Ml Conference Paper Architect

ML Conference Paper Architect

Formatting and structural conventions for NeurIPS, ICML, ICLR, and AAAI submissions. Sits on top of latex-architect, which owns general LaTeX practice — build, preamble order, floats, math, semantic markup. This skill owns only what is venue-specific.

Venue template sources and the classic-look preamble live in assets/VENUE_TEMPLATES.md; worked skeletons in RECIPES.md; pinned template years in STACK.md.

1. The one rule that overrides everything

Read the current year's Call for Papers and author kit before applying anything below. Every number in this skill — page limits, deadline structure, checklist requirements — is revised annually, sometimes substantially. This skill encodes the shape of the requirements and the failure modes; the CFP encodes the current values. Where the two disagree, the CFP wins, and the fix belongs in STACK.md.

Corollary: never hand-copy a preamble from a previous paper. Download the official style file for the target year. Reusing last year's .sty is the single most common cause of a paper that looks right locally and gets flagged by the venue's format checker.

2. Before drafting: confirm

1. Venue and year — determines the .sty, the page limit, and the checklist.
2. Track — main conference, datasets & benchmarks, workshop, or findings-style track. Limits and review criteria differ; workshop tracks are usually 4 pages and often not anonymous.
3. Stage — submission (anonymous), camera-ready (de-anonymized, +1 page at most venues), or preprint (arXiv).
4. Whether code and data will be released — this changes what the checklist and reproducibility statements can honestly claim, and it is not a question to answer at the deadline.

3. Venue matrix

Verify each row against the current CFP — see §1.

AAAI is the outlier. Its style file bans packages the others tolerate (geometry, fullpage, authblk, and historically hyperref), mandates Type 1 fonts and an embedded \pdfinfo block, and runs an automated format checker that rejects on violation. Read AAAI's author kit in full; do not assume NeurIPS habits transfer.

4. Anonymization

Double-blind failures are desk rejects, not review-score penalties. The submission-mode style file handles author blocks automatically — everything else is on you:

• No author names, affiliations, emails, or acknowledgements in submission mode. Acknowledgements go in a section you add at camera-ready.
• Cite your own prior work in the third person. "Prior work [12] showed…", never "In our earlier work [12] we showed…" — the latter deanonymizes even with the citation intact.
• Anonymize links. No GitHub URLs with your username, no personal-domain project pages. Use an anonymized repo host, and check the repo itself for names in the README, LICENSE, and commit history.
• Strip PDF metadata. The compiled PDF carries author fields from your editor or OS. Verify with pdfinfo main.pdf before uploading.
• Check figure provenance — plots exported from a notebook can embed a file path containing your name or institution.
• Know your venue's arXiv policy. ICLR/OpenReview submissions are public; NeurIPS and ICML have specific rules on preprinting during the review window. This is a policy question, not a formatting one — read the CFP.

5. Page-limit accounting

The limit applies to main text; what counts as main text is venue-defined. References are typically excluded and the appendix always is, but figures, tables, and the checklist usually are not.

The failure mode is uniform across venues: papers that hit the limit by shrinking. Every one of these is a rejection risk and some are explicit violations:

• Reducing font size, \baselinestretch, or margins from the style file's values.
• Negative \vspace around floats, sections, or equations to claw back lines.
• \small/\footnotesize on body text (captions and tables are fine, and are usually already set by the .sty).
• Shrinking the reference list font or squeezing \bibsep.

Format checkers detect all of these mechanically. Cut content instead — move derivations, extended results, and ablation details to the appendix, which is unlimited. If the paper does not fit, the paper is too long, not the template.

6. Structure

The ML conference paper has a de-facto fixed skeleton. Deviating costs reviewer goodwill for no gain:

1. Abstract — problem, gap, what you did, headline result with a number. Under ~200 words.
2. Introduction — ending in an explicit contributions list. Reviewers look for this; make it a bulleted list of claims the paper actually substantiates.
3. Related work — position against the closest prior work. Either §2 or after the method, depending on whether the reader needs context to understand the approach.
4. Method — notation defined once, then the approach. A figure of the architecture or pipeline earns its space.
5. Experiments — setup (datasets, baselines, metrics, hyperparameters, compute, seeds), then results, then ablations. Missing ablations is the most common reviewer complaint.
6. Limitations — mandatory at NeurIPS, expected everywhere. Write it honestly; reviewers penalize an obviously sanitized limitations section harder than a real weakness.
7. Conclusion — short. No new claims.
8. Broader impact / ethics — per venue requirements.
9. References, then Appendix.

Writing rules that matter more here than in general LaTeX:

• Every claim in the abstract and intro must map to a specific experiment. Reviewers check this correspondence directly.
• Takeaway-led captions. "Figure 3: \model outperforms all baselines beyond 10k steps" beats "Figure 3: Results". Reviewers read figures and captions before the body.
• Report variance. Single-seed numbers with no error bars invite methodological objections regardless of the result.

7. Citations

Venue templates ship natbib with a venue .bst. Keep it — do not convert to biblatex, which is the standing carve-out in latex-architect §8.

• \citep{x} parenthetical, \citet{x} when the author is a sentence element. Mixing these up is visible and looks careless.
• Cite the published version, not the arXiv preprint, when one exists. Bulk-exported .bib files default to the preprint.
• Brace-protect capitals: {B}ayesian, {ImageNet}, {GPT}. Venue .bst files lowercase titles and will mangle these.
• Fix the export's author lists — and others truncation and inconsistent name forms are the usual damage.

8. Reproducibility

Every one of these venues now requires some form of reproducibility artifact, and the answers are checked against the paper body:

• Answer the checklist honestly. "No" with a stated reason is accepted; "Yes" contradicted by the paper is worse than either.
• Report the full experimental setup: hyperparameters, search ranges and selection method, hardware, wall-clock or GPU-hours, number of seeds, and the exact dataset splits.
• Compute cost is a required disclosure at several venues and increasingly a review criterion. Estimate it rather than omitting it.
• Anonymized code at submission, permanent link at camera-ready. A promise to release code later is weaker evidence than a working anonymous repo.

9. Appendix and supplementary

• Appendix goes after references, introduced by \appendix, with sections numbered A, B, … and referenced from the main text via \cref.
• Whether the appendix ships in the same PDF or a separate supplementary file is venue-specific. Check before restructuring.
• The main text must stand alone. Reviewers are not obligated to read the appendix, so no load-bearing claim may depend on it.
• Appendix content is unlimited but not free — a 40-page appendix signals a paper that was not edited.

10. Pre-submission checks

Run all of these before uploading; the list in RECIPES §4 is the executable version.

• Official style file for the correct year, unmodified, in submission (not final) mode.
• Page count within limit, with no spacing or font hacks.
• No author names, affiliations, acknowledgements, or identifying links; PDF metadata stripped.
• All figures legible at print size and in grayscale; no raster text.
• Every citation resolves; no ? marks in the compiled PDF.
• Checklist / impact statement / limitations section present and consistent with the body.
• Fonts embedded (pdffonts main.pdf shows no non-embedded entries) — AAAI enforces this mechanically.
• Compiles from a clean checkout with latexmk and no manual steps.

11. Out of scope

• General LaTeX practice (build, preamble, floats, math, cross-references) → latex-architect
• Journal and institutional formats → uru-scientific-paper-architect, uru-thesis-architect
• Talk and poster design → demo-presentation-architect
• Research content — this skill formats and structures a paper; it does not generate claims, results, or citations. Never fabricate a number, a baseline, or a reference.