| name | ml-paper-writing |
| description | Write publication-ready ML/AI papers for NeurIPS, ICML, ICLR, ACL, AAAI, COLM. Use when drafting papers from research repos, structuring arguments, verifying citations, or preparing camera-ready submissions. Includes LaTeX templates, reviewer guidelines, and citation verification workflows. |
| version | 1.0.0 |
| author | Orchestra Research |
| license | MIT |
| dependencies | ["semanticscholar","arxiv","habanero","requests"] |
| metadata | {"hermes":{"tags":["Academic Writing","NeurIPS","ICML","ICLR","ACL","AAAI","COLM","LaTeX","Paper Writing","Citations","Research"]}} |
ML Paper Writing for Top AI Conferences
Expert-level guidance for writing publication-ready papers targeting NeurIPS, ICML, ICLR, ACL, AAAI, and COLM. Combines writing philosophy from top researchers (Nanda, Farquhar, Karpathy, Lipton, Steinhardt, Perez) with practical tools: LaTeX templates, citation verification APIs, and conference checklists. This file is the summary; the deep dives live in references/.
When to Use This Skill
- Starting from a research repo to write a paper
- Drafting or revising specific sections
- Finding and verifying citations for related work
- Formatting for conference submission
- Resubmitting to a different venue (format conversion)
- Iterating on drafts with scientist feedback
Always remember: First drafts are starting points for discussion, not final outputs.
Core Philosophy: Collaborative but Proactive
Paper writing is collaborative, but be proactive in delivering drafts. Understand the project (repo, results, docs), deliver a complete first draft when confident, search literature via APIs, then refine through feedback cycles. Don't block waiting for feedback on every section—produce something concrete the scientist can react to.
| Confidence Level | Action |
|---|
| High (clear repo, obvious contribution) | Write full draft, deliver, iterate on feedback |
| Medium (some ambiguity) | Write draft with flagged uncertainties, continue |
| Low (major unknowns) | Ask 1-2 targeted questions, then draft |
Only block for input when: target venue is unclear, contradictory framings seem equally valid, results look incomplete/inconsistent, or review is explicitly requested. Per-section flagging tables and repo-exploration steps: references/paper-workflow.md.
⚠️ CRITICAL: Never Hallucinate Citations
AI-generated citations have a ~40% error rate. Hallucinated references are academic misconduct that can cause desk rejection or retraction.
The Rule: NEVER generate BibTeX from memory. ALWAYS fetch programmatically.
Citation Verification (MANDATORY for every citation) — Workflow 2:
- [ ] Step 1: Search using Exa MCP or Semantic Scholar API
- [ ] Step 2: Verify paper exists in 2+ sources (Semantic Scholar + arXiv/CrossRef)
- [ ] Step 3: Retrieve BibTeX via DOI (programmatically, not from memory)
- [ ] Step 4: Verify the claim you're citing actually appears in the paper
- [ ] Step 5: Add verified BibTeX to bibliography
- [ ] Step 6: If ANY step fails → mark as placeholder, inform scientist
If you cannot verify a citation, mark it explicitly and tell the scientist:
\cite{PLACEHOLDER_author2024_verify_this} % TODO: Verify this citation exists
Summary: Citation Rules
| Situation | Action |
|---|
| Found paper, got DOI, fetched BibTeX | ✅ Use the citation |
| Found paper, no DOI | ✅ Use arXiv BibTeX or manual entry from paper |
| Paper exists but can't fetch BibTeX | ⚠️ Mark placeholder, inform scientist |
| Uncertain if paper exists | ❌ Mark [CITATION NEEDED], inform scientist |
| "I think there's a paper about X" | ❌ NEVER cite - search first or mark placeholder |
APIs, Python code, Exa MCP setup, BibTeX management: references/citation-workflow.md.
The Narrative Principle
Your paper is not a collection of experiments—it's a story with one clear contribution supported by evidence. Neel Nanda: "A paper is a short, rigorous, evidence-based technical story with a takeaway readers care about."
Three pillars must be crystal clear by the end of the introduction:
| Pillar | Description | Example |
|---|
| The What | 1-3 specific novel claims within cohesive theme | "We prove that X achieves Y under condition Z" |
| The Why | Rigorous empirical evidence supporting claims | Strong baselines, experiments distinguishing hypotheses |
| The So What | Why readers should care | Connection to recognized community problems |
If you cannot state your contribution in one sentence, you don't yet have a paper. Full framework: references/writing-guide.md.
Workflow 0: Starting from a Research Repository
Project Understanding:
- [ ] Step 1: Explore the repository structure
- [ ] Step 2: Read README, existing docs, and key results
- [ ] Step 3: Identify the main contribution with the scientist
- [ ] Step 4: Find papers already cited in the codebase
- [ ] Step 5: Search for additional relevant literature
- [ ] Step 6: Outline the paper structure together
- [ ] Step 7: Draft sections iteratively with feedback
Existing citations in the repo are high-signal Related Work starting points. Never assume the narrative—confirm the contribution framing with the scientist. Step-by-step commands and draft-delivery guidance: references/paper-workflow.md.
Workflow 1: Writing a Complete Paper (Iterative)
Each step is draft → feedback → revise:
Paper Writing Progress:
- [ ] Step 1: Define the one-sentence contribution (with scientist)
- [ ] Step 2: Draft Figure 1 → get feedback → revise
- [ ] Step 3: Draft abstract → get feedback → revise
- [ ] Step 4: Draft introduction → get feedback → revise
- [ ] Step 5: Draft methods → get feedback → revise
- [ ] Step 6: Draft experiments → get feedback → revise
- [ ] Step 7: Draft related work → get feedback → revise
- [ ] Step 8: Draft limitations → get feedback → revise
- [ ] Step 9: Complete paper checklist (required)
- [ ] Step 10: Final review cycle and submission
Key constraints per step:
- Contribution: requires explicit scientist confirmation
- Figure 1: many readers skip straight to it—vector graphics, standalone caption, B&W-readable
- Abstract: Farquhar's 5-sentence formula (achievement → why hard → how → evidence → best number); delete generic openings
- Introduction: ≤1.5 pages, 2-4 contribution bullets, methods start by page 2-3
- Methods: enable reimplementation—pseudocode, all hyperparameters, architecture details
- Experiments: each states the claim it supports; error bars with methodology, search ranges, compute, seeds
- Related work: organize methodologically, not paper-by-paper; cite generously
- Limitations: REQUIRED everywhere; honesty helps—reviewers must not penalize it
- Checklist: see references/checklists.md
Full per-step instructions and troubleshooting common issues: references/paper-workflow.md.
Writing Philosophy for Top ML Conferences
Distilled from Nanda, Farquhar, Karpathy, Gopen & Swan, Lipton, Steinhardt, and Perez—these separate accepted papers from rejected ones:
- Time allocation (Nanda): spend equal time on the abstract, the introduction, the figures, and everything else combined—readers go title → abstract → intro → figures → maybe the rest.
- Sentence-level clarity (Gopen & Swan, 7 principles): subject-verb proximity, stress position (emphasis at sentence end), topic position (context first), old-before-new, one unit one function, action in the verb, context before new information.
- Micro-level (Perez): minimize pronouns, verbs early, unfold awkward apostrophes ("X's Y" → "the Y of X"), delete filler words.
- Word choice (Lipton): be specific ("accuracy" not "performance"), eliminate hedging and intensifiers, avoid incremental vocabulary ("combine"/"modify" → "develop"/"propose").
- Precision (Steinhardt): one term per concept, state assumptions formally, intuition alongside rigor.
- What reviewers read: abstract 100%, intro skimmed by 90%+, figures before methods, appendix rarely—front-load value.
Full principles with examples: references/writing-guide.md. Source bibliography with links: references/sources.md.
Conference Requirements Quick Reference
| Conference | Page Limit | Extra for Camera-Ready | Key Requirement |
|---|
| NeurIPS 2025 | 9 pages | +0 | Mandatory checklist, lay summary for accepted |
| ICML 2026 | 8 pages | +1 | Broader Impact Statement required |
| ICLR 2026 | 9 pages | +1 | LLM disclosure required, reciprocal reviewing |
| ACL 2025 | 8 pages (long) | varies | Limitations section mandatory |
| AAAI 2026 | 7 pages | +1 | Strict style file adherence |
| COLM 2025 | 9 pages | +1 | Focus on language models |
Universal Requirements:
- Double-blind review (anonymize submissions)
- References don't count toward page limit
- Appendices unlimited but reviewers not required to read
- LaTeX required for all venues
Per-venue checklists (NeurIPS 16-item, ICML, ICLR, ACL): references/checklists.md.
Workflow 4: Using LaTeX Templates Properly
Template Setup Checklist:
- [ ] Step 1: Copy entire template directory to new project
- [ ] Step 2: Verify template compiles as-is (before any changes)
- [ ] Step 3: Read the template's example content to understand structure
- [ ] Step 4: Replace example content section by section
- [ ] Step 5: Keep template comments/examples as reference until done
- [ ] Step 6: Clean up template artifacts only at the end
Copy the ENTIRE directory (.sty/.bst files are required), never edit style files, compile frequently. Templates for all six venues live in templates/. Step detail, pitfalls table, and per-venue main/style file names: references/latex-templates.md.
Workflow 3: Conference Resubmission & Format Conversion
Format Conversion Checklist:
- [ ] Step 1: Identify source and target template differences
- [ ] Step 2: Create new project with target template
- [ ] Step 3: Copy content sections (not preamble)
- [ ] Step 4: Adjust page limits and content
- [ ] Step 5: Update conference-specific requirements
- [ ] Step 6: Verify compilation and formatting
Never copy LaTeX preambles between templates—start fresh from the target template and migrate content only. Address reviewer concerns in the new version, but never reference the previous submission (blind review). Page-limit deltas, venue-specific additions, and conversion pitfalls: references/latex-templates.md.
Reviewer Evaluation Criteria
Reviewers assess four dimensions: Quality (technical soundness), Clarity (reproducible by experts), Significance (community impact), Originality (new insights—doesn't require a new method). NeurIPS scores on a 6-point scale from Strong Reject (1) to Strong Accept (6). Detailed criteria, scoring rubrics, common concerns, and rebuttal guidance: references/reviewer-guidelines.md.
Tables and Figures
- Tables:
booktabs package; bold the best value, direction arrows (↑/↓), right-align numbers, consistent precision.
- Figures: vector graphics (PDF/EPS) for plots; colorblind-safe palettes (Okabe-Ito/Paul Tol); verify grayscale readability; no title inside the figure; self-contained captions.
LaTeX examples and full design rules: references/writing-guide.md.
References & Resources
| Document | Contents |
|---|
| paper-workflow.md | Repo-to-paper workflow, proactivity tables, per-step writing guide, troubleshooting |
| writing-guide.md | Gopen & Swan 7 principles, Perez micro-tips, word choice, math/figure/table design |
| citation-workflow.md | Citation APIs, Python code, Exa MCP, BibTeX management |
| checklists.md | NeurIPS 16-item, ICML, ICLR, ACL requirements |
| reviewer-guidelines.md | Evaluation criteria, scoring, rebuttals |
| latex-templates.md | Template setup, pitfalls, format conversion |
| sources.md | Complete bibliography of all sources |
LaTeX templates in templates/: ICML 2026, ICLR 2026, NeurIPS 2025, ACL/EMNLP, AAAI 2026, COLM 2025 (see templates/README.md for compilation setup).
