git:20261013.f78768c to git:20261015.6eccba4

30 added, 4 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 three frontier directions — Ontology & Knowledge Representation, Super Agents & Agentic AI, AIGC Content Creation. Use when: evaluating a research idea, finding novelty/创新点, simulating reviewers 审稿, polishing 润色, 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."

---

# 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 for three frontier directions.

> Knowledge base: 955 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 22-method **AIGC Content Creation** categories.

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

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

| R3 | **Simulated Reviewer** | draft complete, or pre-submission self-check | Reviewer-1 (novelty), Reviewer-2 (technical rigor), Reviewer-3 (experiments) + meta-review; every objection paired with a fix | `cg-paper-writing` Gates |

| R5 | **AIGC-Trace Hygiene Expert** | pre-submission hygiene pass | LaTeX-source metadata cleanup (generator fields, AIGC tool markers), AI-cliché and template-rhythm neutralization, style naturalization | §4 red lines |

| R6 | **Framework & Venue Compliance Expert** | structure pass | story-line check, contribution–evidence mapping, venue-specific structure conventions (TVCG/CGF/CAD journal vs. CVPR/ICCV/SIGGRAPH conference) | venue refs in `cg-paper-writing` |

| R7 | **Format & Typesetting Expert** | camera-ready or submission package | strict venue-template compliance: citation style, figure/table/caption rules, page & section limits, anonymization check | venue refs in `cg-paper-writing` |

| R8 | **Experiment Designer** | contribution claims fixed | baseline set (re-query the KB for the right baselines), ablation matrix, metrics, datasets — under the Comparison Fairness Protocol | `3dgs-experiment-planner` |

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

## 2. Three Frontier Directions — Method-Selection Tables

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)

| 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 |

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

| 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 |

### 2.3 AIGC Content Creation (22 methods)

| 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 |

**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).

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

+ Two-layer asset: **`methodology-core.md` is the program** — the six highest-cited exemplars distilled into executable operating procedures, domain-tuned to 3DGS/spatial-intelligence (not bound to those six 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:

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

+ | Need | Section | What you get |

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

+ | Topic selection / innovation (R1/R2) | [§A](references/methodology-core.md) | four topic-selection modes M1-M4 with per-mode yes/no questions · 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 |

+ | Experiment design / analysis (R8/R9) | §C | nine experiment patterns C1-C9 · 3DGS 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 | six value amplifiers (naming rights, citation trio, teaser-as-ad, …) |

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

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

+ | 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) | minimal-paradigm papers · human-practice-transfer papers |

+ 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.

## 3. Role Playbooks (condensed)

- - **R1 idea evaluation**: ground every "already done" verdict in a named library method with arXiv ID; an idea with zero prior art in a 955-method KB is more likely a search failure than a gap — re-query with 2–3 paraphrases before claiming novelty.

- - **R2 innovation mining**: prefer **combination / transfer / protocol** deltas (e.g., neuro-symbolic × AIGC evaluation; super-agent orchestration × scene-graph memory) over "we scale it up" deltas; each innovation claim must name the baseline it moves against.

- - **R8 experiment design + R9 analysis**: enforce the Comparison Fairness Protocol (same dataset version/resolution/split, same metric implementation, rerun-before-cite for mismatched calibers) from `cg-paper-writing` v3.3.0 — one table, one dimension.

- - **R6 framework compliance**: journal papers need a sustained argument thread + related-work that positions; conference papers need a front-loaded contribution punch in the first 1.5 pages; verify contribution-list ↔ evidence-section bijection.

+ - **R1 idea evaluation**: ground every "already done" verdict in a named library method with arXiv ID; an idea with zero prior art in a 955-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-C9 (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.

- **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.

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

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

- 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.

## 5. Red Lines (whole team)

- 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.

## 6. Related Skills

- **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)** — [`methodology-core.md`](references/methodology-core.md): the executable methodology program distilled from the six exemplars (A 选题引擎 / B 论述架构 / C 实验模式库 / D 审稿人预期库 / E 价值放大器); `mp-*.md`: 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 three frontier directions

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