가 sub-Gaussian이라는 조건만으로, $\\mathrm{poly}(d/\\varepsilon)$ 시간에 동작하는 unknown set case의 truncated linear regression에 대한 첫 알고리즘을 제시한다. 우리의 알고리즘은 특정 smoothness condition 하에서 positive example만의 접근(negative example은 전혀 없이)으로 bounded number의 interval들의 union을 효율적으로 학습하는 새로운 subroutine에 의존한다. 이 learning guarantee는 positive-only PAC learning 계열의 연구에 추가되는 결과이며, 독립적인 관심사로도 의미가 있다."},{"id":"61951","en":"Learning Junta Distributions, Quantum Junta States, and QAC$^0$ Circuits","ko":"Junta 분포, Quantum Junta state, QAC$^0$ 회로 학습","authors":"Jinge Bao, Francisco Escudero Gutiérrez","pos":"#4506","link":"https://openreview.net/forum?id=lMieKzi6vx","abs":"In this work, we consider the problems of learning junta distributions, their quantum counterparts (quantum junta states), and $\\mathsf{QAC}^0$ circuits, which we show to be close to juntas. (1) Junta distributions. A probability distribution $p:${-1,1}$^n\\to \\mathbb [0,1]$ is a $k$-junta if it only depends on $k$ bits. We show that they can be learned to within additive error $\\varepsilon$ in total variation distance from $O(2^k\\log(n)/\\varepsilon^2)$ samples, which quadratically improves the upper bound of Aliakbarpour et al. (COLT'16) and matches their lower bound in every parameter. (2) Junta states. We initiate the study of $n$-qubit states that are $k$-juntas, those that are the tensor product of a $k$-qubit state and an $(n-k)$-qubit maximally mixed state. We show that these states can be learned with error $\\varepsilon$ in trace distance with $O(12^{k}\\log(n)/\\varepsilon^2)$ single copies. We also prove a lower bound of $\\Omega((4^k+\\log (n))/\\varepsilon^2)$ copies. Additionally, we show that, for constant $k$, $\\widetilde{\\Theta}(2^n/\\varepsilon^2)$ copies are necessary and sufficient to test whether a state is $\\varepsilon$-close or $7\\varepsilon$-far from being a $k$-junta. (3) $\\mathsf{QAC}^0$ circuits. We show that $n$-qubit $\\mathsf{QAC}^0$ circuits with size $s$, depth $d$ and $a$ auxiliary qubits can be learned from $2^{O(\\log(s^22^a)^d)}\\log(n)$ copies of the Choi state, improving the $n^{O(\\log(s^22^a)^d)}$ by Nadimpalli et al. (STOC'24). Along the way, we give new proof of the optimal performance of Classical Shadows based on Pauli analysis. We also strengthen the lower bounds against $\\mathsf{QAC}^0$ to compute the address function.","absKo":"이 연구에서는 junta distributions, 그 양자 대응물(quantum junta states), 그리고 $\\mathsf{QAC}^0$ circuits를 다루며, 우리는 이들이 juntas에 가깝다는 점을 보인다. (1) Junta distributions. 확률분포 $p:${-1,1}$^n\\to \\mathbb [0,1]$가 $k$-junta라는 것은 $k$개의 비트에만 의존한다는 뜻이다. 우리는 이들이 O(2^k\\log(n)/\\varepsilon^2)개의 샘플로 total variation distance에서 덧셈 오차 $\\varepsilon$ 이내까지 학습될 수 있음을 보인다. 이는 Aliakbarpour et al. (COLT'16)의 상계보다 이차적으로 개선된 것이며, 모든 매개변수에서 그들의 하한과 일치한다. (2) Junta states. 우리는 $n$-qubit 상태 중 $k$-junta인 상태, 즉 $k$-qubit 상태와 $(n-k)$-qubit maximally mixed state의 tensor product인 상태의 연구를 시작한다. 우리는 이러한 상태가 $O(12^{k}\\log(n)/\\varepsilon^2)$개의 single copy로 trace distance 오차 $\\varepsilon$ 이내까지 학습될 수 있음을 보인다. 또한 $\\Omega((4^k+\\log (n))/\\varepsilon^2)$개의 copy에 대한 하한도 증명한다. 추가로, 상수 $k$에 대해 어떤 상태가 $k$-junta에 $\\varepsilon$-close인지 또는 $7\\varepsilon$-far인지 테스트하는 데에는 $\\widetilde{\\Theta}(2^n/\\varepsilon^2)$개의 copy가 필요충분함을 보인다. (3) $\\mathsf{QAC}^0$ circuits. 우리는 size $s$, depth $d$, auxiliary qubits $a$를 갖는 $n$-qubit $\\mathsf{QAC}^0$ circuits가 Choi state의 $2^{O(\\log(s^22^a)^d)}\\log(n)$개 copy로 학습될 수 있음을 보이며, 이는 Nadimpalli et al. (STOC'24)의 $n^{O(\\log(s^22^a)^d)}$를 개선한 것이다. 이 과정에서 우리는 Pauli analysis에 기반한 Classical Shadows의 최적 성능에 대한 새로운 증명을 제시한다. 또한 address function을 계산하는 데 대한 $\\mathsf{QAC}^0$의 하한도 강화한다."},{"id":"62979","en":"High-Dimensional Learning Dynamics of Quantized Models with Straight-Through Estimator","ko":"Straight-Through Estimator를 갖는 양자화 모델의 고차원 학습 동역학","authors":"Yuma Ichikawa, Shuhei Kashiwamura, Ayaka Sakata","pos":"#4509","link":"https://openreview.net/forum?id=bI9moH3UZw","abs":"

Quantized neural network training optimizes a discrete, non-differentiable objective. The straight-through estimator (STE) enables backpropagation through surrogate gradients and is widely used. While previous studies have primarily focused on the properties of surrogate gradients and their convergence, the influence of quantization hyperparameters, such as bit width and quantization range, on learning dynamics remains largely unexplored. We theoretically show that in the high-dimensional limit, STE dynamics converge to a deterministic ordinary differential equation. This reveals that STE training exhibits a plateau followed by a sharp drop in generalization error, with plateau length depending on the quantization range. A fixed-point analysis quantifies the asymptotic deviation from the unquantized linear model. We also extend analytical techniques for stochastic gradient descent to nonlinear transformations of weights and inputs.

","absKo":"

Quantized neural network training은 이산적이고 미분 불가능한 목적함수를 최적화합니다. straight-through estimator (STE)는 surrogate gradient를 통한 backpropagation을 가능하게 하며 널리 사용됩니다. 기존 연구는 주로 surrogate gradient의 성질과 수렴에 초점을 맞췄지만, bit width와 quantization range 같은 quantization hyperparameter가 learning dynamics에 미치는 영향은 여전히 충분히 탐구되지 않았습니다. 우리는 이론적으로 high-dimensional limit에서 STE dynamics가 결정론적 ordinary differential equation으로 수렴함을 보입니다. 이를 통해 STE training은 plateau 뒤에 generalization error가 급격히 감소하는 패턴을 보이며, plateau 길이는 quantization range에 의존함을 알 수 있습니다. fixed-point analysis는 unquantized linear model로부터의 비점근적 편차를 정량화합니다. 또한 우리는 stochastic gradient descent에 대한 해석 기법을 weight와 input의 nonlinear transformation으로 확장합니다.

"},{"id":"64320","en":"Harmful Overfitting in Sobolev Spaces","ko":"Sobolev 공간에서의 해로운 과적합","authors":"Kedar Karhadkar, Alexander Sietsema, Deanna Needell, Guido Montufar","pos":"#4510","link":"https://openreview.net/forum?id=OmmKEy9Ixf","abs":"Motivated by recent work on benign overfitting in overparameterized machine learning, we study the generalization behavior of functions in Sobolev spaces $W^{k, p}(\\mathbb{R}^d)$ that perfectly fit a noisy training data set. Under assumptions of label noise and sufficient regularity in the data distribution, we show that approximately norm-minimizing interpolators, which are canonical solutions selected by smoothness bias, exhibit harmful overfitting: even as the training sample size $n \\to \\infty$, the generalization error remains bounded below by a positive constant with high probability. Our results hold for arbitrary values of $p \\in [1, \\infty)$, in contrast to prior results studying the Hilbert space case ($p = 2$) using kernel methods. Our proof uses a geometric argument which identifies harmful neighborhoods of the training data using Sobolev inequalities.","absKo":"overparameterized machine learning에서의 benign overfitting에 관한 최근 연구에 착안하여, 우리는 noisy training data set에 완벽히 맞는 Sobolev space $W^{k, p}(\\mathbb{R}^d)$의 함수들의 generalization behavior를 연구한다. label noise와 데이터 분포의 충분한 regularity를 가정하면, smoothness bias에 의해 선택되는 전형적인 해인 approximately norm-minimizing interpolator가 harmful overfitting을 보임을 보인다. 즉, training sample size가 $n \\to \\infty$로 커져도 generalization error는 높은 확률로 양의 상수 아래로 내려가지 않는다. 우리의 결과는 임의의 $p \\in [1, \\infty)$에 대해 성립하며, kernel method를 사용해 Hilbert space case($p = 2$)를 연구한 선행 결과와 대조된다. 우리의 증명은 Sobolev inequality를 이용해 training data 주변의 유해한 neighborhood를 식별하는 기하학적 논증을 사용한다."},{"id":"65650","en":"Gradient Flow Through Diagram Expansions: Learning Regimes and Explicit Solutions","ko":"다이어그램 전개를 통한 Gradient Flow: 학습 체제와 명시적 해","authors":"Dmitry Yarotsky, Eugene Golikov, Yaroslav Gusev","pos":"#4511","link":"https://openreview.net/forum?id=BXE3Z0EHCs","abs":"

We develop a general mathematical framework to analyze scaling regimes and derive explicit analytic solutions for gradient flow (GF) in large learning problems. Our key innovation is a formal power series expansion of the loss evolution, with coefficients encoded by diagrams akin to Feynman diagrams. We show that this expansion has a well-defined large-size limit that can be used to reveal different learning phases and, in some cases, to obtain explicit solutions of the nonlinear GF. We focus on learning Canonical Polyadic (CP) decompositions of high-order tensors, and show that this model has several distinct extreme lazy and rich GF regimes such as free evolution, NTK and under- and over-parameterized mean-field. We show that these regimes depend on the parameter scaling, tensor order, and symmetry of the model in a specific and subtle way. Moreover, we propose a general approach to summing the formal loss expansion by reducing it to a PDE; in a wide range of scenarios, it turns out to be 1st order and solvable by the method of characteristics. We observe a very good agreement of our theoretical predictions with experiment.

","absKo":"

우리는 scaling regime을 분석하고 large learning problem에서 gradient flow (GF)의 명시적 해석해를 도출하기 위한 일반적인 수학적 framework를 개발한다. 우리의 핵심 혁신은 loss evolution의 formal power series expansion이며, 그 계수는 Feynman diagram과 유사한 diagram으로 인코딩된다. 우리는 이 전개가 잘 정의된 large-size limit를 가지며, 이를 통해 서로 다른 learning phase를 드러내고 일부 경우에는 nonlinear GF의 명시적 해를 얻을 수 있음을 보인다. 우리는 고차 tensor의 Canonical Polyadic (CP) decomposition을 학습하는 문제에 초점을 맞추며, 이 모델이 free evolution, NTK, 그리고 under- 및 over-parameterized mean-field와 같은 여러 구별되는 극단적 lazy 및 rich GF regime을 가진다는 것을 보인다. 또한 이러한 regime이 parameter scaling, tensor order, 그리고 모델의 symmetry에 특정하고도 미묘한 방식으로 의존함을 보인다. 더 나아가, 우리는 formal loss expansion을 PDE로 환원하여 합산하는 일반적인 접근법을 제안하며, 광범위한 시나리오에서 이것이 1차 PDE가 되어 characteristics method로 풀 수 있음을 보인다. 우리의 이론적 예측은 실험과 매우 잘 일치함을 관찰한다.

"},{"id":"64169","en":"Full-Batch Gradient Descent Outperforms One-Pass SGD: Sample Complexity Separation in Single-Index Learning","ko":"전체 배치 경사 하강법이 단일 패스 SGD를 능가한다: 단일 인덱스 학습에서의 샘플 복잡도 분리","authors":"Filip Kovačević, Hong Chang Ji, Denny Wu, Mahdi Soltanolkotabi, Marco Mondelli","pos":"#4512","link":"https://openreview.net/forum?id=QItZDBVCT0","abs":"It is folklore that reusing training data more than once can improve the statistical efficiency of gradient-based learning. However, beyond linear regression, the theoretical advantage of full-batch gradient descent (GD, which always reuses all the data) over one-pass stochastic gradient descent (online SGD, which uses each data point only once) remains unclear. In this work, we consider learning a $d$-dimensional single-index model with a quadratic activation, for which it is known that one-pass SGD requires $n\\gtrsim d\\log d$ samples to achieve weak recovery. We first show that this $\\log d$ factor in the sample complexity persists for full-batch spherical GD on the correlation loss; however, by simply truncating the activation, full-batch GD exhibits a favorable optimization landscape at $n \\simeq d$ samples, thereby outperforming one-pass SGD (with the same activation) in statistical efficiency. We complement this result with a trajectory analysis of full-batch GD on the squared loss from small initialization, showing that $n \\gtrsim d$ samples and $T \\gtrsim\\log d$ gradient steps suffice to achieve strong (exact) recovery.","absKo":"훈련 데이터를 한 번 이상 재사용하면 gradient-based learning의 통계적 효율이 향상된다는 것은 잘 알려진 사실이다. 그러나 linear regression을 넘어서는 경우, 항상 모든 데이터를 재사용하는 full-batch gradient descent (GD)가 각 데이터 포인트를 한 번만 사용하는 one-pass stochastic gradient descent (online SGD)보다 갖는 이론적 이점은 여전히 명확하지 않다. 본 연구에서는 quadratic activation을 갖는 $d$차원 single-index model의 학습을 고려한다. 이 문제에서 one-pass SGD가 weak recovery를 달성하려면 $n\\gtrsim d\\log d$개의 샘플이 필요하다는 것이 알려져 있다. 우리는 먼저 correlation loss에 대한 full-batch spherical GD에서도 이 sample complexity의 $\\log d$ factor가 지속됨을 보인다. 그러나 activation을 단순히 truncating하면 full-batch GD는 $n \\simeq d$ 샘플에서 유리한 optimization landscape를 보이며, 동일한 activation을 사용하는 one-pass SGD보다 통계적 효율에서 우수함을 보인다. 우리는 작은 초기화에서 squared loss에 대한 full-batch GD의 trajectory analysis로 이 결과를 보완하며, $n \\gtrsim d$개의 샘플과 $T \\gtrsim\\log d$개의 gradient step이면 strong (exact) recovery를 달성하기에 충분함을 보인다."},{"id":"65132","en":"High-accuracy sampling for diffusion models and log-concave distributions","ko":"Diffusion Model과 Log-Concave 분포를 위한 고정밀 샘플링","authors":"Fan Chen, Sinho Chewi, Constantinos Daskalakis, Alexander Rakhlin","pos":"#4625","link":"https://openreview.net/forum?id=GW3umRqsZZ","abs":"We present algorithms for diffusion model sampling which obtain $\\delta$-error in $\\mathrm{polylog}(1/\\delta)$ steps, given access to $\\widetilde O(\\delta)$-accurate score estimates in $L^2$. This is an exponential improvement over all previous results. Specifically, under minimal data assumptions, the complexity is $\\widetilde O(d\\mathrm{polylog}(1/\\delta))$ where $d$ is the dimension of the data; under a non-uniform $L$-Lipschitz condition, the complexity is $\\widetilde O(\\sqrt{dL}\\mathrm{polylog}(1/\\delta))$; and if the data distribution has intrinsic dimension $d_\\star$, then the complexity reduces to $\\widetilde O(d_\\star\\mathrm{polylog}(1/\\delta))$. Our approach also yields the first $\\mathrm{polylog}(1/\\delta)$ complexity sampler for general log-concave distributions using only gradient evaluations.","absKo":"우리는 $L^2$에서 $\\widetilde O(\\delta)$-정확도의 score estimate에 접근할 수 있을 때, $\\mathrm{polylog}(1/\\delta)$ 단계만에 $\\delta$-error를 얻는 diffusion model sampling algorithm을 제시한다. 이는 이전의 모든 결과에 대한 exponential improvement이다. 구체적으로, 최소한의 data assumption 하에서 complexity는 데이터 차원 $d$에 대해 $\\widetilde O(d\\mathrm{polylog}(1/\\delta))$이고, non-uniform $L$-Lipschitz condition 하에서는 $\\widetilde O(\\sqrt{dL}\\mathrm{polylog}(1/\\delta))$이며, data distribution이 intrinsic dimension $d_\\star$를 가지면 complexity는 $\\widetilde O(d_\\star\\mathrm{polylog}(1/\\delta))$로 줄어든다. 우리의 접근법은 또한 gradient evaluation만을 사용하여 general log-concave distribution에 대해 $\\mathrm{polylog}(1/\\delta)$ complexity를 갖는 첫 sampler를 제공한다."},{"id":"62341","en":"A Task-centric Theory for Iterative Self-Improvement with Easy-to-Hard Curricula","ko":"Easy-to-Hard 커리큘럼을 통한 반복적 Self-Improvement의 Task-Centric 이론","authors":"Chenruo Liu, Yijun Dong, Yiqiu Shen, Qi Lei","pos":"#1609","link":"https://openreview.net/forum?id=hUjHHGOocl","abs":"

Iterative self-improvement fine-tunes an autoregressive large language model (LLM) on reward-verified outputs generated by the LLM itself. In contrast to the empirical success of self-improvement, the theoretical foundation of this generative, iterative procedure in a practical, finite-sample setting remains limited. We make progress toward this goal by modeling each round of self-improvement as maximum-likelihood fine-tuning on a reward-filtered distribution and deriving finite-sample guarantees for the expected reward. Our analysis reveals an explicit feedback loop where better models accept more data per iteration, supporting sustained self-improvement while explaining eventual saturation of such improvement. Adopting a task-centric view by considering reasoning tasks with multiple difficulty levels, we further prove quantifiable conditions on model initialization, task difficulty, and sample budget where easy-to-hard curricula provably achieve better guarantees than training on fixed mixtures of tasks. Our analyses are validated via Monte-Carlo simulations and controlled experiments on graph-based reasoning tasks.

","absKo":"

Iterative self-improvement는 autoregressive large language model (LLM)을 LLM 자신이 생성한 reward-verified output으로 fine-tuning하는 방법이다. self-improvement의 경험적 성공과 달리, 실용적인 finite-sample setting에서 이 생성적 반복 절차의 이론적 기반은 여전히 제한적이다. 우리는 각 self-improvement round를 reward-filtered distribution 위의 maximum-likelihood fine-tuning으로 모델링하고 expected reward에 대한 finite-sample guarantee를 유도함으로써 이 목표에 진전을 이룬다. 우리의 분석은 더 나은 모델이 iteration마다 더 많은 데이터를 수용하는 명시적 feedback loop를 드러내며, 지속적인 self-improvement를 뒷받침하는 동시에 이러한 improvement가 결국 포화되는 이유를 설명한다. 여러 난이도 수준을 가진 reasoning task를 고려하는 task-centric view를 채택하여, 우리는 모델 initialization, task difficulty, sample budget에 대한 정량화 가능한 조건을 추가로 증명하며, 이 조건하에서 easy-to-hard curriculum이 고정된 task mixture로 학습하는 것보다 더 나은 guarantee를 달성함을 보인다. 우리의 분석은 Monte-Carlo simulation과 graph-based reasoning task에 대한 통제된 실험으로 검증된다.

"},{"id":"62941","en":"Provable Sample Efficiency of Curriculum Post-Training for Transformer Reasoning","ko":"Transformer Reasoning을 위한 Curriculum Post-Training의 샘플 효율성 보장","authors":"Dake Bu, Wei Huang, Andi Han, Atsushi Nitanda, Hau-San Wong, Qingfu Zhang, Taiji Suzuki","pos":"#1613","link":"https://openreview.net/forum?id=bgM2hmXExQ","abs":"

Recent curriculum techniques in the post-training stage of LLMs have been empirically observed to outperform non-curriculum approaches in improving reasoning performance, yet a principled understanding of their effectiveness and limitations remains incomplete. To bridge this gap, we develop an abstract theoretical framework and identify sufficient conditions under which curriculum post-training yields exponential improvements in sample complexity. To substantiate this framework, we model the base model’s Chain-of-Thought generation as a state-conditioned autoregressive reasoning tree, and formalize curriculum subtasks as either depth-increasing curricula that progressively extend reasoning horizons or hint-decreasing curricula that gradually remove partial hints. Our analysis shows that, under outcome-only reward signals, reinforcement learning finetuning with both curriculum strategies achieves high accuracy with polynomial sample complexity, whereas non-curriculum counterpart encounters an exponential complexity bottleneck. We further establish analogous guarantees for test-time scaling, demonstrating that curriculum-aware querying strategies reduce both reward oracle complexity and sampling cost from exponential to polynomial order. Empirical simulations support our theoretical findings.

","absKo":"

최근 LLM의 post-training 단계에서 사용되는 curriculum technique은 reasoning 성능 향상에서 non-curriculum 접근보다 경험적으로 우수한 것으로 관찰되었지만, 그 효과와 한계에 대한 원리적 이해는 여전히 완전하지 않다. 이러한 간극을 메우기 위해, 우리는 추상적 이론 프레임워크를 구축하고 curriculum post-training이 sample complexity에서 지수적 향상을 가져오는 충분 조건을 식별한다. 이 프레임워크를 뒷받침하기 위해, 우리는 base model의 Chain-of-Thought 생성을 state-conditioned autoregressive reasoning tree로 모델링하고, curriculum subtask를 reasoning horizon을 점진적으로 확장하는 depth-increasing curriculum 또는 부분 힌트를 점진적으로 제거하는 hint-decreasing curriculum으로 형식화한다. 우리의 분석에 따르면, outcome-only reward signal 하에서 두 curriculum 전략을 사용한 reinforcement learning finetuning은 polynomial sample complexity로 높은 정확도를 달성하는 반면, non-curriculum 대응 방식은 지수적 complexity 병목에 직면한다. 우리는 또한 test-time scaling에 대해서도 유사한 보장을 확립하여, curriculum-aware querying strategy가 reward oracle complexity와 sampling cost를 모두 지수 차수에서 다항 차수로 줄임을 보인다. 실험적 simulation은 우리의 이론적 결과를 뒷받침한다.

"},{"id":"60871","en":"Off-Policy Evaluation for Missingness-Aware Policies in MDPs with Rewards Missing Not at Random","ko":"보상이 MNAR로 누락되는 MDP에서 Missingness-Aware Policy의 Off-Policy Evaluation","authors":"Ziheng Wei, Rui Miao, Annie Qu","pos":"#209","link":"https://openreview.net/forum?id=vpSFJoxyDz","abs":"

In offline Reinforcement Learning, immediate rewards in logged batch data are often unobserved due to sparse or irregular record-keeping, or censored beyond certain reward values. This issue frequently arises in practical settings, including health care and marketing. We investigate off-policy evaluation (OPE) in finite-horizon Markov decision processes when rewards are missing not at random (MNAR), which breaks ignorability and induces selection bias even after conditioning on states and actions. To address this, we formalize a reward-dependent propensity model and use future states as shadow variables to identify the full-data conditional mean reward. We further introduce a bridge function that recovers the conditional mean reward without explicitly modeling the MNAR mechanism, and estimate it via a min-max procedure to avoid double sampling. Building upon these identification results, we propose an Fitted-Q-Evaluation-style estimator that propagates the recovered rewards while allowing target policies to depend on past missingness indicators. Finally, we establish consistency and finite-sample error bounds for our OPE estimator, and show through simulations the strong performance of our method compared to existing benchmarks.

","absKo":"

offline Reinforcement Learning에서는 logged batch data의 즉각적 보상이 sparse하거나 불규칙한 기록, 또는 특정 reward 값 이상에서의 censoring 때문에 관측되지 않는 경우가 많다. 이러한 문제는 health care와 marketing을 포함한 실무 환경에서 자주 발생한다. 우리는 reward가 missing not at random (MNAR)인 finite-horizon Markov decision process에서 off-policy evaluation (OPE)을 연구하며, 이는 ignorability를 깨뜨리고 state와 action으로 조건화한 뒤에도 selection bias를 유발한다. 이를 해결하기 위해, 우리는 reward-dependent propensity model을 정식화하고 미래 상태를 shadow variable로 사용하여 full-data conditional mean reward를 식별한다. 또한 MNAR mechanism을 명시적으로 모델링하지 않고도 conditional mean reward를 복원하는 bridge function을 도입하고, double sampling을 피하기 위해 min-max procedure로 이를 추정한다. 이러한 identification 결과를 바탕으로, 우리는 target policy가 과거 missingness indicator에 의존하도록 허용하면서 복원된 reward를 전파하는 Fitted-Q-Evaluation 스타일 estimator를 제안한다. 마지막으로, 우리는 OPE estimator에 대한 consistency와 finite-sample error bound를 확립하고, 시뮬레이션을 통해 기존 benchmark와 비교했을 때 우리의 방법이 강한 성능을 보임을 입증한다.

"},{"id":"64950","en":"Understanding Private Learning From Feature Perspective","ko":"특징 관점에서 프라이빗 학습 이해하기","authors":"Meng Ding, Mingxi Lei, Shaopeng Fu, Shaowei Wang, Di Wang, Jinhui Xu","pos":"#3712","link":"https://openreview.net/forum?id=IP6W8i7XPG","abs":"

Differentially private Stochastic Gradient Descent (DP-SGD) has become integral to privacy-preserving machine learning, ensuring robust privacy guarantees in sensitive domains. Despite notable empirical advances leveraging features from non-private, pre-trained models to enhance DP-SGD training, a theoretical understanding of feature dynamics in private learning remains underexplored. This paper presents the first theoretical framework to analyze private training through a feature learning perspective. Building on the multi-patch data structure from prior work, our analysis distinguishes between label-dependent feature signals and label-independent noise—a critical aspect overlooked by existing analyses in the DP community. Employing a two-layer CNN with polynomial ReLU activation, we theoretically characterize both feature signal learning and data noise memorization in private training via noisy gradient descent. Our findings reveal that (1) Effective private signal learning requires a higher signal-to-noise ratio (SNR) compared to non-private training, and (2) When data noise memorization occurs in non-private learning, it will also occur in private learning, leading to poor generalization despite small training loss. Our findings highlight the challenges of private learning and prove the benefit of feature enhancement to improve SNR. Experiments on synthetic and real-world datasets also validate our theoretical findings.

","absKo":"

Differentially private Stochastic Gradient Descent (DP-SGD)는 privacy-preserving machine learning의 핵심이 되어, 민감한 도메인에서 강력한 privacy guarantee를 보장한다. 비-private pre-trained model에서 추출한 feature를 활용해 DP-SGD 학습을 향상시키는 주목할 만한 경험적 진전이 있었음에도, private learning에서의 feature dynamics에 대한 이론적 이해는 아직 충분히 다뤄지지 않았다. 본 논문은 feature learning 관점에서 private training을 분석하는 최초의 이론적 framework를 제시한다. 이전 연구의 multi-patch data structure를 기반으로, 우리의 분석은 label-dependent feature signal과 label-independent noise를 구분하는데, 이는 DP community의 기존 분석에서 간과된 핵심 요소이다. polynomial ReLU activation을 사용하는 two-layer CNN을 통해, 우리는 noisy gradient descent에서 private training 중의 feature signal learning과 data noise memorization을 이론적으로 특성화한다. 우리의 결과는 (1) 효과적인 private signal learning은 non-private training보다 더 높은 signal-to-noise ratio (SNR)를 요구하며, (2) 비-private learning에서 data noise memorization이 발생하면 private learning에서도 역시 발생하여, training loss는 작음에도 불구하고 일반화 성능이 나빠진다는 점을 보여준다. 이러한 결과는 private learning의 어려움을 강조하며 SNR 향상을 위한 feature enhancement의 이점을 증명한다. synthetic 및 real-world dataset에서의 실험 또한 우리의 이론적 발견을 검증한다.

"},{"id":"62775","en":"On the Power of Source Screening for Learning Shared Feature Extractors","ko":"공유 Feature Extractors 학습에서 Source Screening의 효과","authors":"Muxing Wang, Connor Mclaughlin, Lili Su","pos":"#4013","link":"https://openreview.net/forum?id=dTMrITkTr5","abs":"

Learning with shared representation is widely recognized as an effective way to separate commonalities from heterogeneity across various heterogeneous sources. Most existing work includes all related data sources via simultaneously training a common feature extractor and source-specific heads. It is well understood that data sources with low relevance or poor quality may hinder representation learning. In this paper, we further dive into the question of which data sources should be learned jointly by focusing on the traditionally deemed \"good\" collection of sources, in which individual sources have similar relevance and qualities with respect to the true underlying common structure. Towards tractability, we focus on the linear setting where sources share a low-dimensional subspace. We find that source screening can play a central role in statistically optimal subspace estimation. We show that, for a broad class of problem instances, training on a carefully selected subset of sources suffices to achieve minimax optimality, even when a substantial portion of data is discarded. We formalize the notion of an informative subpopulation, develop algorithms and practical heuristics for identifying such subsets, and validate their effectiveness through both theoretical analysis and empirical evaluations on synthetic and real-world datasets.

","absKo":"shared representation으로 학습하는 것은 다양한 heterogeneous source 전반의 공통성과 이질성을 분리하는 효과적인 방법으로 널리 인식된다. 기존의 대부분 연구는 공통 feature extractor와 source-specific head를 동시에 학습함으로써 모든 관련 data source를 포함한다. relevance가 낮거나 품질이 좋지 않은 data source가 representation learning을 방해할 수 있다는 점은 잘 알려져 있다. 이 논문에서는 true underlying common structure에 대해 각 source가 유사한 relevance와 quality를 갖는 전통적으로 \"좋은\" source 집합에 초점을 맞추어, 어떤 data source를 함께 학습해야 하는지라는 문제를 더 깊이 탐구한다. 계산 가능성을 위해, source들이 저차원 subspace를 공유하는 linear setting에 초점을 맞춘다. 우리는 source screening이 statistically optimal subspace estimation에서 중심적인 역할을 할 수 있음을 보인다. 광범위한 problem instance에 대해, carefully selected subset of source로만 학습해도 상당한 양의 data를 버리더라도 minimax optimality를 달성하기에 충분함을 보인다. 우리는 informative subpopulation의 개념을 formalize하고, 이러한 subset을 식별하기 위한 알고리즘과 실용적 heuristic을 개발하며, theoretical analysis와 synthetic 및 real-world dataset에 대한 empirical evaluation을 통해 그 효과를 검증한다.

"},{"id":"64369","en":"The Fisher Dimension: Instance-Dependent Complexity for Causal Discovery","ko":"Fisher 차원: 인과 발견을 위한 인스턴스 의존적 복잡도","authors":"Luong Doan, Khanh Quoc, Duc Nguyen, Mai Hung, Phong Ho, Nhung Duong, Tuan Do","pos":"#4114","link":"https://openreview.net/forum?id=OIigKb2bnt","abs":"

Classical sample complexity bounds for causal structure learning are minimax in nature, characterizing worst-case difficulty without distinguishing between easy and hard instances. We study instance-specific complexity for Markov equivalence class (MEC) recovery in linear Gaussian structural equation models. We introduce the Fisher dimension, defined as the inverse squared minimum partial correlation that must be detected to recover the MEC. We prove that the Fisher dimension governs sample complexity: it provides both a lower bound and an upper bound (tight up to logarithmic factors) for MEC recovery. A key theoretical finding is that under spectrally well-conditioned models, with bounded noise variances, bounded covariance eigenvalues, and constant-order edge coefficients, the Fisher dimension is uniformly bounded regardless of graph structure. Thus, significant instance-specific variation arises from parametric rather than structural features. Empirical validation shows strong correlation between our predictor and observed sample complexity for structured graph families.

","absKo":"

causal structure learning에 대한 고전적인 sample complexity bound는 minimax 성격을 가지며, 쉬운 경우와 어려운 경우를 구분하지 않고 최악의 경우 난이도만을 특징짓는다. 우리는 linear Gaussian structural equation model에서 Markov equivalence class (MEC) 복구를 위한 instance-specific complexity를 연구한다. 우리는 MEC를 복구하기 위해 탐지되어야 하는 minimum partial correlation의 제곱 역수로 정의되는 Fisher dimension을 도입한다. 우리는 Fisher dimension이 sample complexity를 지배함을 증명한다. 즉, MEC 복구에 대해 lower bound와 upper bound를 모두 제공하며, 이 upper bound는 logarithmic factor를 제외하면 tight하다. 핵심 이론적 발견은 spectrally well-conditioned model에서 noise variance가 bounded이고 covariance eigenvalue가 bounded이며 edge coefficient가 constant order일 때, graph structure와 무관하게 Fisher dimension이 전반적으로 bounded하다는 점이다. 따라서 instance-specific variation의 큰 부분은 structural feature보다 parametric feature에서 비롯된다. 실증적 검증은 우리의 predictor와 구조화된 graph family에서 관측된 sample complexity 사이에 강한 상관관계가 있음을 보여준다.

"},{"id":"65322","en":"Realizable Bayes-Consistency for General Metric Losses","ko":"일반 메트릭 손실에 대한 실현 가능한 Bayes 일관성","authors":"Dan Tsir Cohen, Steve Hanneke, Aryeh Kontorovich","pos":"#4400","link":"https://openreview.net/forum?id=EmqsPzyNHh","abs":"We study strong universal Bayes-consistency in the realizable setting for learning with general metric losses, extending classical characterizations beyond $0$-$1$ classification \\citep{bousquet_theory_2021, hanneke2021universalbayesconsistencymetric} and real-valued regression \\citep{attias_universal_2024}. Given an instance space $(\\mathcal X,\\rho)$, a label space $(\\mathcal Y,\\ell)$ with possibly unbounded loss, and a hypothesis class $\\mathcal H \\subseteq \\mathcal Y^{\\mathcal X}$, we solve an open problem presented in \\cite{pmlr-v178-cohen22a}. Specifically, we find the necessary and sufficient conditions on the hypothesis class $\\mathcal H$ under which there exists a distribution-free learning rule whose risk converges almost surely to the best-in-class risk (which is zero) for every realizable data-generating distribution. Our main contribution is this sharp characterization in terms of a combinatorial obstruction: Similarly to \\citet{attias_universal_2024}, we introduce the notion of an infinite non-decreasing $(\\gamma_k)$-Littlestone tree, where $\\gamma_k \\to \\infty$. This extends the Littlestone tree structure used in \\citep{bousquet_theory_2021} to the metric loss setting.","absKo":"우리는 일반 metric loss를 사용하는 학습에서 realizable setting에 대한 strong universal Bayes-consistency를 연구하며, 고전적인 characterization을 $0$-$1$ classification \\citep{bousquet_theory_2021, hanneke2021universalbayesconsistencymetric}과 실수값 regression \\citep{attias_universal_2024}을 넘어 확장한다. instance space $(\\mathcal X,\\rho)$, 그리고 possibly unbounded loss를 갖는 label space $(\\mathcal Y,\\ell)$, hypothesis class $\\mathcal H \\subseteq \\mathcal Y^{\\mathcal X}$가 주어졌을 때, 우리는 \\cite{pmlr-v178-cohen22a}에서 제기된 열린 문제를 해결한다. 구체적으로, 모든 realizable data-generating distribution에 대해 risk가 almost surely best-in-class risk(여기서는 0)로 수렴하는 distribution-free learning rule이 존재하기 위한 hypothesis class $\\mathcal H$의 필요충분조건을 찾는다. 우리의 주된 기여는 조합론적 장애에 대한 sharp characterization이다. \\citet{attias_universal_2024}와 유사하게, 우리는 $\\gamma_k \\to \\infty$를 만족하는 infinite non-decreasing $(\\gamma_k)$-Littlestone tree라는 개념을 도입한다. 이는 metric loss setting으로 \\citep{bousquet_theory_2021}에서 사용된 Littlestone tree 구조를 확장한 것이다."},{"id":"62204","en":"Sharp Inequalities between Total Variation and Hellinger Distances for Gaussian Mixtures","ko":"Gaussian Mixture에 대한 Total Variation과 Hellinger Distance 간의 Sharp inequality","authors":"Joonhyuk Jung, Chao Gao","pos":"#4402","link":"https://openreview.net/forum?id=ihMB4kA2SQ","abs":"We study the relation between the total variation (TV) and Hellinger distances between two Gaussian location mixtures. Our first result establishes a general upper bound: for any two mixing distributions supported on a compact set, the Hellinger distance between the two mixtures is controlled by the TV distance raised to a power $1-o(1)$, where the $o(1)$ term is of order $1/\\log\\log(1/\\mathrm{TV})$. We also construct two sequences of mixing distributions that demonstrate the sharpness of this bound. Taken together, our results resolve an open problem raised in Jia et al. (2023) and thus lead to an entropic characterization of learning Gaussian mixtures in total variation. Our inequality also yields optimal robust estimation of Gaussian mixtures in Hellinger distance, which has a direct implication for bounding the minimax regret of empirical Bayes under Huber contamination.","absKo":"우리는 두 Gaussian location mixture 사이의 total variation(TV) 거리와 Hellinger 거리의 관계를 연구한다. 첫 번째 결과로, 컴팩트 집합에 지지(support)를 갖는 임의의 두 mixing distribution에 대해, 두 mixture 사이의 Hellinger 거리는 TV 거리의 $1-o(1)$ 제곱에 의해 제어된다는 일반적인 상한을 보인다. 여기서 $o(1)$ 항은 $1/\\log\\log(1/\\mathrm{TV})$의 차수이다. 또한 이 경계의 sharpness를 보여주는 두 개의 mixing distribution 수열을 구성한다. 이 결과들을 종합하면 Jia et al. (2023)에서 제기된 열린 문제를 해결하게 되며, 이에 따라 total variation에서 Gaussian mixture를 학습하는 문제에 대한 entropic characterization을 얻는다. 우리의 부등식은 또한 Hellinger 거리에서 Gaussian mixture의 optimal robust estimation을 도출하며, 이는 Huber contamination 하에서 empirical Bayes의 minimax regret을 상계하는 데 직접적인 함의를 갖는다."},{"id":"62752","en":"Statistical Learning Theory in Lean 4: Empirical Processes from Scratch","ko":"Lean 4의 통계학습이론: Empirical Processes를 기초부터","authors":"Yuanhe Zhang, Jason Lee, Fanghui Liu","pos":"#4404","link":"https://openreview.net/forum?id=dfqmQ9WhCP","abs":"

We present the first comprehensive Lean 4 formalization of statistical learning theory (SLT) grounded in empirical process theory. Our end-to-end formal infrastructure implement the missing contents in latest Lean library, including a complete development of Gaussian Lipschitz concentration, the first formalization of Dudley’s entropy integral theorem for sub-Gaussian processes, and an application to least-squares regression with a sharp rate. The project was carried out using a human–AI collaborative workflow, in which humans design proof strategies and AI agents execute tactical proof construction, resulting in approximately 30,000 lines of human-verified Lean 4 code produced over 500 hours of supervised development. Beyond implementation, the formalization process exposes and resolves implicit assumptions and missing details in standard SLT textbooks, enforcing a granular, line-by-line understanding of the theory. This work establishes a reusable formal foundation for future developments in machine learning theory. The code is provided in the supplementary materials.

","absKo":"

우리는 empirical process theory에 기반한 statistical learning theory(SLT)의 최초의 포괄적인 Lean 4 formalization을 제시한다. 우리의 end-to-end formal infrastructure는 최신 Lean library에 부족했던 내용을 구현하며, 여기에는 Gaussian Lipschitz concentration의 완전한 development, sub-Gaussian processes에 대한 Dudley의 entropy integral theorem의 최초 formalization, 그리고 sharp rate를 갖는 least-squares regression에의 적용이 포함된다. 이 프로젝트는 human–AI collaborative workflow를 통해 수행되었으며, 인간은 proof strategy를 설계하고 AI agents는 tactical proof construction을 실행했다. 그 결과, 약 500시간의 supervised development 동안 인간이 검증한 약 30,000줄의 Lean 4 code가 생성되었다. 구현을 넘어, formalization 과정은 표준 SLT 교재에 내재된 암묵적 가정과 누락된 세부사항을 드러내고 해결함으로써, 이론에 대한 세밀한 line-by-line 이해를 강제한다. 이 작업은 향후 machine learning theory 발전을 위한 재사용 가능한 formal foundation을 확립한다. 코드는 supplementary materials에 제공된다.

"},{"id":"61938","en":"Tight Stability Bounds for Robust Distributed Learning: Byzantine Failures Hurt Generalization More than Data Poisoning","ko":"강건한 분산 학습을 위한 tight stability bounds: Byzantine failure가 데이터 포이즈닝보다 일반화 성능을 더 크게 저해","authors":"Thomas Boudou, Batiste Le Bars, Nirupam Gupta, Aurélien Bellet","pos":"#4409","link":"https://openreview.net/forum?id=lWyXszYR58","abs":"Robust distributed learning algorithms aim to maintain reliable performance despite the presence of misbehaving workers. Such misbehaviors are commonly modeled as *Byzantine failures*, allowing arbitrarily corrupted communication, or as *data poisoning*, a weaker form of corruption restricted to local training data. While prior work shows similar optimization guarantees for both models, an important question remains: *How do these threat models impact generalization?* We show, for the first time, a fundamental gap in generalization guarantees between the two threat models: Byzantine failures yield strictly worse rates than those achievable under data poisoning. Our findings leverage a tight algorithmic stability analysis of robust distributed learning. Specifically, we prove that: *(i)* under data poisoning, the uniform algorithmic stability of an algorithm with optimal optimization guarantees degrades by an additive factor of $\\varTheta ( \\frac{f}{n-f} )$, with $f$ out of $n$ workers misbehaving; whereas *(ii)* under Byzantine failures, the degradation is in $\\Omega \\big( \\sqrt{ \\frac{f}{n-2f}} \\big)$.","absKo":"강건한 distributed learning 알고리즘은 비정상적으로 행동하는 worker가 존재하더라도 신뢰할 수 있는 성능을 유지하는 것을 목표로 한다. 이러한 비정상 행위는 흔히 communication이 임의로 손상될 수 있는 *Byzantine failures*로 모델링되거나, local training data에 국한된 더 약한 형태의 corruption인 *data poisoning*으로 모델링된다. 이전 연구는 두 모델에 대해 유사한 optimization guarantee를 보였지만, 중요한 질문은 여전히 남아 있다: *이러한 threat model이 generalization에 어떤 영향을 미치는가?* 우리는 처음으로, 두 threat model 사이의 generalization guarantee에 근본적인 격차가 있음을 보인다. 즉, Byzantine failures는 data poisoning 하에서 달성 가능한 것보다 더 나쁜 rate를 초래한다. 우리의 결과는 robust distributed learning에 대한 정밀한 algorithmic stability 분석에 기반한다. 구체적으로, 우리는 다음을 증명한다. *(i)* data poisoning 하에서는, $n$개의 worker 중 $f$명이 비정상적으로 행동할 때 optimal optimization guarantee를 가진 알고리즘의 uniform algorithmic stability는 $\\varTheta ( \\frac{f}{n-f} )$의 additive factor만큼 악화된다. 반면 *(ii)* Byzantine failures 하에서는 그 악화가 $\\Omega \\big( \\sqrt{ \\frac{f}{n-2f}} \\big)$이다."},{"id":"62520","en":"Next-Token Prediction and Regret Minimization","ko":"Next-Token Prediction과 Regret Minimization","authors":"Mehryar Mohri, Clayton Sanford, Jon Schneider, Kiran Vodrahalli, Yifan Wu","pos":"#4411","link":"https://openreview.net/forum?id=frRHI1zK86","abs":"We consider the question of how to employ next-token prediction algorithms in adversarial online decision making environments. Specifically, if we train a next-token prediction model on a distribution $\\mathcal{D}$ over sequences of opponent actions, when is it the case that the induced online decision making algorithm (by approximately best responding to the model's predictions) has low adversarial regret (i.e., when is $\\mathcal{D}$ a \\emph{low-regret distribution})? For unbounded context windows (where the prediction made by the model can depend on all the actions taken by the adversary thus far), we show that although not every distribution $\\mathcal{D}$ is a low-regret distribution, every distribution $\\mathcal{D}$ is exponentially close (in TV distance) to one low-regret distribution, and hence sublinear regret can always be achieved at negligible cost to the accuracy of the original next-token prediction model. In contrast to this, for bounded context windows (where the prediction made by the model can depend only on the past $w$ actions taken by the adversary, as may be the case in modern transformer architectures), we show that there are some distributions $\\mathcal{D}$ of opponent play that are $\\Theta(1)$-far from any low-regret distribution $\\mathcal{D'}$ (even when $w = \\Omega(T)$ and such distributions exist). Finally, we complement these results by showing that the unbounded context robustification procedure can be implemented by layers of a standard transformer architecture, and provide empirical evidence that transformer models can be efficiently trained to represent these new low-regret distributions.","absKo":"우리는 적대적 online decision making 환경에서 next-token prediction 알고리즘을 어떻게 활용할지에 대한 문제를 다룬다. 구체적으로, 상대의 행동 시퀀스에 대한 분포 $\\mathcal{D}$ 위에서 next-token prediction model을 학습할 때, 그 model의 예측에 approximately best responding함으로써 유도된 online decision making algorithm이 언제 낮은 adversarial regret를 가지는가(즉, 언제 $\\mathcal{D}$가 \\emph{low-regret distribution}인가)? unbounded context window(즉, model이 내린 예측이 지금까지 adversary가 취한 모든 행동에 의존할 수 있는 경우)에서는, 모든 분포 $\\mathcal{D}$가 low-regret distribution은 아니지만, 모든 분포 $\\mathcal{D}$는 하나의 low-regret distribution에 지수적으로 가깝게(TV distance 기준) 놓일 수 있음을 보인다. 따라서 원래 next-token prediction model의 정확도에 거의 비용을 들이지 않고도 항상 sublinear regret를 달성할 수 있다. 반면 bounded context window(즉, 현대 transformer architecture에서처럼 model의 예측이 adversary가 취한 과거 $w$개의 행동에만 의존할 수 있는 경우)에서는, 일부 상대의 행동 분포 $\\mathcal{D}$가 어떤 low-regret distribution $\\mathcal{D'}$와도 $\\Theta(1)$만큼 떨어져 있음을 보인다($w = \\Omega(T)$이고 그러한 분포가 존재하더라도). 마지막으로, 우리는 unbounded context robustification 절차가 standard transformer architecture의 layers로 구현될 수 있음을 보이고, transformer model이 이러한 새로운 low-regret distribution을 효율적으로 표현하도록 학습될 수 있다는 실증적 증거를 제시한다."},{"id":"61046","en":"Rational Transductors","ko":"합리적 트랜스덕터","authors":"Mehryar Mohri","pos":"#4500","link":"https://openreview.net/forum?id=uEZpyELNuB","abs":"Standard Transformers excel at semantic modeling but struggle with rigid sequential logic and state tracking. Theoretical work establishes that self-attention is limited to $\\AC^0$ (under hard attention) or $\\TC^0$ (under soft attention), complexity classes that often fail to support robust length generalization on sequential problems without intermediate chain-of-thought \\citep{hahn2020theoretical, merrill2022saturated}. In this work, we introduce \\emph{Rational Transductors}, a dual-stream architecture that augments the Transformer with a matrix-valued recurrence derived from Weighted Finite Automata (WFA). By injecting rational state information into the attention mechanism via a \\emph{Deep Rational Injection} scheme, our framework strictly generalizes Transformers to capture all Regular Languages, $\\NC^1$-complete problems (such as Boolean Formula Evaluation), and fundamental separations like Parity and Modular Counting, while preserving $O(\\log T)$ parallel training efficiency. Theoretical analysis and empirical results demonstrate that Rational Transductors solve the \"Regular Gap,\" enabling robust length generalization on algorithmic tasks where standard Transformers fail, without the sequential computational bottlenecks of traditional RNNs.","absKo":"표준 Transformer는 semantic modeling에는 뛰어나지만, 엄격한 sequential logic과 state tracking에서는 어려움을 겪는다. 이론 연구는 self-attention이 hard attention 하에서는 $\\AC^0$, soft attention 하에서는 $\\TC^0$에 한정됨을 보였으며, 이러한 complexity class는 intermediate chain-of-thought 없이는 sequential 문제에서 강건한 length generalization을 지원하지 못하는 경우가 많다 \\citep{hahn2020theoretical, merrill2022saturated}. 본 연구에서는 Transformer를 Weighted Finite Automata(WFA)에서 유도된 matrix-valued recurrence로 확장하는 dual-stream architecture인 \\emph{Rational Transductors}를 제안한다. \\emph{Deep Rational Injection} scheme을 통해 rational state information을 attention mechanism에 주입함으로써, 우리의 framework는 Transformers를 엄밀히 일반화하여 모든 Regular Language, $\\NC^1$-complete 문제(예: Boolean Formula Evaluation), 그리고 Parity 및 Modular Counting 같은 근본적 분리를 포착할 수 있으며, 동시에 $O(\\log T)$ parallel training efficiency를 유지한다. 이론 분석과 실험 결과는 Rational Transductors가 \"Regular Gap\"을 해결하여, 표준 Transformers가 실패하는 algorithmic task에서 강건한 length generalization을 가능하게 하며, 전통적인 RNN의 순차적 계산 병목도 피함을 보여준다."},{"id":"64788","en":"Provably Data-driven Multiple Hyper-parameter Tuning with Structured Loss Function","ko":"구조화된 손실 함수를 활용한 증명 가능한 데이터 기반 다중 하이퍼파라미터 튜닝","authors":"QUOC TUNG LE, Anh Nguyen, Viet Anh Nguyen","pos":"#4501","link":"https://openreview.net/forum?id=JnuwpwbZ8D","abs":"

Data-driven algorithm design automates hyperparameter tuning, but its statistical foundations remain limited because model performance can depend on hyperparameters in implicit and highly non-smooth ways. Existing guarantees focus on the simple case of a one-dimensional (scalar) hyperparameter. This leaves the practically important, multi-dimensional hyperparameter tuning setting unresolved. We address this open question by establishing the first general framework for establishing generalization guarantees for tuning multi-dimensional hyperparameters in data-driven settings. Our approach strengthens the generalization guarantee framework for semi-algebraic function classes by exploiting tools from real algebraic geometry, yielding sharper, more broadly applicable guarantees. We then extend the analysis to hyperparameter tuning using the validation loss under minimal assumptions, and derive improved bounds when additional structure is available. Finally, we demonstrate the scope of the framework with new learnability results, including data-driven weighted group lasso and weighted fused lasso.

","absKo":"

Data-driven algorithm design은 hyperparameter tuning을 자동화하지만, model performance가 hyperparameter에 대해 암묵적이고 매우 비매끄러운 방식으로 의존할 수 있기 때문에 그 통계적 기반은 여전히 제한적이다. 기존의 보장은 1차원(scalar) hyperparameter라는 단순한 경우에 초점을 맞춘다. 이로 인해 실제로 중요한 다차원 hyperparameter tuning 설정은 여전히 해결되지 않은 상태로 남아 있다. 우리는 data-driven setting에서 다차원 hyperparameter tuning에 대한 generalization guarantee를 확립하는 첫 번째 일반 프레임워크를 제시함으로써 이 열린 문제를 해결한다. 우리의 접근은 real algebraic geometry의 도구를 활용하여 semi-algebraic function class에 대한 generalization guarantee 프레임워크를 강화하고, 더 날카롭고 더 폭넓게 적용 가능한 보장을 도출한다. 이어서 최소한의 가정 하에서 validation loss를 사용한 hyperparameter tuning으로 분석을 확장하고, 추가적인 구조가 있을 때 개선된 bound를 유도한다. 마지막으로 data-driven weighted group lasso와 weighted fused lasso를 포함한 새로운 learnability 결과로 이 프레임워크의 적용 범위를 입증한다.

"},{"id":"61213","en":"Language Generation with Replay: A Learning-Theoretic View of Model Collapse","ko":"Replay를 활용한 언어 생성: Model Collapse에 대한 학습 이론적 관점","authors":"Giorgio Racca, Michal Valko, Amartya Sanyal","pos":"#4507","link":"https://openreview.net/forum?id=scnRgI2hhX","abs":"

As scaling laws push the training of frontier large language models (LLMs) toward ever larger data requirements, training pipelines are approaching a regime where much of the publicly available online text may be consumed. At the same time, widespread LLM usage increases the volume of machine-generated content on the web; together, these trends raise the likelihood of generated content re-entering future training corpora, increasing the associated risk of model collapse. In practice, model developers address this concern through data cleaning, watermarking, synthetic-data policies, or by simply ignoring the issue. However, the problem of model collapse in generative models has not been examined from a learning-theoretic perspective: we study it through the theoretical lens of the language generation in the limit framework, introducing a replay adversary that augments the example stream with the generator's own past outputs. Our main contribution is a fine-grained learning-theoretic characterization of when replay fundamentally limits generation: while replay is benign for the strongest notion of uniform generation, it provably creates separations for the weaker notions of non-uniform generation and generation in the limit. Interestingly, our positive results mirror heuristics widely used in practice, such as data cleaning, watermarking, and output filtering, while our separations show when these ideas can fail.

","absKo":"

scaling law가 frontier large language models (LLMs)의 training을 점점 더 큰 data requirement로 밀어붙이면서, training pipeline은 공개적으로 이용 가능한 online text의 상당 부분이 소진되는 regime에 가까워지고 있다. 동시에 LLM 사용의 확산은 web상 machine-generated content의 양을 증가시키며, 이 두 흐름은 생성된 content가 미래의 training corpus로 다시 유입될 가능성을 높여, 이와 관련된 model collapse의 위험을 키운다. 실제로 model developer들은 data cleaning, watermarking, synthetic-data policy, 또는 단순히 이 문제를 무시하는 방식으로 이 우려를 다룬다. 그러나 generative model에서의 model collapse 문제는 learning-theoretic 관점에서 검토된 바가 없다. 우리는 이를 language generation in the limit framework의 이론적 관점에서 연구하며, example stream에 generator의 과거 output을 추가하는 replay adversary를 도입한다. 우리의 주요 기여는 replay가 언제 generation을 근본적으로 제한하는지에 대한 세밀한(fine-grained) learning-theoretic characterization이다. replay는 uniform generation의 가장 강한 개념에서는 무해하지만, non-uniform generation과 generation in the limit의 더 약한 개념에서는 분리를 명확히 만들어냄을 증명한다. 흥미롭게도, 우리의 positive result는 data cleaning, watermarking, output filtering과 같이 실제로 널리 사용되는 heuristic을 반영하며, 우리의 separation은 이러한 아이디어가 언제 실패할 수 있는지 보여준다.

"},{"id":"63336","en":"Feature Bagging Provides Stability","ko":"특징 배깅은 안정성을 제공한다","authors":"Yuheng Ma, Qiang Sun","pos":"#4513","link":"https://openreview.net/forum?id=XztZIs3k61","abs":"

We study the stability properties of feature bagging, an ensemble technique that improves robustness by training each learner on a randomly selected subset of features. We introduce feature stability (FS), a notion that quantifies the sensitivity of an algorithm’s output to the removal of a single feature. This notion complements classical instance stability (IS) and together provides a more comprehensive framework for evaluating algorithmic stability. Within this framework, we analyze feature bagging in both a parametric linear model and an assumption-free setting inspired by recursive feature subsampling in random forests. In both cases, we establish formal stability guarantees showing that feature bagging strictly outperforms its non-bagged counterpart, with larger gains achieved at smaller subsampling ratios, and that only a modest number of bagging rounds is sufficient to attain near-optimal stability.

","absKo":"

우리는 feature bagging의 안정성 특성을 연구한다. feature bagging은 각 learner를 무작위로 선택된 feature 부분집합 위에서 학습시켜 robustness를 향상하는 ensemble 기법이다. 우리는 feature stability(FS)를 도입하는데, 이는 단일 feature 하나를 제거했을 때 알고리즘 출력이 얼마나 민감하게 변하는지를 정량화하는 개념이다. 이 개념은 고전적인 instance stability(IS)를 보완하며, 함께 알고리즘 안정성을 평가하기 위한 더 포괄적인 framework를 제공한다. 이 framework 내에서 우리는 parametric linear model과 random forests에서의 recursive feature subsampling에 영감을 받은 assumption-free setting 모두에서 feature bagging을 분석한다. 두 경우 모두에서 우리는 feature bagging이 비-bagged 대응 방법보다 엄밀하게 더 우수함을 보이는 formal stability guarantee를 확립하며, subsampling ratio가 더 작을수록 더 큰 향상이 나타나고, 최적에 가까운 안정성을 달성하는 데에는 적은 수의 bagging round만으로도 충분함을 보인다.

"},{"id":"61253","en":"Coverage Improvement and Fast Convergence of On-policy Preference Learning","ko":"On-Policy Preference Learning의 Coverage 개선과 빠른 수렴","authors":"Juno Kim, Jihun Yun, Jason Lee, Kwang-Sung Jun","pos":"#4514","link":"https://openreview.net/forum?id=s7wV3xGQFL","abs":"

On-policy preference learning algorithms for language model alignment such as online direct policy optimization (DPO) can significantly outperform their offline counterparts. We provide a theoretical explanation for this phenomenon by analyzing how the sampling policy's coverage evolves throughout on-policy training. We propose and rigorously justify the \\emph{coverage improvement principle}: with sufficient batch size, each update moves into a region around the target where coverage is uniformly better, making subsequent data increasingly informative and enabling rapid convergence. In the contextual bandit setting with Bradley-Terry preferences and linear softmax policy class, we show that on-policy DPO converges exponentially in the number of iterations for batch size exceeding a generalized coverage threshold. In contrast, any learner restricted to offline samples from the initial policy suffers a slower minimax rate, leading to a sharp separation in total sample complexity. Motivated by this analysis, we further propose a simple hybrid sampler based on a novel \\emph{preferential} G-optimal design, which removes dependence on coverage and guarantees convergence in just two rounds. Finally, we develop principled on-policy schemes for reward distillation in the general function class setting, and show faster noiseless rates under an alternative deviation-based notion of coverage.

","absKo":"

online direct policy optimization (DPO)과 같은 language model alignment를 위한 on-policy preference learning algorithm은 offline counterpart보다 상당히 뛰어난 성능을 보일 수 있다. 우리는 sampling policy의 coverage가 on-policy training 전반에 걸쳐 어떻게 진화하는지를 분석함으로써 이 현상에 대한 이론적 설명을 제시한다. 우리는 \\emph{coverage improvement principle}을 제안하고 엄밀히 정당화한다. 충분한 batch size가 주어지면, 각 update는 target 주변의 영역으로 이동하여 coverage가 일관되게 더 좋아지며, 이후 data는 점점 더 informative해져 빠른 수렴을 가능하게 한다. Bradley-Terry preference와 linear softmax policy class를 갖는 contextual bandit setting에서, on-policy DPO가 generalized coverage threshold를 초과하는 batch size에 대해 iteration 수에 대해 지수적으로 수렴함을 보인다. 반면, 초기 policy에서 나온 offline sample로 제한된 어떤 learner도 더 느린 minimax rate를 겪게 되며, 총 sample complexity에서 뚜렷한 차이가 발생한다. 이러한 분석에 동기를 얻어, 우리는 coverage에 대한 의존성을 제거하고 단 두 라운드 만에 수렴을 보장하는 새로운 \\emph{preferential} G-optimal design에 기반한 간단한 hybrid sampler를 추가로 제안한다. 마지막으로, 일반 function class setting에서 reward distillation을 위한 원리 기반 on-policy scheme을 개발하고, coverage의 대안적 deviation-based 개념 아래 더 빠른 noiseless rate를 보인다.

"},{"id":"64696","en":"Combinatorial Sparse PCA Beyond the Spiked Identity Model","ko":"Spiked Identity Model을 넘어선 조합적 Sparse PCA","authors":"Peiyuan Zhang, Syamantak Kumar, Kevin Tian, Purnamrita Sarkar","pos":"#4515","link":"https://openreview.net/forum?id=Kk5UZgkWFx","abs":"Sparse PCA is one of the most well-studied problems in high-dimensional statistics. In this problem, we are given samples from a distribution with covariance $\\\\mathbf{\\Sigma}$, whose top eigenvector $\\\\mathbf{v} \\in \\\\mathbb{R}^d$ is $s$-sparse. Existing sparse PCA algorithms can be broadly categorized into (1) combinatorial algorithms (e.g., diagonal or elementwise covariance thresholding) and (2) SDP-based algorithms. While combinatorial algorithms are much simpler, they are typically only analyzed under the spiked identity model (where $\\\\mathbf{\\Sigma} \\propto \\\\mathbf{I}_d + \\gamma \\\\mathbf{v}\\\\mathbf{v}^\\top$ for some $\\gamma > 0$), whereas SDP-based algorithms require no additional assumptions on $\\\\mathbf{\\Sigma}$. We demonstrate explicit counterexample covariances $\\\\mathbf{\\Sigma}$ against the success of standard combinatorial algorithms for sparse PCA, when moving beyond the spiked identity model. In light of this discrepancy, we give the first combinatorial method for sparse PCA that provably succeeds for general $\\\\mathbf{\\Sigma}$ using $\\\\mathsf{poly}(s, \\log(d))$ samples and $d^2 \\cdot \\\\mathsf{poly}(s, \\log(d))$ time, by providing a global convergence guarantee on the truncated power method of Yuan and Zhang (JMLR, 2013). We provide a natural generalization of our method to recovering $k$ sparse principal components. Finally, we evaluate our method on synthetic and real-world sparse PCA datasets.","absKo":"Sparse PCA는 고차원 통계에서 가장 많이 연구된 문제 중 하나이다. 이 문제에서는 covariance가 $\\\\mathbf{\\Sigma}$인 분포에서 sample을 받으며, 그 top eigenvector $\\\\mathbf{v} \\in \\\\mathbb{R}^d$는 $s$-sparse이다. 기존 sparse PCA algorithm은 크게 (1) combinatorial algorithm(예: diagonal 또는 elementwise covariance thresholding)과 (2) SDP-based algorithm으로 분류할 수 있다. Combinatorial algorithm은 훨씬 단순하지만, 일반적으로는 spiked identity model(즉, 어떤 $\\\\gamma > 0$에 대해 $\\\\mathbf{\\Sigma} \\propto \\\\mathbf{I}_d + \\gamma \\\\mathbf{v}\\\\mathbf{v}^\\top$인 경우)에서만 분석되는 반면, SDP-based algorithm은 $\\\\mathbf{\\Sigma}$에 대한 추가 가정을 요구하지 않는다. 우리는 spiked identity model을 넘어설 때 sparse PCA에 대한 표준 combinatorial algorithm의 성공을 반박하는 명시적 counterexample covariance $\\\\mathbf{\\Sigma}$를 제시한다. 이러한 불일치를 고려하여, 우리는 Yuan과 Zhang (JMLR, 2013)의 truncated power method에 대해 global convergence guarantee를 제공함으로써, 일반적인 $\\\\mathbf{\\Sigma}$에 대해 provably 성공하고 $\\\\mathsf{poly}(s, \\log(d))$ sample과 $d^2 \\cdot \\\\mathsf{poly}(s, \\log(d))$ time을 사용하는 sparse PCA용 최초의 combinatorial method를 제시한다. 또한 우리는 이 방법을 $k$개의 sparse principal component를 복원하는 문제로 자연스럽게 일반화한다. 마지막으로 synthetic 및 실제 sparse PCA dataset에서 우리의 method를 평가한다."},{"id":"64677","en":"On the Learning Dynamics of RLVR at the Edge of Competence","ko":"역량의 경계에서 RLVR의 학습 동역학에 대하여","authors":"Yu Huang, Zixin Wen, Yuejie Chi, Yuting Wei, Aarti Singh, Yingbin LIANG, Yuxin Chen","pos":"#4608","link":"https://openreview.net/forum?id=KxYCE98u1d","abs":"

Reinforcement Learning with Verifiable Rewards (RLVR) has been a main driver of recent breakthroughs in large reasoning models. Yet it remains a mystery how rewards based solely on final outcomes can help overcome the long-horizon barrier to extended reasoning. To understand this, we develop a theory of the training dynamics of RL for transformers on compositional reasoning tasks. Our theory characterizes how the effectiveness of RLVR is governed by the smoothness of the difficulty spectrum. When data contains abrupt discontinuities in difficulty, learning undergoes grokking-type phase transitions, giving rise to prolonged plateaus before progress recurs. In contrast, a smooth difficulty spectrum leads to a relay effect: stable gradient signals on easier problems elevate the model's capabilities to the point where harder ones become tractable, resulting in steady and continuous learning. Our theory explains how RLVR can improve performance at the edge of competence, and suggests that appropriately designed data mixtures can yield scalable gains. As a technical contribution, our analysis develops and adapts tools from Fourier analysis on finite groups to our setting. We validate the predicted mechanisms empirically via synthetic experiments.

","absKo":"

Reinforcement Learning with Verifiable Rewards (RLVR)는 최근 large reasoning model에서의 돌파구를 이끈 주요 동력이었다. 그러나 최종 결과에만 기반한 reward가 어떻게 extended reasoning의 long-horizon barrier를 극복하는 데 도움을 줄 수 있는지는 여전히 미스터리로 남아 있다. 이를 이해하기 위해 우리는 compositional reasoning task에서 transformer에 대한 RL의 training dynamics 이론을 개발한다. 우리의 이론은 RLVR의 효과가 difficulty spectrum의 smoothness에 의해 어떻게 좌우되는지를 특징짓는다. 데이터에 difficulty의 급격한 불연속이 존재하면, 학습은 grokking-type phase transition을 겪으며, 진전이 재개되기 전까지 긴 plateau가 나타난다. 반대로, smooth difficulty spectrum은 relay effect를 유도한다. 즉, 더 쉬운 문제에서의 안정적인 gradient signal이 모델의 능력을 어려운 문제도 다룰 수 있는 수준까지 끌어올려, 꾸준하고 연속적인 학습을 가능하게 한다. 우리의 이론은 RLVR이 어떻게 competence의 경계에서 성능을 향상시킬 수 있는지 설명하며, 적절히 설계된 data mixture가 확장 가능한 성능 향상을 가져올 수 있음을 시사한다. 기술적 기여로서, 우리의 분석은 finite group 위의 Fourier analysis 도구를 이 setting에 맞게 개발하고 적용한다. 우리는 synthetic experiment를 통해 예측된 메커니즘을 경험적으로 검증한다.

"},{"id":"65429","en":"Positive Distribution Shift as a Framework for Understanding Tractable Learning","ko":"다루기 쉬운 학습 이해를 위한 프레임워크로서의 긍정적 분포 이동","authors":"Marko Medvedev, Idan Attias, Elisabetta Cornacchia, Theodor Misiakiewicz, Gal Vardi, Nati Srebro","pos":"#4609","link":"https://openreview.net/forum?id=DkLQ40hTlt","abs":"

We study a setting where the goal is to learn a target function f(x) with respect to a target distribution D(x), but training is done on i.i.d. samples from a different training distribution D’(x), labeled by the true target f(x). Such a distribution shift (here in the form of covariate shift) is usually viewed negatively, as hurting or making learning harder, and the traditional distribution shift literature is mostly concerned with limiting or avoiding this negative effect. In contrast, we argue that with a well-chosen D'(x), the shift can be positive and make learning easier -- a perspective called Positive Distribution Shift (PDS). Such a perspective is central to contemporary machine learning, where much of the innovation is in finding good training distributions D’(x), rather than changing the training algorithm. We further argue that the benefit is often computational rather than statistical, and that PDS allows computationally hard problems to become tractable even using standard gradient-based training. We formalize different variants of PDS, show how certain hard classes are easily learnable under PDS, and make connections with membership query learning.

","absKo":"

우리는 목표 분포 D(x)에 대한 타깃 함수 f(x)를 학습하는 것이 목표이지만, 학습은 진짜 타깃 f(x)로 라벨링된 서로 다른 학습 분포 D’(x)에서 i.i.d. 샘플을 사용해 수행되는 설정을 연구한다. 이러한 분포 이동(여기서는 covariate shift의 형태)은 일반적으로 학습을 저해하거나 더 어렵게 만드는 부정적 현상으로 여겨지며, 전통적인 distribution shift 문헌은 대부분 이 부정적 효과를 제한하거나 회피하는 데 초점을 맞춰 왔다. 이에 반해 우리는 적절히 선택된 D'(x)를 사용하면 이 shift가 긍정적으로 작용하여 학습을 더 쉽게 만들 수 있다고 주장한다. 이러한 관점은 Positive Distribution Shift (PDS)라 불리며, 현대 machine learning에서 핵심적이다. 많은 혁신이 학습 알고리즘을 바꾸는 것이 아니라 좋은 학습 분포 D’(x)를 찾는 데 있기 때문이다. 우리는 더 나아가 이 이점이 종종 통계적이라기보다 계산적 성격을 띠며, PDS가 표준 gradient-based training만으로도 계산적으로 어려운 문제를 다룰 수 있게 한다고 주장한다. 우리는 PDS의 여러 변형을 정식화하고, 특정 어려운 클래스가 PDS 하에서 어떻게 쉽게 학습 가능한지 보이며, membership query learning과의 연결도 제시한다.

"},{"id":"62171","en":"DPO Unchained: Your Training Algorithm is Secretly Disentangled in Human Choice Theory (and Its Loss' Convexity is Dispensable)","ko":"DPO Unchained: 학습 알고리즘은 인간 선택 이론에서 본질적으로 disentangled됨(손실의 Convexity는 불필요)","authors":"Wenxuan Zhou, Shujian Zhang, brice magdalou, John Lambert, Ehsan Amid, Richard Nock, Andrew Hard","pos":"#4618","link":"https://openreview.net/forum?id=j4c3i3a5kH","abs":"

Normative theories allow one to elicit key parts of a ML algorithm from first principles, which is crucial at a time of championed scrutiny for ML work. Direct Preference Optimization (DPO) cleverly bypasses reward modeling by making an explicit link with a specific normative model of human choice. Our paper elevates this connection to the full generality of DPO's normative framework. Getting there requires reworking social choice theory's textbook path for a better RLHF/ML fit. It elevates the connection to a remarkably broad viewpoint on preference optimization, considering the current panorama of DPO follow-ups. It also unveils unexpected riches for ML, chief among which the support for non-convex losses, the fact that any compliant ML analytical choice can be embedded with any human choice model, and a normative framework's umbrella wide enough to safeguard DPO's extensions (margins, length correction, ...). A toy experiment ``far away'' from the DPO crowd is given.

","absKo":"

Normative theory는 원리 수준에서 ML algorithm의 핵심 부분을 끌어낼 수 있게 해 주며, 이는 ML 연구에 대한 championed scrutiny가 강화되는 지금 특히 중요하다. Direct Preference Optimization (DPO)는 인간 선택의 특정 normative model과의 명시적 연결을 통해 reward modeling을 영리하게 우회한다. 우리의 논문은 이 연결을 DPO의 normative framework가 가지는 완전한 일반성으로 끌어올린다. 이를 위해서는 RLHF/ML에 더 잘 맞는 형태로 social choice theory의 교과서적 전개를 다시 구성해야 한다. 그 결과, 현재 DPO follow-up의 전반적 panorama를 고려하는, preference optimization에 대한 놀라울 정도로 폭넓은 관점을 제시한다. 또한 ML에 대해 예상치 못한 풍부함도 드러내는데, 그중 핵심은 non-convex loss를 지원한다는 점, 어떠한 compliant ML analytical choice도 어떠한 human choice model과도 embedding될 수 있다는 점, 그리고 DPO의 extensions (margins, length correction, ...)을 보호할 만큼 충분히 넓은 normative framework의 우산이다. DPO 계열과는 ``멀리 떨어진'' toy experiment도 제시한다.

"},{"id":"66615","en":"A Fine-Grained Understanding of Uniform Convergence for Halfspaces","ko":"Halfspace에 대한 균등 수렴의 세밀한 이해","authors":"Aryeh Kontorovich, Kasper Green Larsen","pos":"#4619","link":"https://openreview.net/forum?id=28nj9x9hLa","abs":"We study the fine-grainded uniform convergence behavior of halfspaces beyond worst-case VC bounds. For inhomogeneous halfspaces in $\\mathbb{R}^d$ with $d\\ge 2$, we show that standard first-order VC bounds are essentially tight: even consistent hypotheses can incur population error $\\Theta(d\\log(n/d)/n)$, and in the agnostic setting the deviation scales as $\\sqrt{\\tau\\log(1/\\tau)}$ at true error $\\tau$. In contrast, homogeneous halfspaces in $\\mathbb{R}^2$ exhibit a markedly different behavior. In the realizable case, every hypothesis consistent with the sample has error $O(1/n)$. In the agnostic case, we prove a bandwise, log-free deviation bound on each dyadic risk band via a critical-wedge localization argument. Unioning over bands incurs only a $\\log\\log n$ overhead, and we establish a matching lower bound showing this overhead is unavoidable. Together, these results give a fine-grained and nearly complete picture of uniform convergence for halfspaces, revealing sharp dimensional and structural thresholds.","absKo":"우리는 최악의 경우 VC 경계를 넘어서는 halfspaces의 미세한 uniform convergence 거동을 연구한다. $\\mathbb{R}^d$에서 $d\\ge 2$인 비동차(inhomogeneous) halfspaces에 대해, 표준적인 1차 VC 경계가 본질적으로 타이트함을 보인다. 즉, 일관적인 가설조차 population error가 $\\Theta(d\\log(n/d)/n)$에 이를 수 있으며, agnostic 설정에서는 실제 오류가 $\\tau$일 때 deviation이 $\\sqrt{\\tau\\log(1/\\tau)}$로 스케일한다. 반면 $\\mathbb{R}^2$의 동차(homogeneous) halfspaces는 현저히 다른 거동을 보인다. realizable 경우, 샘플과 일치하는 모든 가설의 error는 $O(1/n)$이다. agnostic 경우에는 critical-wedge localization 논증을 통해 각 dyadic risk band에서 log-free deviation bound를 증명한다. bands에 대해 union bound를 적용하면 $\\log\\log n$의 추가 비용만 발생하며, 이 overhead가 불가피함을 보이는 matching lower bound도 제시한다. 종합하면, 이 결과들은 halfspaces에 대한 uniform convergence의 미세한 관점을 거의 완전하게 제공하며, 날카로운 차원적 및 구조적 임계점을 드러낸다."},{"id":"63659","en":"A Random Matrix Theory of Masked Self-Supervised Learning","ko":"마스크된 자기지도 학습의 랜덤 행렬 이론","authors":"Arie Zurich, Bruno Loureiro, Federica Gerace, Yue Lu","pos":"#4620","link":"https://openreview.net/forum?id=UuH4mmAGPJ","abs":"

In the era of transformer models, masked self-supervised learning (SSL) has become a foundational training paradigm. A defining feature of masked SSL is that training aggregates predictions across many masking patterns, giving rise to a joint, matrix-valued predictor rather than a single vector-valued estimator. This object encodes how coordinates condition on one another and poses new analytical challenges. We develop a precise high-dimensional analysis of masked modeling objectives in the proportional regime where the number of samples scales with the ambient dimension. Our results provide explicit expressions for the generalization error and characterize the spectral structure of the learned predictor, revealing how masked modeling extracts structure from data. For spiked covariance models, we show that the joint predictor undergoes a Baik–Ben Arous–Péché (BBP)-type phase transition, identifying when masked SSL begins to recover latent signals. Finally, we identify structured regimes in which masked self-supervised learning provably outperforms PCA, highlighting potential advantages of SSL objectives over classical unsupervised methods.

","absKo":"Transformer model 시대에 masked self-supervised learning(SSL)은 근본적인 training paradigm이 되었다. masked SSL의 정의적 특징은 training이 여러 masking pattern에 걸친 예측을 aggregate하여, 단일 vector-valued estimator가 아니라 joint한 matrix-valued predictor를 만들어낸다는 점이다. 이 객체는 coordinate가 서로를 어떻게 조건화하는지 인코딩하며 새로운 분석적 도전을 제기한다. 우리는 샘플 수가 ambient dimension에 비례해 증가하는 proportional regime에서 masked modeling objective에 대한 정밀한 high-dimensional analysis를 전개한다. 우리의 결과는 generalization error에 대한 명시적 표현을 제공하고 학습된 predictor의 spectral structure를 특징짓는다. 이를 통해 masked modeling이 데이터로부터 구조를 추출하는 방식을 드러낸다. Spiked covariance model에 대해서는 joint predictor가 Baik–Ben Arous–Péché(BBP)형 phase transition을 겪음을 보이며, masked SSL이 언제 latent signal을 회복하기 시작하는지를 식별한다. 마지막으로, 구조화된 regime에서 masked self-supervised learning이 PCA를 엄밀하게 능가함을 밝혀, classical unsupervised method에 비해 SSL objective가 가질 수 있는 잠재적 이점을 강조한다.

"},{"id":"64963","en":"A Refined Generalization Analysis for Extreme Multi-class Supervised Contrastive Representation Learning","ko":"극단적 다중클래스 지도 Contrastive Representation Learning을 위한 정교화된 일반화 분석","authors":"Minh Hieu Nong, Antoine Ledent","pos":"#4621","link":"https://openreview.net/forum?id=IL6lhb2r1e","abs":"Contrastive Representation Learning (CRL) has achieved strong empirical success in multiple machine learning disciplines, yet its theoretical sample complexity remains poorly understood. Existing analyses usually assume that input tuples are identically and independently distributed, an assumption violated in most practical settings where contrastive tuples are constructed from a finite pool of labeled data, inducing dependencies among tuples. While one recent work analyzed this learning setting using U-Statistics to estimate the population risk, the techniques used therein require the risk of each class to concentrate uniformly, making excess risk bounds scale in the order of $\\rho_{\\min}^{-{1}/{2}}$ where $\\rho_{\\min}$ denotes the probability of the rarest class. Such a dependency can be overly pessimistic in the extreme multiclass settings where there are many tail classes which contribute minimally to the overall population risk. Our contributions are two-fold. Firstly, we improve upon the previous work and prove a bound with a sample complexity of the same order as the number of classes $R$, regardless of the distribution over classes. Furthermore, we formulate a different estimator that captures the concentration of the risk \\textit{across classes}, enabling sharper bounds in extreme multi-class learning scenarios, especially where class distributions are long-tailed. Under mild assumptions on the class distributions, the resulting sample complexity is $\\mathcal{{O}}(k)$ where $k$ is the number of samples per tuple.","absKo":"Contrastive Representation Learning (CRL)은 여러 machine learning 분야에서 강한 경험적 성과를 거두었지만, 그 theoretical sample complexity는 여전히 잘 이해되지 않고 있다. 기존 분석은 보통 입력 tuple이 independent and identically distributed라고 가정하는데, 이는 대부분의 실제 설정에서 위배된다. 실제로 contrastive tuple은 유한한 labeled data pool로부터 구성되며, 이로 인해 tuple 간 dependency가 유도된다. 최근 한 연구는 population risk를 추정하기 위해 U-Statistics를 사용하여 이 learning setting을 분석했지만, 그 기술은 각 class의 risk가 uniform하게 concentration된다는 점을 요구한다. 이로 인해 excess risk bound가 가장 희귀한 class의 확률인 $\\rho_{\\min}$에 대해 $\\rho_{\\min}^{-{1}/{2}}$ 차수로 스케일되는데, 이러한 의존성은 전체 population risk에 거의 기여하지 않는 tail class가 많은 극단적인 multiclass setting에서는 지나치게 비관적일 수 있다. 우리의 기여는 두 가지이다. 첫째, 우리는 이전 연구를 개선하여 class 분포와 무관하게 class 수 $R$의 차수와 같은 sample complexity를 갖는 bound를 증명한다. 더 나아가, 우리는 \\textit{across classes}에서의 risk concentration을 포착하는 다른 estimator를 정식화하여, 특히 class 분포가 long-tailed인 극단적 multi-class learning scenario에서 더 날카로운 bound를 가능하게 한다. class 분포에 대한 완화된 가정 하에서, 그 결과의 sample complexity는 $\\mathcal{{O}}(k)$이며, 여기서 $k$는 tuple당 sample 수이다."},{"id":"65788","en":"A Theoretical Framework for Statistical Evaluability of Generative Models","ko":"생성 모델의 통계적 평가 가능성을 위한 이론적 프레임워크","authors":"Shashaank Aiyer, Yishay Mansour, Shay Moran, Han Shao","pos":"#4622","link":"https://openreview.net/forum?id=A8AxU1GUUl","abs":"

Statistical evaluation aims to estimate the generalization performance of a model using held-out i.i.d. test data sampled from the ground-truth distribution. In supervised learning settings such as classification, performance metrics such as error rate are well-defined, and test error reliably approximates population error given sufficiently large datasets. In contrast, evaluation is more challenging for generative models due to their open-ended nature: it is unclear which metrics are appropriate and whether such metrics can be reliably evaluated from finite samples. In this work, we introduce a theoretical framework for evaluating language models and establish evaluability results for commonly used metrics. We study two categories of metrics: test-based metrics, including integral probability metrics (IPMs), and similarity-based metrics, including Rényi and KL divergences. We show that IPMs with respect to any bounded test class can be evaluated from finite samples up to multiplicative and additive approximation errors. Moreover, when the test class has finite fat-shattering dimension, IPMs can be evaluated with arbitrary precision. In contrast, similarity-based metrics, including Rényi and KL divergences, are not evaluable from finite samples, as their values can be critically determined by rare events. We also analyze the potential and limitations of perplexity as an evaluation method.

","absKo":"

통계적 평가는 ground-truth distribution에서 샘플링한 held-out i.i.d. test data를 사용해 모델의 generalization performance를 추정하는 것을 목표로 한다. classification과 같은 supervised learning setting에서는 error rate와 같은 performance metric이 잘 정의되어 있으며, 충분히 큰 dataset이 주어지면 test error는 population error를 신뢰성 있게 근사한다. 반면, generative model은 open-ended한 특성 때문에 평가가 더 어렵다. 어떤 metric이 적절한지, 그리고 그러한 metric을 유한한 sample로 신뢰성 있게 평가할 수 있는지가 명확하지 않다. 본 연구에서는 language model을 평가하기 위한 이론적 framework를 도입하고, 널리 사용되는 metric들에 대한 evaluability 결과를 확립한다. 우리는 두 부류의 metric을 연구한다: integral probability metrics (IPMs)를 포함한 test-based metric과 Rényi 및 KL divergence를 포함한 similarity-based metric이다. 우리는 임의의 bounded test class에 대한 IPM은 multiplicative 및 additive approximation error까지 유한한 sample로 평가할 수 있음을 보인다. 더 나아가 test class가 finite fat-shattering dimension을 가지면, IPM은 임의의 정밀도로 평가 가능하다. 반면 Rényi 및 KL divergence를 포함한 similarity-based metric은 드문 사건에 의해 값이 결정될 수 있으므로 유한한 sample로는 평가할 수 없다. 또한 우리는 평가 방법으로서 perplexity의 가능성과 한계를 분석한다.

"},{"id":"65984","en":"Beyond Binary: Continuous State Optimization with Graph-Structured Objectives","ko":"이진을 넘어서: 그래프 구조화 목표를 활용한 연속 상태 최적화","authors":"Corinna Cortes, Yishay Mansour, Mehryar Mohri","pos":"#4623","link":"https://openreview.net/forum?id=8CGNXWeK4U","abs":"

Large-scale learning systems often face the challenge of balancing multiple, potentially competing objectives, such as fairness, accuracy, and latency. While recent work has formalized this as an optimization problem over binary states, many real-world control parameters—such as fairness thresholds, diversity mixing rates, or resource budgets—are continuous. In this work, we extend the framework to continuous state spaces. We model the problem as minimizing a sum of linear objectives subject to movement costs that penalize system instability. We capture the local structure of the objectives using a dependency graph (or factor graph), where each objective is determined by a subset of the state attributes. To address the tension between exploration and stability, we propose Lazy Graph-LinUCB, an algorithm that performs lazy updates to minimize switching costs while maintaining near-optimal regret. Beyond stability, we introduce three advanced mechanisms to exploit the underlying graph structure: (1) an asynchronous update schedule that eliminates synchronization overhead in sparse graphs; (2) an adaptive algorithm that learns the graph structure from data; and (3) a joint estimator that leverages data sharing among correlated objectives to significantly tighten regret bounds. Empirically, we demonstrate that these structural exploitations reduce movement costs by more than a factor of three in heterogeneous systems while maintaining similar cumulative losses.

","absKo":"

대규모 학습 시스템은 공정성, 정확도, 지연시간과 같이 서로 잠재적으로 경쟁하는 여러 목표 사이의 균형을 맞추는 과제에 자주 직면한다. 최근 연구는 이를 binary state 위의 최적화 문제로 formalize했지만, 공정성 임계값, 다양성 mixing rate, 자원 예산과 같은 많은 실제 제어 매개변수는 연속적이다. 본 연구에서는 이 프레임워크를 연속 상태 공간으로 확장한다. 우리는 시스템 불안정성을 패널티하는 이동 비용을 제약으로 하여 선형 objective의 합을 최소화하는 문제로 모델링한다. 각 objective가 state attribute의 부분집합에 의해 결정되는 dependency graph(또는 factor graph)를 사용하여 objective의 국소 구조를 포착한다. 탐색과 안정성 사이의 긴장을 다루기 위해, 우리는 switching cost를 최소화하면서 near-optimal regret를 유지하는 lazy update를 수행하는 알고리즘 Lazy Graph-LinUCB를 제안한다. 안정성을 넘어, 우리는 underlying graph structure를 활용하기 위한 세 가지 고급 메커니즘을 도입한다: (1) sparse graph에서 synchronization overhead를 제거하는 asynchronous update schedule, (2) 데이터로부터 graph structure를 학습하는 adaptive algorithm, (3) 상관된 objective들 사이의 data sharing을 활용하여 regret bound를 크게 조이는 joint estimator. 실험적으로, 이러한 구조 활용은 이질적 시스템에서 누적 손실은 유사하게 유지하면서 이동 비용을 3배 이상 줄임을 보여준다.

"}];window.dispatchEvent(new CustomEvent("papers-loaded",{detail:"68689"}));