brycewang-stanford AISTATS Writing Style
Use this when revising the main paper. AISTATS papers need compact statements of why a statistical or ML contribution matters and enough detail for technical validation.
Revision rules
- Put the AI/statistics contribution in the first page: problem, gap, method or theorem, evidence, and why existing methods are insufficient.
- Make assumptions explicit. AISTATS reviewers are sensitive to hidden distributional, asymptotic, optimization, or data-generating assumptions.
- Pair every major claim with either proof structure, simulation design, empirical table, or reproducibility detail.
- Use the 8-page body for core logic; move long derivations, extra simulations, and extended ablations to the appendix without making the main paper unintelligible.
- Avoid overclaiming empirical wins when differences are small or variance is unreported.
- Maintain double-blind style in self-citations, prior code references, acknowledgements, funding, and artifact descriptions.
Theorem-presentation discipline
- Number assumptions and cite them by label inside every theorem statement; AISTATS readers audit assumption flows the way software reviewers audit imports.
- Give each major theorem a proof sketch in the body; the appendix proof supports the sketch but never replaces it.
- Keep problem-dependent constants visible or explicitly deferred; silently absorbing dimension or condition numbers into O-notation is a standing statistician complaint.
- Define notation once in one location — two-column pages punish redundant redefinition.
- Phrase remarks after theorems as interpretation, not as extra unproved claims; AISTATS reviewers treat every declarative sentence near a theorem as something they may verify.
- When a result is conjectured or only empirically supported, label it so; mixing proved and observed statements in one paragraph is a credibility leak at this venue.
Sentence-level rewrites
| Draft pattern | AISTATS-safe rewrite |
|---|---|
| "Our method significantly outperforms..." | "reduces mean error by X (SE Y) over Z seeded runs" |
| "Under mild conditions..." | "Under Assumptions 1-3 (boundedness, ...)" |
| "It is easy to see that..." | "By Lemma 2 and the triangle inequality..." |
| "Achieves state-of-the-art..." | Claim scoped to the regimes actually tested |
Vignette: compressing into eight two-column pages
A draft with three theorems, six figures, and a sprawling related-work section: keep all theorem statements, one sketch each, and the two decision-critical figures (the rate plot and the coverage plot); compress related work into contribution contrasts; move secondary lemmas and four figures to the appendix with explicit forward references. The test of a good cut: a reviewer should reconstruct the whole argument without ever opening the supplement.
Output format
[Writing diagnosis] clear / under-justified / overclaimed / overloaded
[First-page fix] <new framing>
[Claim discipline] <claim -> proof/experiment/limitation>
[Compression cuts] <move/delete/merge>
[Anonymity edits] <phrases to rewrite>