research-paper-writing

Follow an iterative pipeline: define a one-sentence contribution, map each claim to a supporting experiment, run experiments with incremental result 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.

Fetch BibTeX programmatically via DOI content negotiation from CrossRef or the Semantic Scholar API rather than generating from memory. Verify each paper exists in at least two sources, and mark anything unverifiable as a placeholder citation for manual checking.

Report error bars with the method specified, 95% confidence intervals via bootstrapping, McNemar's test for paired method comparisons, and effect sizes such as Cohen's h. Always state sample sizes and number of independent runs.

Yes, the pipeline covers theory, survey, benchmark, and position papers in addition to empirical ML work. It also includes guidance for human evaluation design, negative results framing, and post-acceptance deliverables like posters and code releases.

ICML allows 8 pages for submission, ICLR and NeurIPS allow 9 pages, ACL long papers allow 8, AAAI allows 7, and COLM allows 9. References and appendices are unlimited at all of these venues, and camera-ready versions typically get one extra page.

AI-generated citations have roughly a 40% error rate, producing fabricated titles, wrong venues, and non-existent DOIs. The mitigation is a mandatory verification workflow: search, verify in two sources, retrieve BibTeX via DOI, validate the claim, then add to the bibliography.