# research-expert-team — Master Paper Studies (Oct 15, 2026)

**Scope**: Deep-study corpus for the `research-expert-team` skill — per-direction full-text dissections of the field's highest-cited exemplar papers, distilled into reusable submission playbooks, plus a two-layer methodology program (Wave 1: 6 papers / 3 directions; Wave 2: +9 papers / +2 directions + Agent expansion). Wave 1 triggered by user request: "每个方向各找 1-2 篇顶刊引用量最高的核心论文，认真研究学习……沉淀到相应论文团队中".

## Selection & Verification Pipeline

1. Candidate pool per direction (18 papers) → **Semantic Scholar batch API** verified citation counts + venues (2026-10-15, script `.temp/cite_batch.py`)
2. User confirmed selection (2 per direction, combination-coverage rationale):

| Direction | Paper | Venue | Citations |
|---|---|---|---|
| Ontology & KR | Kimera | IJRR 2021 (journal) | 407 |
| Ontology & KR | ConceptGraphs | ICRA 2024 | 671 |
| Super Agents | ReAct | ICLR 2023 | **12,540** |
| Super Agents | MetaGPT | ICLR 2024 Oral | 2,626 |
| AIGC Content Creation | DreamFusion | ICLR 2023 | **3,929** |
| AIGC Content Creation | DreamGaussian | ICLR 2024 | 1,098 |

(Rejected alternates logged for future rotations: ViperGPT 908, AutoGen 2,918, Voyager 2,622 TMLR, SWE-agent 2,163 NeurIPS, LRM 1,032, Shap-E 520.)

3. **Full-text deep read** of all 6 papers via ar5iv (abstract / intro / contributions / method / full experiment tables / ablations / limitations / appendices)

## Files Added (skill structure: SKILL.md + references/)

- `skills/research-expert-team/references/mp-ontology.md` — Kimera + ConceptGraphs six-dimension dissections (选题·论述·论文骨架·实验·价值·打法) + direction playbook
- `skills/research-expert-team/references/mp-super-agents.md` — ReAct + MetaGPT dissections + direction playbook
- `skills/research-expert-team/references/mp-aigc.md` — DreamFusion + DreamGaussian dissections + direction playbook
- `skills/research-expert-team/SKILL.md` — new §2.4 "Master Paper Studies" routing table (load-by-direction before R1/R2 and R6); R1 and R6 playbooks now reference §2.4 patterns; related-skills list updated

## Distilled Cross-Cutting Lessons (embedded in the files)

- Three proven topic-selection patterns: **combination gap** (ReAct), **obstacle-bypass + problem reformulation** (DreamFusion), **ecosystem bridge / human-practice transfer** (ConceptGraphs, MetaGPT); pain-point-with-mechanism for representation swaps (DreamGaussian)
- Experiment design gems catalogued: same-trajectory-subtraction controlled ablations (ReAct), self-built stronger baselines (ConceptGraphs), anti-cheating metrics + oracle anchors (DreamFusion), speed-quality paired tables + per-stage ablation rows (DreamGaussian), module-level + embedded real-time (Kimera), multi-dimensional cost/productivity matrices (MetaGPT)
- **Paper-skeleton tables**: each exemplar ships a section-by-section structural template (venue-specific: IJRR journal arc vs. ICRA capability-matrix vs. ICLR minimal-paradigm) writers can mirror directly
- Each direction file ends with a submission playbook mapping lessons onto the library's own baselines (SKILL.md §2 selection tables) and the Comparison Fairness Protocol

## Methodology Core Refactoring (same day, follow-up)

**Trigger**: user follow-up instruction — "不要被这几篇论文限制，聚焦这个领域方向，学习他们的方式，变成这领域的强有力的技能". The master-paper studies were reading notes; the skill needed an **executable methodology** instead.

**Architecture change**: two-layer asset, enforced division of labor —

- `references/methodology-core.md` (NEW, the program): topic-selection engine (§A: modes M1-M4 + red-flag checklist + delta typology 5 levels) · argumentation architecture (§B: skeleton templates T1-T3 + contribution–evidence bijection + related-work organization + limitations style) · experiment-design pattern library (§C: C1-C9, all benchmark-proven) · reviewer-expectation library (§D: D1-D8 must-ask questions with preempt actions) · value amplifiers (§E: 6 items) · team operating conventions (§F: mode-number tagging for traceability). Every procedure cites its source paper and is domain-tuned to 3DGS/spatial-intelligence, not bound to the six exemplars.
- `references/mp-*.md` (3 files, repositioned as evidence): headers now declare "案例证据库" status pointing to methodology-core; no methodology duplication.

**SKILL.md wiring**: §2.4 rewritten as the two-layer load-by-need/load-by-direction router; §3 playbooks now bind R1→§A, R2→§A.3 (+M/delta tagging), R3→§D (+per-item D1-D8 coverage), R6→§B, R8/R9→§C (+pattern numbering and domain iron rule); related-skills list updated.

## Wave 2: Three More Directions + Agent Expansion (same day, follow-up)

**Trigger**: user request — “按照同样方式也补充 3DGS、空间智能、Agent 等相关领域方向的，都选择领域影响力最高的 2-3 篇论文进行分析沉淀”。

**Selection & verification**: 27-paper candidate pool → Semantic Scholar batch API (2026-10-15, script `.temp/cite_batch2.py`); NeRF citation fixed via title-match endpoint (5,900) with OpenAlex cross-check (5,113 LNCS + 524 arXiv, split versions). User confirmed 3 per direction (alternates logged: Mip-NeRF 2,979, VGGT 1,995, SpatialVLM 1,114, Mip-Splatting 971, SplaTAM 749, LangSplat 588, GS-SLAM 571, Reflexion 6,044, Tree of Thoughts 5,130, AutoGen 2,918, SWE-agent 2,164, DreamerV3 1,504):

| Direction | Paper | Venue | Citations |
|---|---|---|---|
| 3DGS | 3DGS (Kerbl et al.) | TOG/SIGGRAPH 2023 | **10,887** |
| 3DGS | 4DGS (Wu et al.) | CVPR 2024 | 1,676 |
| 3DGS | SuGaR | CVPR 2024 | 982 |
| Spatial Intelligence | NeRF | ECCV 2020 | 5,900 |
| Spatial Intelligence | Instant-NGP | TOG/SIGGRAPH 2022 | 6,456 |
| Spatial Intelligence | DUSt3R | CVPR 2024 | 2,022 |
| Super Agents (+3) | Toolformer | NeurIPS 2023 | 6,146 |
| Super Agents (+3) | Generative Agents | UIST 2023 | 5,987 |
| Super Agents (+3) | Voyager | TMLR | 2,622 |

**Full-text deep reads**: all 9 papers via ar5iv (three-element structure / math-equivalence anchors / Pareto sweeps / data bootstrapping / interview protocol / skill-library design, among others).

## Files Changed (Wave 2)

- NEW `references/mp-3dgs.md` — 3DGS/4DGS/SuGaR six-dimension dissections + direction playbook (three proven routes: representation-revolution / extension-paradigm / downstream-gap)
- NEW `references/mp-spatial.md` — NeRF/Instant-NGP/DUSt3R dissections + playbook (routes: representation-founding / efficiency-mechanism / pipeline-replacement)
- EXPANDED `references/mp-super-agents.md` (2→5 exemplars) — Toolformer (paradox opening + data bootstrapping with loss-diff filtering + disabled-tier baseline), Generative Agents (memory architecture + interview evaluation protocol + 40-year agenda revival), Voyager (skill library + percentage-quantified ablations + cross-method component evidence); direction playbook upgraded from 2 to 5 proven routes
- UPDATED `references/methodology-core.md` (program layer, now distilled from 15 exemplars) — new §A.1 opening-sentence variants (clash / paradox / agenda-revival / pipeline-dissection / legacy-gap); new §B.5 representation-paper techniques (math anchor, bounded first-claim, ideal-case loss derivation, default-tool refutation, free-extraction lists, counter-intuitive preempt, cross-community primer); §C expanded C1→C16 (dual operating points, mechanism-tiered ablations, disabled-tier baselines, Pareto sweeps with default recommendations, single-model zero-finetune multi-task protocol, evaluation-hygiene clause set, cross-method component evidence) + C4/C5/C8 sources enriched; §D preempt actions and sources expanded for all 8 questions; §E +3 amplifiers (dataset-as-community-standard, future-work-as-topic-interface, naming-chain extensibility)
- UPDATED `SKILL.md` — §2.4 Layer 2 evidence table now 5 directions / 15 exemplar papers; Layer 1 program rows updated (A.1 variants, B.5, C1-C16, nine amplifiers); R8 playbook references C1-C16; related-skills entry updated

## Wave 3: 3DGS Paper Writing Blueprint (same day, follow-up)

**Trigger**: user supplied a 3DGS-specific paper-writing methodology (motivation-upgrade rewriting to avoid "tuning-paper" feel · absolute symbol unification · painpoint-targeted resource-metric evaluation · five-section structure) and requested it be learned, distilled into a drafting skill, deposited into the existing asset system, and **self-consistently fused** ("总结形成一种撰写skill沉淀到已有的，并进行自洽融合").

**Fusion design** (division of labor enforced to avoid duplication):

- NEW `references/3dgs-paper-blueprint.md` — the 3DGS domain extension of the program layer: §1 three writing rules (motivation upgrade with weak→strong rewriting table + template sentence · symbol unification delegating to cg-paper-writing's `symbols-terminology.md` · painpoint→resource-metric mapping with a 3DGS inherent-defect library) · §2 five-section placement table (intro 4-paragraph arc with the mandatory "maintains real-time rendering" contribution sentence · related-work 3-way moat · method total-then-split with the 3DGS math preliminaries G/Σ'/α-blending · experiment 4-table spec incl. zoom-in artifact-annotated qualitative figures · conclusion+limitation) · §3 3DGS limitations library (honest-limitation sentence skeletons) · §4 cross-asset mapping table (every procedure mapped to methodology-core codes / cg-paper-writing files / mp case libraries, with an explicit conflict-resolution rule) · §5 pre-submission checklist.
- UPDATED `references/methodology-core.md` — §B.5 +2 generalizable techniques (B.5-8 motivation-upgrade program; B.5-9 symbol-system unification, delegating execution to cg-paper-writing's symbol table); NEW §B.6 domain-skeleton router (core stays general, blueprint holds the 3DGS skeleton); C4 renamed 速度-质量-资源配对表 with the five-resource caliber made explicit (PSNR/SSIM/LPIPS + FPS + VRAM + MB + Training time; quality-only tables = D5 rejection).
- UPDATED `SKILL.md` — §2.4 Layer 1 gains a domain-extension row routing 3DGS paper drafting to the blueprint; R6 playbook appends the blueprint walk; §6 related-skills entry and frontmatter description updated with 3DGS-drafting trigger phrases.
- UPDATED `references/mp-3dgs.md` — direction playbook now points to the blueprint for execution-level drafting (evidence stays in mp, program in blueprint).

**Self-consistency checks**: no symbol table duplicated (delegated to cg-paper-writing); no methodology duplicated (blueprint §4 maps every procedure to core codes); cg-paper-writing boundary preserved (it keeps section templates/venue formats/symbol table; blueprint owns the 3DGS skeleton program and domain iron rules).

## Wave 4: Expert-Review Audit (same day, follow-up)

**Trigger**: user requested an expert-researcher + top-journal-reviewer audit of the generated md files for factual errors or inaccurate descriptions.

**Method**: re-fetched the full texts of all 9 Wave-2 exemplar papers + DreamGaussian (ar5iv) and re-verified every numeric/mechanism claim line-by-line against the sources; searched award claims; audited cross-file consistency (SKILL.md routing vs file inventory).

**Findings & fixes** (all applied):

- **A-class (factual)**: ① INGP-vs-NGLOD IoU numbers were cross-scene spliced (0.9911 is ours on one scene, 0.9812 is NGLOD on the other; each side wins one scene) — rewritten to per-scene pairs with the paper's own "similar numeric quality" framing; ② NeRF "best paper honorable mention" could not be verified from any authoritative source — removed per the skill's own no-unverified-claims red line (oral status retained); ③ 3DGS §7.3 has 5 ablation discussion subsections, not 3 — fixed; ④ 3DGS Table 1 is six metric columns (3 quality + 3 resource), not five — fixed in mp-3dgs + methodology-core C4; ⑤ 4DGS Neu3D pivot misstated: Im4D renders ~5 FPS with 93MB storage (ours 30 FPS / 90MB — storage is a tie, not an advantage) — rewritten to an FPS + multi-camera-caliber pivot.
- **B-class (imprecision)**: NeRF positional-encoding frequency range corrected to 2⁰…2^(L−1); first-claim quote restored with the "high-resolution" / "real objects and scenes" qualifiers; "(INGP、 probing)" typography fixed; Toolformer scaling-law chain now lists all four GPT-2 sizes; blueprint defect-library rows remapped (Mip-Splatting → aliasing/multi-scale, not background sparsity; SuGaR moved to a dedicated surface-alignment row; textureless → curvature/surface-aware regularization).
- **C-class (internal inconsistency)**: SKILL.md still said "three frontier directions" in four places while §2.4 ships five — description / scope / §2 header / related-skills updated to distinguish the three KB-category selection tables from the five master-paper directions.
- **Verified correct (no change needed)**: ~60 numeric claims — incl. 3DGS (48h / 6m25s / 160FPS / 41m33s / 134FPS / 33.32-vs-33.09 / λ=0.2 / α-reset-3000 / clone-split volume phrasing / Eq.2-Eq.3 equivalence); SuGaR (Erratum-Eq.8 / BakedSDF-48h / NeRFMeshing-8×V100 / licensing-7-of-9 / R-SuGaR-2K-7K-15K / λ=0.1-0.3-0.5 / binding params); 4DGS (82FPS / 30FPS@1352×1014 / O(t·N)-vs-O(N+ℱ) / warm-up-3000 / ablations 27.05-26.67-33.08-33.02-31.91-vs-34.05); NeRF (LPIPS 0.212-vs-0.250 / 5MB-15GB-3000× / 100-300k-iters-V100 / 9-row ablation / L=5-10-15); INGP (ACORN 36.9h-38.59dB-2.5min / ~10×-CUDA + 10-100×-MLP split / L2-6MB-at-2^19 / F=2-L=16 / ε=10⁻¹⁵ / skip-zero-gradient-10%); Toolformer (53.5-vs-39.8 / disabled-14.8-vs-9.6 / w_t=max(0,1−0.2t) / LAMA-QA dual-disable / 98.1%-0.2% attribution / 775M emergence); DUSt3R (8 datasets / 8.5M pairs / CroCo init / 224→512 / confidence loss / global alignment convergence); DreamGaussian (2min-0.738 / 45s-0.594 / 27s-0.591 / 20min-0.647 / user study 4.31-3.92); award claims INGP SIGGRAPH-2022-best-paper and VGGT CVPR-2025-best-paper (both search-verified); Voyager and Generative Agents sections fully verified.

## Wave 5: Latest-Paper Review & Caliber Unification (same day, follow-up)

**Trigger**: user request — "再去查阅涉及到的方向的最新论文，并进行更新，保证准确以及整个项目数据口径一致".

**Method**: citation-count recheck (S2 throttled → OpenAlex cross-reference; database-caliber difference confirmed as expected, S2-2026-10 declarations stand) + five-direction 2025-2026 progress searches + targeted re-verification of time-sensitive claims (VGGT abstract re-fetched from arXiv).

**Updates applied**:

- `mp-spatial.md` — ① NeRF word-root count "+500 papers" (stale) → "数以千计的衍生工作" (with 2026 quantum-NeRF / surgical-NeRF examples still emerging); ② VGGT caliber aligned to the paper's own wording "hundreds of views" (was "100+ 视角", conservative but imprecise); ③ DUSt3R lineage extended with DGGT (CVPR 2026, dynamic pose-free reconstruction) — search-verified lineage continuation.
- `mp-3dgs.md` §4 — candidate-gap list annotated with 2026-10 search evidence: industry-recognized three deployment bottlenecks (sparse views / artifacts / edge compute), physical-interactivity (RAF, CVPR 2026), GS-SLAM real-time fusion, large-scale aerial (STARS-GS), NeRF↔GS cross-representation distillation (RouteBridge) — original gaps confirmed still valid and aligned with industry pain points.
- `mp-super-agents.md` §6 — 2026-10 battlefield note: Deep Research Agent / long-horizon systems as the research center of gravity; evaluation-side gaps (IDRBench, BrowseComp-Plus, ACL 2026 best-resource-paper expert-gap) tied back to M1 combination-gap judgment.
- `mp-aigc.md` §3 — 2026-10 landscape note: large-3D-generative-model route established (TRELLIS SLAT multi-format incl. 3D Gaussians; Hunyuan3D-2; Oxford SynCity 3000 city-scale world generation) — "per-sample SDS optimization" paradigm superseded by feed-forward large-model generation; comparison baseline landscape changed accordingly.
- `mp-ontology.md` §3 — 2026-10 direction-evolution note: open-vocabulary 3D scene graphs matured to TPAMI-journal level (Hyperbolic hierarchical 3D, TPAMI 2026); active axes = hierarchical / dynamic-environment (LOST-3DSG) / monocular feed-forward (LEXI-SG) / functional scene graphs — OntoGS-type submissions should differentiate on these axes.
- `methodology-core.md` E-9 — naming chain extended with DGGT.
- **Caliber unification**: all five mp-*.md footnotes now uniformly "Semantic Scholar batch API 2026-10-15 核实" (was a 10-14/10-15 mix); SKILL.md Layer 2 Voyager venue unified to "TMLR 2024" (matching the mp table); no citation-count values changed (same-day recheck; OpenAlex runs a different aggregation caliber by design, which is why the declared source stays S2).

## Verification (cumulative, all waves)

- All 15 papers' facts drawn from full-text reads; citation counts from Semantic Scholar batch API (single request each wave, 2026-10-15) + NeRF title-match endpoint with OpenAlex cross-check
- Wave 4 audit: 10 papers re-fetched and every numeric claim line-verified (2026-10-15); all A/B/C-class findings fixed in the same pass
- Wave 5 review: five-direction 2025-2026 progress searches; VGGT abstract re-fetched; stale claims updated; citation-date and venue calipers unified across all files
- Dual-copy sync to `~/.workbuddy/skills/research-expert-team/` — SKILL.md + 7 references (incl. 3dgs-paper-blueprint.md) MD5-identical
- TeleAgent skill directory (`~/.config/TeleAgent/users/.../skills/research-expert-team/`) refreshed in the same pass — local install stays in sync with the project tree
- No carrier-count changes (955/26/18 untouched — this is skill-content-only increment)

## State

Uncommitted in working tree — user commits manually, per established workflow.