Paper Review Pipeline (ML Top Conferences)
Run a two-view paper review for ML conference submissions:
- Section-by-section review (Abstract → Intro → Method → Experiments → …) with concrete edits.
- Prioritized issue list with P0/P1/P2 severity, grouped by category, including recommended fixes and verification notes.
This skill also supports rebuttal / review response: parse reviewer comments, classify, choose a strategy, and draft a professional point-by-point response.
Parity Guarantee (No-Omission)
This skill is a consolidation layer. It must not omit any distinctive workflow, constraints, or output formats from the legacy skills it replaces.
Use:
references/parity-matrix.md as the feature-parity contract and regression scenarios.
references/modules/ for full imported workflows and checklists.
Execution Modes
This skill supports two modes:
- Default:
targeted — run only the most relevant tracks based on the user request and inputs, but always produce the Final Synthesis.
- Optional:
full-parallel — run all tracks as independent outputs (acceptable redundancy), then produce the Final Synthesis.
Trigger full-parallel when the user says: “全量”, “并行”, “run all tracks”, “run every skill”, “逐个 skill 测试”, “full pipeline”.
Protocol + required report template:
references/full-parallel-protocol.md
references/report-template.md
Optional External Second Opinion (paperreview)
If the user provides a near-final or final PDF, complete the local review first and then explicitly ask whether they also want an external second opinion via paperreview.
- Ask only when the input is clearly a near-final or final PDF.
- Do not auto-submit to external services. Get explicit confirmation first.
- If the user agrees, hand off to
paperreview as a follow-up step; otherwise finish with the local pipeline result only.
When to Use
- Pre-submission quality check (ICML/ICLR/NeurIPS/AAAI).
- After a draft is “mostly done” but clarity/logic is shaky.
- When you suspect citation problems (missing, inconsistent, unverified, or claim-to-citation mismatch).
- Before sending to advisor/collaborators for feedback (to reduce “obvious issues”).
- After receiving reviews: draft rebuttal and a revision plan.
When NOT to Use
- If the user wants new research or new experiments invented: require the user to provide results/artifacts.
- If the user asks for verbatim PDF-to-LaTeX copying or large-scale reformatting without source.
- If the task is pure BibTeX generation from memory: do not do it; use verified metadata workflows.
Non-Negotiable Guardrails
- No hallucinated citations.
- If a citation cannot be verified, mark it as
[CITATION NEEDED] / placeholder and tell the user explicitly.
- Do not fabricate authors/years/venues/DOIs.
- Do not change technical meaning.
- When proposing edits, preserve claims and numbers unless the user provides corrected data.
- Preserve LaTeX semantics when editing source.
- Do not break
\\cite{}, \\ref{}, \\label{}, math environments, figures/tables, or bibliography hooks.
- Respect blind review constraints (if applicable).
- Avoid identity-revealing self-citations, acknowledgments, or repo links unless the user confirms it is camera-ready.
Inputs (Ask for the Minimum Needed)
For pre-submission review
- Paper source: LaTeX section text, or the relevant excerpts pasted in chat (preferred), and optionally the PDF for context.
- Target venue + track (ICML/ICLR/NeurIPS/AAAI) and any required sections (e.g., limitations / broader impact / ethics).
- One-sentence contribution (if the user has it). If not, infer and ask for confirmation.
For rebuttal / review response
- Reviewer comments (verbatim text, ideally grouped by reviewer).
- Rebuttal constraints: word/page limit, formatting (ICLR OpenReview vs PDF), and timeline.
- What the user is willing to change: “clarify only” vs “add experiments” vs “major rewrite”.
Output Contract (Always Produce Both Views)
View A — Section-by-Section Review
For each section, output:
- What works (1–3 bullets)
- What’s missing / unclear (P0/P1/P2 tagged bullets)
- Concrete fixes (rewrite suggestions or structural moves)
Use the checklists in:
references/section-review-checklist.md
View B — P0/P1/P2 Issue List (Prioritized)
Format each issue like:
- Priority: P0 (blocking) / P1 (important) / P2 (nice-to-have)
- Category: Narrative / Evidence / Experimental Design / Statistics / Reproducibility / Citations / Writing / Figures-Tables / Format
- Where: section name + a short quote anchor (or LaTeX label if available)
- Problem
- Fix
- Verification (if needed): what evidence/log/search is required before claiming it is correct
Use the taxonomy in:
references/p0-p2-taxonomy.md
Modules (Routing Rules)
Use these modules to preserve legacy feature parity:
- Paper-level QA:
references/modules/paper-self-review.md
- Rebuttal / review response:
references/modules/review-response.md and references/rebuttal-workflow.md
- LaTeX + BibTeX toolbox:
references/modules/academic-paper-helper.md
- Claim-level citation audit:
references/modules/citation-validator.md and references/citation-integrity.md
- Anti-AI polish:
references/modules/writing-anti-ai.md
- LaTeX rhythm pass:
references/modules/latex-rhythm-refiner.md
- Literature discovery (extended):
references/modules/claude-scholar-ml-paper-writing.md
Workflow (Default)
Pass 0 — Triage (5–10 minutes)
- Identify the one-sentence contribution and confirm it with the user.
- Extract the top 3–7 claims the paper relies on.
- For each claim, note the current support:
- empirical result (table/figure)
- theorem/proof
- citation / prior work
- ablation / analysis
- Flag immediate P0 risks (typical: missing baselines, unclear experimental protocol, unverified citations, paper “about X” but experiments test Y).
Pass 1 — Section Review (primary deliverable)
Review sections in order (and check alignment between them):
- Abstract
- Introduction (motivation → gap → contribution bullets)
- Related Work (positioning + not a bibliography dump)
- Method (reproducible description + design justification)
- Experiments / Results (fair baselines + full setup + statistical reporting)
- Analysis / Ablations (claim-driven, not exploratory noise)
- Limitations / Broader Impact / Ethics (venue-dependent)
- Conclusion (tight restatement + constraints + future work without overclaim)
Pass 2 — Consolidate into P0/P1/P2
Turn findings into an actionable issue list:
- P0 first (blockers / likely desk-reject causes)
- P1 next (acceptance probability movers)
- P2 last (polish)
Pass 3 (Optional) — Revision Plan
If the user wants an execution plan, produce:
- 3–8 tasks with measurable acceptance criteria
- suggested order (dependency-aware)
- what can be done in parallel (writing vs experiments vs citations)
Full-Parallel Workflow (Comprehensive)
When mode is full-parallel, do not collapse everything into one voice. Instead:
- Run all tracks (A–G) as independent “mini-reviewers” and keep each track output visible.
- Then produce the Final Synthesis:
- View A: consolidated section-by-section review
- View B: consolidated P0/P1/P2 issue list
- minimal revision plan
- conflicts & resolutions between tracks
Use:
references/full-parallel-protocol.md
references/report-template.md
Rebuttal / Review Response Module
When reviews arrive, do:
- Parse and classify each comment: Major / Minor / Clarification / Missing baseline / Missing experiment / Writing / Citation / Misunderstanding.
- Choose a strategy per item: Accept + change, Clarify, Defend, Add experiment, Defer (explain constraints).
- Draft point-by-point responses with:
- gratitude + precise restatement
- what changed (or why not)
- where to find it (section/figure/table)
- evidence-based tone (no overpromising)
Use:
references/rebuttal-workflow.md
Citation Integrity Module (Hard Requirement)
Before submission, ensure:
- every non-trivial factual claim has a citation or empirical evidence
- every citation key resolves in the bibliography
- any newly added citations are verified (paper exists; BibTeX not fabricated)
If deep citation audit is requested, follow:
references/citation-integrity.md
1---
2name: paper-review-pipeline
3description: Use when a mostly complete ML conference paper needs self-review, pre-submission QA, camera-ready checking, section-by-section critique, citation-risk inspection, or rebuttal/review-response drafting. Skip this for initial drafting and use `paperreview` only when the user explicitly wants external submission.
4metadata:
5 source: https://skillmd.com/skills/gabrielmoreira/paper-review-pipeline
6---
7
8# Paper Review Pipeline (ML Top Conferences)
9
10Run a *two-view* paper review for ML conference submissions:
11
121) **Section-by-section review** (Abstract → Intro → Method → Experiments → …) with concrete edits.
132) **Prioritized issue list** with **P0/P1/P2** severity, grouped by category, including recommended fixes and verification notes.
14
15This skill also supports **rebuttal / review response**: parse reviewer comments, classify, choose a strategy, and draft a professional point-by-point response.
16
17## Parity Guarantee (No-Omission)
18
19This skill is a consolidation layer. It must **not** omit any distinctive workflow, constraints, or output formats from the legacy skills it replaces.
20
21Use:
22- `references/parity-matrix.md` as the feature-parity contract and regression scenarios.
23- `references/modules/` for full imported workflows and checklists.
24
25## Execution Modes
26
27This skill supports two modes:
28
29- **Default: `targeted`** — run only the most relevant tracks based on the user request and inputs, but **always** produce the Final Synthesis.
30- **Optional: `full-parallel`** — run **all tracks** as independent outputs (acceptable redundancy), then produce the Final Synthesis.
31
32Trigger `full-parallel` when the user says: “全量”, “并行”, “run all tracks”, “run every skill”, “逐个 skill 测试”, “full pipeline”.
33
34Protocol + required report template:
35- `references/full-parallel-protocol.md`
36- `references/report-template.md`
37
38## Optional External Second Opinion (`paperreview`)
39
40If the user provides a near-final or final PDF, complete the local review first and then explicitly ask whether they also want an external second opinion via `paperreview`.
41
42- Ask only when the input is clearly a near-final or final PDF.
43- Do **not** auto-submit to external services. Get explicit confirmation first.
44- If the user agrees, hand off to `paperreview` as a follow-up step; otherwise finish with the local pipeline result only.
45
46## When to Use
47
48- Pre-submission quality check (ICML/ICLR/NeurIPS/AAAI).
49- After a draft is “mostly done” but clarity/logic is shaky.
50- When you suspect citation problems (missing, inconsistent, unverified, or claim-to-citation mismatch).
51- Before sending to advisor/collaborators for feedback (to reduce “obvious issues”).
52- After receiving reviews: draft rebuttal and a revision plan.
53
54## When NOT to Use
55
56- If the user wants **new research** or **new experiments** invented: require the user to provide results/artifacts.
57- If the user asks for **verbatim PDF-to-LaTeX copying** or large-scale reformatting without source.
58- If the task is **pure BibTeX generation from memory**: do not do it; use verified metadata workflows.
59
60## Non-Negotiable Guardrails
61
621) **No hallucinated citations.**
63 - If a citation cannot be verified, mark it as `[CITATION NEEDED]` / placeholder and tell the user explicitly.
64 - Do not fabricate authors/years/venues/DOIs.
652) **Do not change technical meaning.**
66 - When proposing edits, preserve claims and numbers unless the user provides corrected data.
673) **Preserve LaTeX semantics when editing source.**
68 - Do not break `\\cite{}`, `\\ref{}`, `\\label{}`, math environments, figures/tables, or bibliography hooks.
694) **Respect blind review constraints (if applicable).**
70 - Avoid identity-revealing self-citations, acknowledgments, or repo links unless the user confirms it is camera-ready.
71
72## Inputs (Ask for the Minimum Needed)
73
74### For pre-submission review
75- Paper source: LaTeX section text, or the relevant excerpts pasted in chat (preferred), and optionally the PDF for context.
76- Target venue + track (ICML/ICLR/NeurIPS/AAAI) and any required sections (e.g., limitations / broader impact / ethics).
77- One-sentence contribution (if the user has it). If not, infer and ask for confirmation.
78
79### For rebuttal / review response
80- Reviewer comments (verbatim text, ideally grouped by reviewer).
81- Rebuttal constraints: word/page limit, formatting (ICLR OpenReview vs PDF), and timeline.
82- What the user is willing to change: “clarify only” vs “add experiments” vs “major rewrite”.
83
84## Output Contract (Always Produce Both Views)
85
86### View A — Section-by-Section Review
87
88For each section, output:
89- **What works** (1–3 bullets)
90- **What’s missing / unclear** (P0/P1/P2 tagged bullets)
91- **Concrete fixes** (rewrite suggestions or structural moves)
92
93Use the checklists in:
94- `references/section-review-checklist.md`
95
96### View B — P0/P1/P2 Issue List (Prioritized)
97
98Format each issue like:
99- **Priority**: P0 (blocking) / P1 (important) / P2 (nice-to-have)
100- **Category**: Narrative / Evidence / Experimental Design / Statistics / Reproducibility / Citations / Writing / Figures-Tables / Format
101- **Where**: section name + a short quote anchor (or LaTeX label if available)
102- **Problem**
103- **Fix**
104- **Verification** (if needed): what evidence/log/search is required before claiming it is correct
105
106Use the taxonomy in:
107- `references/p0-p2-taxonomy.md`
108
109## Modules (Routing Rules)
110
111Use these modules to preserve legacy feature parity:
112
113- **Paper-level QA**: `references/modules/paper-self-review.md`
114- **Rebuttal / review response**: `references/modules/review-response.md` and `references/rebuttal-workflow.md`
115- **LaTeX + BibTeX toolbox**: `references/modules/academic-paper-helper.md`
116- **Claim-level citation audit**: `references/modules/citation-validator.md` and `references/citation-integrity.md`
117- **Anti-AI polish**: `references/modules/writing-anti-ai.md`
118- **LaTeX rhythm pass**: `references/modules/latex-rhythm-refiner.md`
119- **Literature discovery (extended)**: `references/modules/claude-scholar-ml-paper-writing.md`
120
121## Workflow (Default)
122
123### Pass 0 — Triage (5–10 minutes)
124
1251) Identify the **one-sentence contribution** and confirm it with the user.
1262) Extract the **top 3–7 claims** the paper relies on.
1273) For each claim, note the current support:
128 - empirical result (table/figure)
129 - theorem/proof
130 - citation / prior work
131 - ablation / analysis
1324) Flag immediate P0 risks (typical: missing baselines, unclear experimental protocol, unverified citations, paper “about X” but experiments test Y).
133
134### Pass 1 — Section Review (primary deliverable)
135
136Review sections in order (and check alignment between them):
1371) Abstract
1382) Introduction (motivation → gap → contribution bullets)
1393) Related Work (positioning + not a bibliography dump)
1404) Method (reproducible description + design justification)
1415) Experiments / Results (fair baselines + full setup + statistical reporting)
1426) Analysis / Ablations (claim-driven, not exploratory noise)
1437) Limitations / Broader Impact / Ethics (venue-dependent)
1448) Conclusion (tight restatement + constraints + future work without overclaim)
145
146### Pass 2 — Consolidate into P0/P1/P2
147
148Turn findings into an actionable issue list:
149- P0 first (blockers / likely desk-reject causes)
150- P1 next (acceptance probability movers)
151- P2 last (polish)
152
153### Pass 3 (Optional) — Revision Plan
154
155If the user wants an execution plan, produce:
156- 3–8 tasks with measurable acceptance criteria
157- suggested order (dependency-aware)
158- what can be done in parallel (writing vs experiments vs citations)
159
160## Full-Parallel Workflow (Comprehensive)
161
162When mode is `full-parallel`, do not collapse everything into one voice. Instead:
163
1641) Run all tracks (A–G) as independent “mini-reviewers” and keep each track output visible.
1652) Then produce the Final Synthesis:
166 - View A: consolidated section-by-section review
167 - View B: consolidated P0/P1/P2 issue list
168 - minimal revision plan
169 - conflicts & resolutions between tracks
170
171Use:
172- `references/full-parallel-protocol.md`
173- `references/report-template.md`
174
175## Rebuttal / Review Response Module
176
177When reviews arrive, do:
1781) Parse and classify each comment: **Major / Minor / Clarification / Missing baseline / Missing experiment / Writing / Citation / Misunderstanding**.
1792) Choose a strategy per item: **Accept + change**, **Clarify**, **Defend**, **Add experiment**, **Defer (explain constraints)**.
1803) Draft point-by-point responses with:
181 - gratitude + precise restatement
182 - what changed (or why not)
183 - where to find it (section/figure/table)
184 - evidence-based tone (no overpromising)
185
186Use:
187- `references/rebuttal-workflow.md`
188
189## Citation Integrity Module (Hard Requirement)
190
191Before submission, ensure:
192- every non-trivial factual claim has a citation or empirical evidence
193- every citation key resolves in the bibliography
194- any newly added citations are verified (paper exists; BibTeX not fabricated)
195
196If deep citation audit is requested, follow:
197- `references/citation-integrity.md`