git:20261019.c9f7986 to v2.0.0

99 added, 97 removed. Audit A to A.

---

name: research-expert-team

- description: "Full-cycle research expert team for 3DGS/spatial-intelligence papers: idea evaluation, innovation mining, simulated peer review, academic polishing, AIGC-trace hygiene (publication hygiene), framework & venue compliance, strict venue-format writing, experiment design, and experiment analysis. Ships with method-selection tables for the Ontology & Knowledge Representation, Super Agents & Agentic AI, and AIGC Content Creation KB categories, master-paper studies for five directions (adding 3D Gaussian Splatting and Spatial Intelligence), plus a 3DGS paper-drafting blueprint (five-section skeleton, motivation-upgrade rewriting, painpoint→resource-metric evaluation mapping). Use when: evaluating a research idea, finding novelty/创新点, simulating reviewers 审稿, polishing 润色, drafting or structuring a 3DGS paper / 3DGS论文撰写, removing AI-generated traces 消除AI生成痕迹 (metadata/style hygiene), checking paper-framework logic against venue expectations, formatting strictly to a journal/conference template, designing or analyzing experiments, 科研专家团."

+ description: 'Evidence-led research assistance for 3DGS, spatial intelligence and agent papers: evaluate an idea,

+ locate related work, plan experiments, interpret results, review a draft, improve prose and check venue requirements.

+ Use for 科研专家团、选题评估、创新点分析、实验设计、结果分析、模拟审稿、论文润色和投稿检查。'

+ license: Apache-2.0

+ metadata:

+ version: 2.0.0

+ reviewed-at: '2026-10-21'

---

- # Research Expert Team — 科研全流程专家团

- > Scope: the full research cycle for a 3DGS / spatial-intelligence paper — from idea evaluation to rebuttal — orchestrated as nine specialist roles, with knowledge-backed method-selection tables and five-direction master-paper studies (§2.4).

- > Knowledge base: 976 arXiv-verified methods in `data/methods.json` (single source of truth — always re-query, never answer from memory), including the 12-method **Ontology & Knowledge Representation**, 20-method **Super Agents & Agentic AI**, and 23-method **AIGC Content Creation** categories.

- ## 1. The Nine Roles (assign by stage)

- |---|------|-------------|--------|----------|

- | R1 | **Idea Evaluator** | a raw idea / "is this worth doing?" | 4-axis verdict: novelty / feasibility / impact / cost, each backed by ≥3 library methods as prior art | → R2 |

- | R2 | **Innovation Miner** | idea approved, need contribution claims | 2–4 ranked contribution statements, each mapped to a measurable delta vs. a named library baseline | → R8 |

- | R4 | **Academic Polisher** | prose-level pass | language, terminology consistency, paragraph rhythm, claim-evidence alignment edits (tracked, never silent rewrites of technical claims) | → R5 |

- **Pipeline**: R1 → R2 → R8 → (write) → R6 → R7 → R4 → R5 → R3 → (revise) → R9 → (rebuttal). Any role can be invoked standalone.

- ## 2. Method-Selection Tables — KB Categories & Direction Coverage

- The tables below cover the three KB categories built for these directions; the other two covered directions (3D Gaussian Splatting, Spatial Intelligence) pick baselines by re-querying the KB's Reconstruction / Surface / Dynamic / Generation categories — all five directions have master-paper evidence in §2.4. When the paper sits in (or borrows from) one of these directions, pick baselines, comparisons, and citations from the table — then re-query `data/methods.json` for the full category before locking the experiment plan.

- ### 2.1 Ontology & Knowledge Representation (12 methods)

+ # Research Expert Team — 以证据为依据的科研协作

- | Paper task | Anchor methods | Mechanism to cite/compare |

- |-----------|----------------|---------------------------|

- | Scene-graph construction over 3DGS | GraphWrit3R, DynamicGSG, Graph-GSReg, Prior-SG | end-to-end writing vs. dynamic memory vs. cross-scene registration |

- | Ontology-grounded planning | OntoPlan | shared symbolic vocabulary before planning |

- | Neuro-symbolic spatial reasoning | SATURN, NEUROSYMLAND, GRASP, Vision-TL-Action | soft predicates + symbolic executor; PDDL grounding; temporal-logic graphs |

- | KG-augmented QA | KLDrive, EmbodiedVSR | structured facts cut LLM hallucination |

+ Help the researcher make a defensible next decision. Role names organize checks;

+ they do not imply independent human experts, additional tools, or actual peer review.

+ Use only the roles needed for the current request. A short idea question does not

+ require a nine-stage report or a set of generated files.

- ### 2.2 Super Agents & Agentic AI (20 methods)

+ ## Start with the actual task

- | Paper task | Anchor methods | Mechanism to cite/compare |

- |-----------|----------------|---------------------------|

- | Agent architecture pattern | Octopus v2 (on-device), E-Ledger (safe harness), SwarmWorld (decentralized) | topology + policy layers |

- | 3D/spatial agent integration | COLMAR, GRAND-SLAM, GATOR, CADFather, DesignAgent3D, GASE, PhotoAgent, GN0 | cooperative reconstruction, SLAM, CAD, editing, VLN |

- | Paper-to-code / physical-feedback pipelines | SPLATIFY, Artisan | grammar-constrained synthesis; simulation gates |

- | Benchmarks & evaluation | GISAgentBench, WorldBench, E3VS-Bench, MeMoREval, SkillScriptBench | match the benchmark to the claim |

+ 1. Identify the research question, contribution type, available evidence, and current

+ stage. Use information already supplied; ask only for a missing fact that would

+ materially change the work. Proceed with explicit assumptions when possible.

+ 2. Query `data/methods.json` and `data/evidence-audit.json`. Read the snapshot date

+ and per-record evidence state. Counts come from data, never remembered prose.

+ 3. Search primary sources for the closest competing approaches and relevant new

+ work. The catalog is a starting index, not an exhaustive literature search.

+ 4. Separate observation, literature statement, hypothesis, proposed experiment,

+ measured result, and editorial suggestion throughout the answer.

- ### 2.3 AIGC Content Creation (23 methods)

+ ## Choose the relevant roles

- | Paper task | Anchor methods | Mechanism to cite/compare |

- |-----------|----------------|---------------------------|

- | Text → 3DGS scene | PSGS (panorama sliding), CGGS (ego-centric), Drag4D (motion control) | text-driven generation variants |

- | Generative scene reconstruction | Lyra, GaussVideoDreamer, UniWorld-View | video-diffusion self-distillation / inconsistency-aware GS |

- | Single-image → 3D asset | DiffSplat, Dragen3D, DD3G, GSV3D, GS-RGBN, DirectGaussian, DreamArt (articulated), T2Bs (character) | diffusion repurposing, distillation, RGBN volumes, surfels |

- | Text-to-3D optimization | CSD | coupled score distillation fixing SDS geometry drift |

- | Stylization & artistic creation | InStyle, OmniStyle-INR, SplatFont3D | appearance stylization, multimodal transfer, 3D fonts |

- | Avatar / 4D human generation | GaussianMotion, Flex4DHuman | pose-guided avatars; multi-view video diffusion |

- | Content composition & VR dynamics | LightHarmony3D, VirSqueezer | insertion harmonization; squeezing dynamics |

+ | Role | Question to answer | Expected evidence or output |

+ |---|---|---|

+ | R1 — Idea evaluator | What would this teach us, and what could make it fail? | Closest work, documented search boundary, assumptions, feasibility and unresolved risks |

+ | R2 — Contribution analyst | What precisely changes relative to the closest approach? | A bounded difference claim and a test that could refute it; no universal novelty ranking |

+ | R3 — Critical reviewer | Which conclusions does the draft support? | Located observations, consequences, uncertainty and feasible remedies; revise objections when evidence changes |

+ | R4 — Academic editor | Can a reader understand the argument? | Clearer prose with tracked substantive changes; no silent strengthening of claims |

+ | R5 — Publication integrity editor | Are authorship, disclosure and provenance accurate? | Consistent citations, anonymization and permitted source cleanup; preserve required AI-use disclosures |

+ | R6 — Argumentation reviewer | Do the claims and evidence form a coherent argument? | Claim–evidence map; structure chosen for this paper, not imposed from a successful exemplar |

+ | R7 — Venue checker | Which current official requirements apply? | Official URL, retrieval date and requirement-to-draft matrix; mark unconfirmed rules |

+ | R8 — Experiment planner | Which experiment distinguishes the explanations? | Controls, independent experimental unit, metrics, splits, repeats, budget, failure criteria and analysis plan |

+ | R9 — Result analyst | What do the measurements support, and what remains open? | Traceable run artifacts, effect sizes, uncertainty, alternative explanations and limitations |

- **Boundary note for classification**: generic 3D generative models → `Generation`; editing pipelines → `Editing`; watermark/copyright protection → `Security`. AIGC Content Creation = the creation-side application layer (text/image → 3DGS content, stylization, generative assets).

+ Suggested sequence when applicable: scope → related-work check → contribution

+ and test design → measurements → interpretation → drafting → review → revision.

+ Writing-only tasks can start at R4; existing results can start at R9.

- ### 2.4 Methodology Core + Master-Paper Evidence Base（双层加载）

+ ## Evidence ledger

- Two-layer asset: **`methodology-core.md` is the program** — the fifteen highest-cited exemplars distilled into executable operating procedures, domain-tuned to 3DGS/spatial-intelligence (not bound to those papers); **`mp-*.md` are the evidence** — per-direction full-text dissections tracing each procedure to its source paper. Load the program layer by need, the evidence layer to trace a provenance:

+ For each consequential claim, record:

- **Layer 1 — Program (methodology-core.md, load by role)**:

+ | Field | Content |

+ |---|---|

+ | Claim and type | Exact claim; literature fact, interpretation, hypothesis, measurement or proposal |

+ | Primary source | Paper/proceedings URL, dataset release, run log or supplied manuscript |

+ | Location | Page/section/table/figure, file/line, or exact run artifact |

+ | Checked scope | Identity, publication venue, technical claim, numerical result, or replication |

+ | Conditions | Dataset/version/split, resolution, hardware, prior/model, metric implementation and repeats as relevant |

+ | Assessment | Supported, partially supported, contradicted, unresolved or not applicable |

+ | Time and limits | Retrieval date, version and what was not checked |

- | Need | Section | What you get |

- |---|---|---|

- | Topic selection / innovation (R1/R2) | [§A](references/methodology-core.md) | four topic-selection modes M1-M4 + five opening-sentence variants A.1 (clash/paradox/agenda-revival/pipeline-dissection/legacy-gap) · red-flag checklist · delta typology (5 power levels) |

- | Argumentation framework (R6) | §B | three venue-validated skeleton templates T1-T3 · contribution–evidence bijection procedure · related-work organization · limitations-section style · §B.5 representation-paper techniques (incl. motivation-upgrade rewriting + symbol-system unification) |

- | 3DGS paper drafting (domain extension) | [`references/3dgs-paper-blueprint.md`](references/3dgs-paper-blueprint.md) | 3DGS-specific skeleton program: three writing rules (motivation upgrade / symbol unification / painpoint→resource-metric mapping) · five-section placement table (intro 4-paragraph arc · related-work 3-way moat · method total-then-split with 3DGS math preliminaries · experiment 4-table spec) · painpoint library · limitations library · pre-submission checklist |

- | Experiment design / analysis (R8/R9) | §C | sixteen experiment patterns C1-C16 · domain iron rule (same diffusion prior / resolution / prompt set / evaluation protocol) |

- | Simulated review (R3) | §D | eight reviewer must-ask questions D1-D8 with preempt actions |

- | Polishing / value packaging | §E | nine value amplifiers (naming rights, citation trio, dataset-as-asset, future-work interfaces, …) |

+ An arXiv title match checks metadata only. It does not validate acceptance, a

+ performance claim, causal attribution or replication. Treat lexical matches as

+ candidates. A paper outside the catalog is not fabricated, and a bounded search

+ that finds no match does not establish novelty.

- **Layer 2 — Evidence (mp-*.md, load by direction)**:

+ Do not require an arbitrary number of citations per claim. Use the sources needed

+ to establish the claim; report when the relevant literature is sparse. Citation

+ counts, awards and venue prestige are not substitutes for evidence quality.

- |---|---|---|---|

- | Super Agents & Agentic AI | [`references/mp-super-agents.md`](references/mp-super-agents.md) | ReAct (ICLR 2023, 12,540) · MetaGPT (ICLR 2024 Oral, 2,626) · Toolformer (NeurIPS 2023, 6,146) · Generative Agents (UIST 2023, 5,987) · Voyager (TMLR 2024, 2,622) | minimal-paradigm · human-practice-transfer · data-bootstrapping · memory-architecture · lifelong-skill-library papers |

- | 3D Gaussian Splatting | [`references/mp-3dgs.md`](references/mp-3dgs.md) | 3DGS (TOG 2023, 10,887) · 4DGS (CVPR 2024, 1,676) · SuGaR (CVPR 2024, 982) | representation-revolution · extension-paradigm · downstream-gap (surface/editing) papers |

- | Spatial Intelligence | [`references/mp-spatial.md`](references/mp-spatial.md) | NeRF (ECCV 2020, 5,900) · Instant-NGP (TOG 2022, 6,456) · DUSt3R (CVPR 2024, 2,022) | representation-founding · efficiency-mechanism · pipeline-replacement (geometry foundation model) papers |

+ ## Experiments and comparisons

- Citation counts: Semantic Scholar, verified 2026-10. Division of labor (enforced): new lessons go into methodology-core as generalizable procedures; mp-*.md only ever grows by adding case evidence — never duplicate methodology there.

+ - Map each contribution to an informative test, including alternatives and failure

+ conditions. Benchmark, replication, negative-result and system contributions can

+ be valuable when they answer a real research question.

+ - Choose datasets and baselines by relevance and coverage, not by a fixed quota or

+ publication-age cutoff. Include strong older baselines where still appropriate.

+ - For a component claim, hold relevant confounders constant or measure their effect.

+ For an end-to-end system claim, report changes in priors, training data, budgets

+ and hardware. Do not silently claim that unmatched settings isolate a component.

+ - Multiple metrics and resource dimensions can share a table. Keep each compared

+ quantity on a compatible protocol; label paper-reported and rerun values, and

+ avoid ranking incomparable entries as a controlled experiment.

+ - Choose independent repeats and sample sizes from variability, desired precision,

+ effect sizes and resources. Separate scenes, views, seeds, tasks and episodes;

+ do not count correlated observations as independent replication.

+ - Never invent missing measurements or quietly remove negative results. An

+ unavailable experiment leads to a narrower claim or an explicit open question.

- ## 3. Role Playbooks (condensed)

+ See `references/methodology-core.md` for reusable procedures and

+ `references/3dgs-paper-blueprint.md` for 3DGS-specific planning.

- - **R1 idea evaluation**: ground every "already done" verdict in a named library method with arXiv ID; an idea with zero prior art in a 976-method KB is more likely a search failure than a gap — re-query with 2–3 paraphrases before claiming novelty. Load methodology-core **§A** (§2.4): classify the idea into M1-M4 (combination gap / obstacle-bypass / ecosystem-bridge / pain-point-with-mechanism) via the per-mode yes/no questions, then run the red-flag checklist — any hit → rework the topic, not the writing.

- - **R2 innovation mining**: load methodology-core **§A.3** delta typology — rank claims by power (representation-level > mechanism-level > combination-level > protocol-level > system-level), make each pass the one-sentence delta test (which named baseline, which measurable delta), and prefer combination/transfer/protocol deltas over "we scale it up". Tag every verdict with its mode number (M1-M4) and delta level (1-5) for R6/R8 traceability.

- - **R3 simulated review**: beyond domain-standard scrutiny, walk methodology-core **§D** D1-D8 (representation arbitrariness · own-prior-work overlap · baseline parity · metric blind spots · efficiency caliber · cost disadvantage · mechanism patchwork · hidden negatives); every question the draft cannot answer becomes an objection **with a fix**. The review report must state per-item D1-D8 coverage.

- - **R8 experiment design + R9 analysis**: compose the plan from methodology-core **§C** patterns C1-C16 (state which pattern each experiment table uses) and enforce the Comparison Fairness Protocol from `cg-paper-writing` v3.3.0 — same dataset version/resolution/split, same metric implementation, rerun-before-cite for mismatched calibers, one table one dimension; for 3DGS/generative work the domain iron rule (same diffusion prior / prompt set) is non-negotiable. R9 interprets failures through the failure-mode taxonomy (C5), not anecdotes.

- - **R6 framework compliance**: load methodology-core **§B** — pick the skeleton template matching the paper type (T1 journal system-arc / T2 conference capability-matrix / T3 paradigm-creating minimal-method), run the contribution–evidence bijection (every contribution ↔ unique evidence section; orphan experiments deleted or demoted to appendix), organize related work per B.3 (per-paper verdicts or grouped-gap prose), and write the limitations section per B.4 until Reviewer-2 has nothing left to add. For 3DGS-direction papers, additionally load the §2.4 domain blueprint (`3dgs-paper-blueprint.md`) and walk its five-section placement table + pre-submission checklist.

- - **R7 format**: pull the venue's actual template and checklist (see `cg-paper-writing/references/` venue files) — never format from memory; citation style, anonymization, and figure rules are checked item by item.

+ ## Review and release checks

- ## 4. AIGC-Trace Hygiene — Position and Red Lines

+ Review the supplied work, not an imagined manuscript. Distinguish a scientific

+ error from missing reporting, and an empirical requirement from a style preference.

+ For an objection, provide its location, why it matters, confidence and a remedy.

+ Accept a well-supported rebuttal; do not enforce concession quotas, fixed attack

+ scores or a predetermined accept/reject outcome. Simulated review is preparation,

+ not a forecast of a venue's decision.

- Positioned as **publication hygiene**, not detection-evasion:

+ For a finished artifact, use supported / unresolved / not-applicable checks:

- - Clean: generator/tool metadata left in source files, template-sentence rhythm, unexplained acronym inconsistency, boilerplate transitions, and artifacts of the writing toolchain — things that make the manuscript read as machine-drafted when it is author-owned.

- - Neutralize AI-cliché patterns: empty hedging, "delve/leverage/in the realm of" filler, symmetrical bullet-only rhythm, unwarranted superlatives.

- - **Red line**: do NOT help conceal, misstate, or evade a venue's AI-usage disclosure policy; do NOT fabricate human-authorship evidence. Where a journal requires an AI-use statement, the correct move is compliance — hygiene applies to style and metadata, never to disclosure obligations.

+ 1. Citation identity, venue and claim support are separately checked.

+ 2. Numerical claims trace to original tables or measured artifacts and settings.

+ 3. Contribution wording does not exceed the available evidence.

+ 4. Relevant uncertainty, limitations, failure cases and deviations are disclosed.

+ 5. Formatting and AI-use rules are verified against current official instructions.

- ## 5. Red Lines (whole team)

+ For a draft, unresolved items may remain as explicit placeholders. Never mark a

+ submission-ready gate as passed merely because a checklist or schema is complete.

+ Structured review can use `skills/_contracts/research-review.schema.json`; schema

+ validation checks structure and declared provenance, not the truth of a conclusion.

- - Every capability/prior-art claim maps to a library method with arXiv ID; no fabricated methods, IDs, or numbers; unverified items are marked `unverified`.

- - Re-query `data/methods.json` before any baseline/related-work recommendation — never answer from memory.

- - No cross-dimension comparisons in one table (Comparison Fairness Protocol).

- - Reviewer simulation must produce objections **with fixes**, not verdicts alone; no "reject without reasons".

- - Polishing may tighten language, never silently weaken or upgrade a technical claim.

+ ## Supporting skills and examples

- ## 6. Related Skills

+ Use `3dgs-paper-reader`, `3dgs-method-compare`, `3dgs-experiment-planner`,

+ `cg-paper-writing` and `3dgs-visualizer` where their work is needed. Availability of

+ external tools and compute must be checked rather than assumed.

- - **cg-paper-writing** — writing templates, claim-verification chain, Gate 1/2 integrity checks, comparison-fairness protocol, venue format references (the deep assets this team orchestrates)

- - **references/ (this skill, two layers + domain extension)** — [`methodology-core.md`](references/methodology-core.md): the executable methodology program distilled from the 15 exemplars (A 选题引擎+开题变体 / B 论述架构+表示论文技法+骨架路由 / C 实验模式库 C1-C16 / D 审稿人预期库 / E 价值放大器 / F 团队约定); [`3dgs-paper-blueprint.md`](references/3dgs-paper-blueprint.md): 3DGS-specific drafting skeleton (three writing rules, five-section placement, painpoint/limitation libraries, cross-referenced with cg-paper-writing templates); `mp-*.md` (5 files, 5 directions, 15 exemplar papers): per-direction full-text evidence base tracing every program step to its source paper

- - **3dgs-experiment-planner** — detailed experiment design for CVPR/SIGGRAPH/TVCG

- - **3dgs-method-compare** — 10+ dimension method comparison tables

- - **3dgs-paper-reader** — structured reading of the anchor papers above

- - **ontology** / **super-agent** — deep-dives into two of the five covered directions (Ontology, Super Agents)

- - **3dgs-visualizer** — publication charts for R9 analysis output

+ `references/mp-3dgs.md`, `mp-spatial.md`, `mp-ontology.md`, `mp-aigc.md` and

+ `mp-super-agents.md` are primary-source case indexes with explicitly labeled

+ methodological interpretations. They are not universal writing rules or evidence

+ that copying a successful paper's structure improves acceptance odds.