research-paper-writing

Follow a phased pipeline: define a one-sentence contribution, map each claim to a supporting experiment, run baselines and ablations with statistical significance tests, then draft using the venue's official LaTeX template. Complete the mandatory paper checklist before submission to avoid desk rejection.

Never generate BibTeX from memory — fetch it programmatically via DOI content negotiation from CrossRef or the Semantic Scholar API. Verify each paper exists in at least two sources and confirm the cited claim actually appears in the paper, marking unverifiable entries as placeholders.

Use McNemar's test for paired binary outcomes on the same problems, bootstrapped confidence intervals for aggregate metrics, and Cohen's h for effect sizes on proportions. Always report sample sizes, number of runs, and whether error bars show standard deviation or standard error.

Yes, it includes official LaTeX templates and requirement guides for six venues: NeurIPS, ICML, ICLR, ACL, AAAI, and COLM. Each has documented page limits, formatting rules, and mandatory sections such as ACL's Limitations section and ICLR's LLM disclosure policy.

Use human evaluation when automated metrics cannot capture your claims, such as fluency, helpfulness, or preference in generation tasks at NLP venues. Design it early with pairwise comparisons, at least 3 annotators, and report Krippendorff's alpha for inter-annotator agreement.

Iterative refinement fails on unconstrained tasks with frontier models due to synthesis drift, and weak models cannot generate diverse enough candidates to benefit. It works best with mid-tier models on constrained tasks, using a ground-truth critic with access to actual experimental data.