research-paper-writing

Follow a phased pipeline: define a one-sentence contribution, map every experiment to a claim, run experiments with incremental saving, compute statistical significance, then draft in the official venue LaTeX template. Finish with a simulated self-review and the venue's mandatory checklist before submission.

Never generate BibTeX from memory—fetch it programmatically via DOI content negotiation from CrossRef. Verify each paper exists in at least two sources such as Semantic Scholar and arXiv, and confirm the cited claim actually appears in the paper before adding it to the bibliography.

Report McNemar's test for paired method comparisons, bootstrapped 95% confidence intervals for key metrics, and effect sizes such as Cohen's h for proportions. Always state sample sizes, number of runs, and whether error bars show standard deviation or standard error.

Yes, NeurIPS requires a completed 16-item paper checklist covering claims alignment, limitations, reproducibility, compute resources, and ethics. Papers missing the checklist face automatic desk rejection, though it appears after references and does not count toward the page limit.

Use autoreason-style iterative refinement for mid-tier models on constrained tasks, where the generation-evaluation gap is largest. Frontier models on unconstrained tasks do better with critique-and-revise or a single pass, since unconstrained iteration causes synthesis drift.

ICML 2026 allows 8 pages, ICLR 2026 allows 9 pages, and NeurIPS 2025 allows 9 pages of main content. All three permit unlimited references and appendices, with camera-ready versions typically gaining one additional page.