NeurIPS Supplementary

Use this skill to prevent two common NeurIPS failures: hiding the main contribution in the appendix and overloading the main paper with details that belong in supplemental material.

Placement rules

• Main paper: problem, contribution, method, essential theory, core experiments, limitations, and enough detail for reviewers to judge correctness.
• References: complete, verified citations; do not rely on AI-generated references without manual verification.
• Text appendices: proofs, derivations, additional ablations, dataset details, model cards, extended limitations, and supporting analyses.
• Checklist: mandatory and placed according to current template instructions.
• Separate ZIP: code, data, scripts, configuration files, logs, lightweight demo assets, or other review artifacts when allowed by current rules.

NeurIPS-specific guardrails

• The supplement does not rescue an unclear main claim. Reviewers are not required to read every supplemental detail.
• Supplementary material must be anonymized during double-blind review.
• Links must preserve anonymous browsing and comply with current policy.
• Code and data that support central claims should include exact commands and environment details, not just a repository dump.

Placement decision table

NeurIPS assembles the final PDF from distinct zones plus a separate review ZIP. Misplacing material is a recurring weakness: a proof that a reviewer cannot find, or a contribution buried where it does not count. Map each item to exactly one home before submission; reopen the current template for the authoritative zone list.

Datasets & Benchmarks supplement note

For a Datasets & Benchmarks track submission, the datasheet, license, intended-use statement, and hosting/maintenance plan are not optional appendix garnish; they are part of what reviewers score. Keep them anonymized for review and stage a de-anonymized public version for release. Hedge the exact required documentation to the current year's track call.

Worked vignette: a benchmark paper

Authors put their 40-page datasheet, croissant metadata, and loading code only in the appendix and leave the main body to leaderboard tables. A NeurIPS reviewer cannot judge dataset quality from tables alone. The fix: lift provenance, licensing, consent, and intended-use sentences into the main body so the contribution is visible; keep the full datasheet and loaders in the appendix and ZIP; and add one navigation sentence telling reviewers exactly where the documentation lives.

Output format

[Main-paper must stay] <items>
[Move to appendix] <items>
[Move to ZIP/artifact] <items>
[Anonymity fixes] <items>
[Reviewer navigation] <one sentence telling reviewers where to find support>