iclr-supplementary

ICLR Supplementary

Use this when deciding what belongs in the main PDF, appendix, supplementary files, or private discussion-period links. The goal is to make extra evidence easy to find without hiding the paper's core argument outside the main text.

Structure

• Keep central method, claims, and minimum evidence in the main text.
• Put derivations, extended ablations, robustness, extra qualitative examples, hyperparameters, model cards, dataset cards, and ethics details in the appendix.
• Use supplementary files for code, data, large tables, logs, demos, and artifacts that do not fit cleanly in the PDF.
• Include a one-page appendix map at the start when the appendix is long.
• Label discussion-period revisions explicitly so reviewers can find changes quickly.

Anonymity and access

• Remove author names, institutions, usernames, Git remotes, file paths, cloud buckets, and license headers that identify the team.
• Avoid links that reveal visitors or owner accounts. Prefer anonymized repositories, static files, or private OpenReview-visible links permitted by the current guide.
• Make supplementary filenames descriptive but neutral.
• If code cannot be released, state the legal or privacy constraint and provide a smaller reproducibility substitute.

Main-text versus appendix placement

ICLR's single-track format means the main PDF carries the argument and appendices come after references with generous room. The risk is burying decisive evidence where a time-pressed reviewer never looks, while padding the appendix with material that proves nothing.

Content	Belongs in	Why
Central claim and its key result	Main text	Reviewers may not open the appendix
Full proof / extended derivation	Appendix	Needed for rigor, not for the headline
Hyperparameters, model/dataset cards	Appendix	Reference material, not argument
Code, logs, large tables, checkpoints	Supplementary ZIP	Too big or interactive for the PDF
Decisive ablation a claim depends on	Main text	Moving it weakens the narrative

Worked vignette

A generative-model paper hides its only fairness-of-comparison ablation in Appendix G, so a reviewer concludes the headline gain is unsupported. The fix: promote a compact version into the main text near the claim, leave the full grid in the appendix, and add an appendix map. During discussion the authors upload a revision and label the changelog so the relocation is obvious.

Reviewer-pushback patterns

• "Key evidence is buried." Promote it; an appendix map is not a substitute for main-text placement.
• "Cannot tell what changed in the revision." Add a dated changelog at the top of the revised PDF.
• "Supplement leaks identity." Re-check file paths, license headers, and remotes before upload.

Output format

[Supplement plan] PDF appendix / ZIP / anonymous repo / private link / no supplement
[Main-text dependencies] <claims that must move back into main text>
[Reviewer navigation] <appendix map and file names>
[Anonymity risks] <paths, metadata, links, licenses>
[Discussion update note] <short changelog if revised>