# FrontierCS 2 tasks

Generated by `python3 scripts/task_metadata.py table` from `tasks/metadata/`.
Resources are the limits the platform enforces for each run: CPU cores / memory GiB /
workspace GiB, GPU type and total GPU-hours, the Agent's answering time, the number
of train (development) evaluations and the final evaluation timeout.

| Task | Category | Paper | Agent CPU / mem / disk | Agent GPU | Verifier CPU / mem | Verifier GPU | Agent time | Train calls | Eval timeout |
| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
| `clusterwise_spgemm`<br>Fixed-Length Clustered Sparse Multiplication | Architecture | [Improving SpGEMM Performance Through Matrix-Reordering and Cluster-wise Computation](https://arxiv.org/pdf/2507.21253v1) (SC 2025) | 2 / 16 / 32 | - | 60 / 64 | - | 5 h | 3 | 120 min |
| `kami_fp64_block_gemm`<br>KAMI FP64 Block GEMM — trusted batched harness adaptation | Architecture | [KAMI](https://www.ssslab.cn/assets/papers/2025-wang-KAMI.pdf) (SC 2025) | 3 / 4 / 4 | 1x H100, 0.25 GPU-h | 3 / 4 | 1x H100, 0.5 GPU-h | 5 h | 12 | 30 min |
| `runlts_predictor`<br>Register-aware Branch Prediction | Architecture | [RUNLTS](https://ericrotenberg.wordpress.ncsu.edu/files/2025/06/cbp2025-final44-Koizumi.pdf) (CBP 2025) | 1 / 4 / 40 | - | 8 / 32 | - | 5 h | 100 | 120 min |
| `edlib_exact_alignment`<br>Exact Sequence Alignment | Bioinformatics | [Edlib](https://doi.org/10.1093/bioinformatics/btw753) (Bioinformatics 2017) | 2 / 4 / 5 | - | 2 / 4 | - | 5 h | 20 | 10 min |
| `ckks_composable_noise`<br>CKKS Precision Propagation at 2,048 Slots | Cryptography | [Accurate and Composable Noise Estimates for CKKS with Application to Exact HE Computation](https://eprint.iacr.org/2024/853.pdf) (IACR CiC 2025) | 2 / 2 / 8 | - | 2 / 4 | - | 5 h | 100 | 10 min |
| `quarter_pir`<br>Private Set List Decoding | Cryptography | [Efficient Pre-processing PIR Without Public-Key Cryptography](https://eprint.iacr.org/2023/1574) (EUROCRYPT 2024) | 4 / 8 / 8 | - | 4 / 8 | - | 5 h | 8 | 30 min |
| `lwe_structured_recovery`<br>Structured-LWE Public Witness Recovery | Cryptography | [A Systematic Study of Sparse LWE](https://eprint.iacr.org/2024/1589) (CRYPTO 2024) | 8 / 32 / 32 | - | 8 / 2 | - | 5 h | 100 | 2 min |
| `approximate_mips_query_stopping`<br>Approximate-MIPS Query Stopping | Database | [FARGO](https://www.vldb.org/pvldb/vol16/p1100-zheng.pdf) (PVLDB 2023) | 4 / 4 / 4 | - | 4 / 4 | - | 5 h | 100 | 15 min |
| `batch_local_counting`<br>Batch Local Counting | Database | [Fast Local Subgraph Counting](https://www.vldb.org/pvldb/vol17/p1967-li.pdf) (PVLDB 2024) | 1 / 4 / 4 | - | 1 / 4 | - | 5 h | 100 | 30 min |
| `bounded_evm_path_prediction`<br>Bounded EVM Path Prediction | Database | [Seer](https://www.vldb.org/pvldb/vol18/p822-xiao.pdf) (PVLDB 2024) | 8 / 4 / 4 | - | 8 / 8 | - | 5 h | 50 | 30 min |
| `evm_state_checkpoints`<br>EVM State Checkpoints | Database | [Spectrum](https://www.vldb.org/pvldb/vol17/p2541-zhang.pdf) (PVLDB 2024) | 4 / 8 / 8 | - | 4 / 8 | - | 5 h | 100 | 60 min |
| `approximate_mips_global_probing`<br>FARGO paper-aligned query core | Database | [FARGO](https://www.vldb.org/pvldb/vol16/p1100-zheng.pdf) (PVLDB 2023) | 4 / 4 / 4 | - | 4 / 4 | - | 5 h | 50 | 15 min |
| `graph_mining_method_selection`<br>Graph Mining Method Selection | Database | [Accurate and Fast Approximate Graph Pattern Mining at Scale](https://www.vldb.org/pvldb/vol18/p93-chen.pdf) (PVLDB 2024) | 8 / 16 / 4 | - | 8 / 16 | - | 5 h | 100 | 15 min |
| `join_tree_rewriting`<br>Join-Tree Rewriting | Database | [Instance-Optimal Acyclic Join Processing Without Regret](https://www.vldb.org/pvldb/vol18/p2413-vansummeren.pdf) (PVLDB 2025) | 2 / 8 / 1 | - | 2 / 8 | - | 5 h | 100 | 15 min |
| `persistent_treap_update_engine`<br>Persistent Treap Update Engine | Database | [Efficient Concurrent Updates to Persistent Randomized Binary Search Trees](https://www.vldb.org/pvldb/vol18/p1481-wang.pdf) (PVLDB 2025) | 8 / 16 / 8 | - | 8 / 16 | - | 5 h | 100 | 30 min |
| `residual_state_advancement`<br>Residual State Advancement | Database | [JoinBoost](https://www.vldb.org/pvldb/vol16/p3071-huang.pdf) (PVLDB 2023) | 2 / 16 / 4 | - | 2 / 16 | - | 5 h | 100 | 120 min |
| `parallel_dbscan`<br>Exact DBSCAN Connectivity | HPC | [Theoretically-Efficient and Practical Parallel DBSCAN](https://doi.org/10.1145/3318464.3380582) (SIGMOD 2020) | 8 / 4 / 2 | - | 8 / 4 | - | 5 h | 100 | 10 min |
| `hypergraph_core_numbers`<br>Hypergraph Core Numbers | HPC | [Practical Parallel Hypergraph Algorithms](https://people.csail.mit.edu/jshun/hygra.pdf) (PPoPP 2020) | 8 / 4 / 2 | - | 8 / 4 | - | 5 h | 100 | 30 min |
| `parallel_radix_sort`<br>In-place Integer Redistribution | HPC | [Theoretically-Efficient and Practical Parallel In-Place Radix Sorting](https://people.csail.mit.edu/jshun/RegionsSort.pdf) (SPAA 2019) | 4 / 2 / 2 | - | 4 / 2 | - | 5 h | 100 | 15 min |
| `parbutterfly`<br>ParButterfly Exact Counting | HPC | [Parallel Algorithms for Butterfly Computations](https://jshun.csail.mit.edu/butterfly.pdf) (APOCS 2020) | 4 / 8 / 4 | - | 4 / 8 | - | 5 h | 100 | 30 min |
| `knn_gpu_kernel_optimization`<br>GPU Brute-Force k-NN Kernel Optimization | MLSys | [FlashLib](https://flashml-org.github.io/) (Blog) | 4 / 8 / 16 | 1x H100, 3 GPU-h | 3 / 8 | 1x H100, 4.5 GPU-h | 5 h | 25 | 90 min |
| `dbscan_gpu_kernel_optimization`<br>GPU DBSCAN Kernel Optimization | MLSys | [FlashLib](https://flashml-org.github.io/) (Blog) | 4 / 8 / 16 | 1x H100, 3 GPU-h | 4 / 8 | 1x H100, 3.67 GPU-h | 5 h | 25 | 40 min |
| `ivf_pq_gpu_kernel_optimization`<br>GPU IVF-PQ Search Kernel Optimization | MLSys | [FlashLib](https://flashml-org.github.io/) (Blog) | 4 / 8 / 16 | 1x H100, 3 GPU-h | 3 / 8 | 1x H100, 7 GPU-h | 5 h | 25 | 240 min |
| `kmeans_gpu_kernel_optimization`<br>GPU K-Means Kernel Optimization | MLSys | [Flash-KMeans](https://arxiv.org/abs/2603.09229) (arXiv 2026) | 4 / 8 / 16 | 1x H100, 3 GPU-h | 3 / 8 | 1x H100, 5 GPU-h | 5 h | 25 | 30 min |
| `vllm_llm_serving_optimization`<br>GPU vLLM Serving Optimization | MLSys | [Continuum](https://arxiv.org/pdf/2511.02230) (arXiv 2025) | 8 / 32 / 64 | 1x H100, 5 GPU-h | 8 / 32 | 4x H100, 192 GPU-h | 5 h | 25 | 1800 min |
| `sparse_video_attention_policy`<br>Sparse Video Attention Policy | MLSys | [SVG-EAR](https://arxiv.org/abs/2603.08982) (ECCV 2026) | 3 / 12 / 10 | 1x H100, 15 GPU-h | 3 / 12 | 4x H100, 364 GPU-h | 5 h | 25 | 210 min |
| `representation_frechet_loss`<br>Differentiable Population Statistics | Machine Learning | [Representation Fréchet Loss for Visual Generation](https://arxiv.org/abs/2604.28190) (arXiv 2026) | 3 / 12 / 10 | 1x A100-80GB, 2 GPU-h | 3 / 12 | 1x A100-80GB, 9 GPU-h | 5 h | 3 | 180 min |
| `nano_world_model`<br>Future Video Sampling | Machine Learning | [Nano World Models](https://arxiv.org/abs/2605.23993) (arXiv 2026) | 3 / 12 / 10 | 1x A10G, 0.5 GPU-h | 3 / 12 | 1x A10G, 5.67 GPU-h | 5 h | 3 | 170 min |
| `fast_kv_compaction`<br>KV Cache Compaction | Machine Learning | [Fast KV Compaction via Attention Matching](https://arxiv.org/pdf/2602.16284) (ICML 2026) | 4 / 16 / 10 | 1x H100, 5 GPU-h | 3 / 12 | 8x H100, 24 GPU-h | 5 h | 8 | 90 min |
| `lejepa_world_model`<br>Latent Objective Recovery | Machine Learning | [When Does LeJEPA Learn a World Model?](https://arxiv.org/abs/2605.26379) (arXiv 2026) | 2 / 12 / 4 | - | 2 / 2 | - | 5 h | 8 | 6.67 min |
| `dynamic_tanh`<br>Learned Transformer Layer Mapping | Machine Learning | [Transformers without Normalization](https://arxiv.org/abs/2503.10622) (CVPR 2025) | 3 / 12 / 10 | 1x L4, 0.5 GPU-h | 3 / 12 | 1x L4, 1 GPU-h | 5 h | 3 | 15 min |
| `ragen_multi_turn_rl`<br>Multi-turn Trajectory Batches | Machine Learning | [RAGEN](https://arxiv.org/abs/2504.20073) (arXiv 2025) | 3 / 12 / 10 | 1x A100-80GB, 0.5 GPU-h | 3 / 12 | 1x A100-80GB, 2 GPU-h | 5 h | 3 | 30 min |
| `flow_opd`<br>OCR Distillation Trajectory Sampling | Machine Learning | [Flow-OPD](https://arxiv.org/abs/2605.08063) (arXiv 2026) | 3 / 12 / 10 | 1x H100, 0.83 GPU-h | 3 / 12 | 1x H100, 4.17 GPU-h | 5 h | 4 | 50 min |
| `polar_express`<br>Polar-Factor Approximation | Machine Learning | [The Polar Express](https://arxiv.org/abs/2505.16932) (ICLR 2026) | 2 / 4 / 4 | - | 2 / 4 | - | 5 h | 100 | 30 min |
| `detok_latent_denoising`<br>Robust Image Tokenizer | Machine Learning | [Latent Denoising Makes Good Tokenizers](https://arxiv.org/abs/2507.15856) (ICLR 2026) | 3 / 12 / 10 | 1x L4, 1.5 GPU-h | 3 / 12 | 1x L4, 1.5 GPU-h | 5 h | 6 | 90 min |
| `bbq_packet_scheduling`<br>BBQ bitmap traversal | Network | [BBQ](https://www.usenix.org/system/files/nsdi24-atre.pdf) (NSDI 2024) | 2 / 4 / 16 | - | 2 / 4 | - | 5 h | 100 | 30 min |
| `shockwave_dynamic_gpu_allocation`<br>Dynamic GPU Allocation | Network | [Shockwave](https://arxiv.org/pdf/2210.00093) (NSDI 2023) | 4 / 4 / 4 | - | 4 / 4 | - | 5 h | 3 | 30 min |
| `metaopt_combined_te_gap_search`<br>MetaOpt Combined-TE Gap Search | Network | [Finding Adversarial Inputs for Heuristics using Multi-level Optimization](https://www.usenix.org/system/files/nsdi24-namyar-finding.pdf) (NSDI 2024) | 4 / 8 / 8 | - | 4 / 8 | - | 5 h | 100 | 60 min |
| `glenside`<br>Access-Pattern Rewriting | Programming Languages | [Pure Tensor Program Rewriting via Access Patterns (Representation Pearl)](https://arxiv.org/abs/2105.09377) (MAPS 2021) | 2 / 4 / 8 | - | 2 / 4 | - | 5 h | 100 | 5 min |
| `egraph_invariant_restoration`<br>E-Graph Invariant Restoration | Programming Languages | [egg](https://arxiv.org/pdf/2004.03082) (POPL 2021) | 8 / 4 / 4 | - | 8 / 4 | - | 5 h | 100 | 30 min |
| `babble_anti_unification`<br>Reusable Library Candidates | Programming Languages | [babble](https://arxiv.org/abs/2212.04596) (POPL 2023) | 2 / 8 / 4 | - | 2 / 2 | - | 5 h | 100 | 10 min |
| `polygon`<br>SQL Counterexample Search | Programming Languages | [Polygon](https://arxiv.org/pdf/2504.06542v1) (PLDI 2025) | 2 / 8 / 4 | - | 2 / 8 | - | 5 h | 100 | 5.5 min |
| `diffusion_policy_cpu`<br>Closed-loop action sampling | Robotics | [Diffusion Policy](https://arxiv.org/abs/2303.04137) (RSS 2023) | 2 / 4 / 4 | - | 2 / 4 | - | 5 h | 100 | 62 min |
| `mujoco_playground_cpu`<br>Go1 Joystick Policy Training | Robotics | [Demonstrating MuJoCo Playground](https://arxiv.org/abs/2502.08844) (RSS 2025) | 3 / 12 / 10 | 1x H100, 4 GPU-h | 3 / 12 | 1x H100, 8.5 GPU-h | 5 h | 3 | 240 min |
| `lrr`<br>Localized Regex Repair | Security | [Repairing Regex Vulnerabilities via Localization-Guided Instructions](https://aclanthology.org/2026.eacl-long.379.pdf) (EACL 2026) | 4 / 12 / 8 | - | 4 / 12 | 1x L4, 2 GPU-h | 5 h | 100 | 60 min |
| `progent_privilege_control`<br>Privilege Control for Tool-Using LLM Agents | Security | [Progent](https://arxiv.org/pdf/2504.11703) (arXiv 2025) | 4 / 8 / 4 | - | 8 / 16 | - | 5 h | 5 | 180 min |
| `cute_crest`<br>Search Strategy for Concolic Testing | Software Engineering | [Heuristics for Scalable Dynamic Test Generation](https://www2.eecs.berkeley.edu/Pubs/TechRpts/2008/EECS-2008-123.pdf) (ASE 2008) | 8 / 16 / 8 | - | 8 / 32 | - | 5 h | 100 | 180 min |
| `binary_fuse_filter`<br>Binary Fuse Filter Construction and Lookup | Theory | [Binary Fuse Filters](https://arxiv.org/abs/2201.01174) (ACM JEA 2022) | 2 / 4 / 4 | - | 2 / 4 | - | 5 h | 100 | 3 min |

48 tasks: 32 CPU-only, 16 using Modal GPUs.
The Agent's model calls go through the Candidate's provider route.
`progent_privilege_control` also uses a task model: the site must serve the LiteLLM alias `progent-agent-gpt-4o-mini` (`features.task_model_aliases`), with a total key budget of $65 per run.