Ml Paper Writing

ML Paper Writing

Expert guidance for NeurIPS, ICML, ICLR, ACL, AAAI, COLM. For systems venues (OSDI, NSDI, ASPLOS, SOSP), use the venue's own authoring guidelines instead.

When NOT to use

• Need internal prose (minutes, reviews, updates) — use internal-writing.
• Need software docs (tutorials/how-tos) — use documentation-writer.
• No research repo or contribution yet — use brainstorming → research to establish one first.
• Paper is for a non-ML venue without LaTeX — use venue template directly.

⚠️ CRITICAL: Never Hallucinate Citations

AI-generated citations have a ~40% error rate. Never write BibTeX from memory. Always fetch programmatically.

• Found paper with DOI → fetch BibTeX via CrossRef/arXiv API ✅
• Cannot verify → mark \cite{PLACEHOLDER_author2024_verify} and tell the scientist ✅
• Guessing a plausible reference → never ❌

Full citation workflow (Semantic Scholar API, DOI fetch, placeholder rules): references/citation-workflow.md. Bibliography: references/sources.md.

Post-draft check: grep -R "PLACEHOLDER" --include="*.tex" --include="*.bib" . must return 0 lines before submission; any \\cite{PLACEHOLDER is a blocker.

Setup

Before drafting, run first-time setup:

1. Explore the repo (README, results, .bib files, experiment configs)
2. Confirm the one-sentence contribution with the scientist
3. Identify the target venue (→ Conference Quick Reference below)

Core Principle

Be proactive. Deliver a draft, then iterate.

• Flag uncertainties inline, don't block on them
• Iterate on feedback

Paper Structure

Write sections in the order specified by references/paper-structure.md. The reference defines the complete checklist (contribution → Figure 1 → Abstract → Introduction → Methods → Experiments → Related Work → Limitations → checklist) with timing notes and the abstract formula. Read it when you start drafting.

Writing Principles

Key rules — full guide with examples: references/writing-guide.md

• One contribution: state it in one sentence; if you can't, you don't have a paper yet
• Stress position: place the key result at the end of the sentence
• Verbs not nominalizations: "we analyzed" not "we performed an analysis"
• No hedging: drop "may," "can," "essentially," "basically," "very"
• Consistent terminology: pick one term per concept and never vary it
• Specific over vague: "accuracy" not "performance"; "15% lower latency" not "faster"
• Minimize pronouns: "this result shows" not "this shows"

Conference Quick Reference

Snapshot only — verify against the venue's current Call for Papers before submitting (limits change every cycle). Full table lives in references/conference-quick-ref.md.

Universal: double-blind, references don't count, LaTeX required, appendices unlimited. Templates in templates/ — always copy the entire directory, not just main.tex. Venue submission checklists: references/checklists.md.

Format Conversion (Resubmission)

Never merge preambles. Start with the target template; copy only content sections.

When cutting: move proofs to appendix, cite surveys instead of individual papers, combine tables, tighten prose.

Rebuttal and reviewer evaluation criteria: references/reviewer-guidelines.md.

Tables and Figures

• Figures: vector (PDF/EPS) for plots; raster (PNG 600 DPI) only for photos; colorblind-safe palette (Okabe-Ito); verify grayscale readability; captions must be self-contained
• Tables: use booktabs; bold best value per metric; include direction symbols (↑/↓); consistent decimal precision

Related skills

• research — source gathering before writing.
• internal-writing — internal prose, not papers.
• documentation-writer — software docs.

Completion criteria

• Contribution one-liner + venue confirmed; no BibTeX hallucinated
• grep -R PLACEHOLDER returns 0; all cites fetched via CrossRef/arXiv
• Draft sections follow references/paper-structure.md order with venue limits respected