NeurIPS writing must make a technical contribution legible to reviewers outside the authors' exact subfield. Good style is precise, evidence-calibrated, and checklist-aware.
The mandatory NeurIPS Paper Checklist asks each "yes" answer to point at where the paper supports it. Treat checklist items as a prose target, not a postscript: every "yes" should already have a sentence reviewers can find. Drafting against these anchors removes most "claim broader than evidence" comments before review.
| Checklist theme | Sentence the body should already contain |
|---|---|
| Claims match results | "We show X under assumptions A; we do not claim X holds when B." |
| Limitations | a labeled limitations paragraph naming failure modes and unscaled regimes |
| Theory assumptions | each theorem states its assumptions and forward-references the proof |
| Reproducibility | one sentence pointing at code, configs, seeds, and compute |
| Broader impact | concrete, non-boilerplate discussion of misuse or societal effect, not "no impact" |
| Datasets & Benchmarks | provenance, license, consent, and intended-use sentences for any released data |
The broader-impact statement is a NeurIPS norm, not a generic ethics paragraph: hedge to the current year's checklist for exact wording and required placement.
| Reviewer comment | Root cause in prose | NeurIPS-specific fix |
|---|---|---|
| "Claims exceed evidence" | adjectives where a baseline should be | scope the claim to the tested regime; cite the table |
| "Where is the limitation?" | limitations hidden or absent | a visible labeled paragraph reviewers can reward |
| "Impact statement is boilerplate" | generic "no societal impact" | a specific misuse/benefit tied to this method or dataset |
| "Contribution type unclear" | intro reads as generic ML | name the declared type and what evidence it demands |
| "Novelty over a topic, not a mechanism" | abstract sells a label | state the technical delta against the closest method |
A paper proposes a curvature-aware optimizer and writes "our method achieves state-of-the-art on language modeling." A NeurIPS reviewer reads this as an unbounded claim. Rewrite as: "On three autoregressive benchmarks at the 1B-parameter scale, our optimizer reaches the validation loss of tuned AdamW in 30% fewer steps; we do not test beyond 1B or on vision, and the gain narrows under aggressive learning-rate schedules (Appendix C)." The rewrite names the comparison set, the scale boundary, the failure regime, and forward-references evidence, pre-answering the claims-match-results and limitations checklist items in one move.
[Claim before] <original claim>
[Claim after] <NeurIPS-calibrated claim>
[Evidence anchor] <experiment/proof/ablation/dataset/checklist>
[Limitation sentence] <plain limitation text>
[Reviewer concern reduced] <novelty/clarity/significance/reproducibility/ethics>