Sigir Submission

SIGIR Submission

Run this audit before anything is uploaded to a SIGIR OpenReview group. SIGIR is a multi-track venue where the same manuscript can be legal in one track and desk-rejectable in another, so the audit starts with "which track is this PDF for?" and only then checks format. Reopen the current edition's track page — every number below is a 2026-cycle anchor (verified 2026-07-08, exact deadlines 待核实), not a rule.

The two-format trap

SIGIR's full and short formats are not a long and a short version of the same bar. A 4-page short paper is a complete argument about one focused finding; a 9-page full paper is expected to carry a rounded evidence program (multiple collections, ablations, significance tests, efficiency notes). The most common self-inflicted wound is submitting a compressed full paper as a short paper — reviewers see missing evidence, not admirable focus.

The appendix rule is the sharpest edge on this table: unlike NeurIPS-style venues, SIGIR gives you no unlimited appendix after the references. Proofs, prompt listings, extra tables — either they fit in the budget or they move to the cited repository.

Anonymization for IR papers specifically

Generic double-blind advice misses the leaks IR papers actually have:

• The system name. If your ranker is already public under the same name (arXiv, a leaderboard, a plugin for Anserini/PyTerrier), the name itself de-anonymizes. Rename for review or argue from the current policy page that preprints are allowed.
• Leaderboard submissions. MS MARCO / TREC-style leaderboards show team names and dates. Do not cite your own leaderboard row in a way that only the submitting team could.
• Proprietary query logs. "Queries sampled from our commercial search engine's logs in country X" can identify the company. Genericize the description, keep the statistics.
• Self-citations in third person; repository links anonymized (e.g., an anonymous.4open.science mirror); PDF metadata stripped.

Policy gates (2026, source: submission-policies page)

• Parallel submission = desk reject, regardless of how the other venue decides.
• Cross-track double submission inside SIGIR is banned — e.g., the dataset you submit as a Resource paper cannot also be the novel contribution of your full paper. Decide which track owns each contribution before the abstract deadline.
• Generative-AI use must be disclosed; AI tools cannot be authors; SIGIR reserves automated compliance checks. Assume your PDF will be machine-screened.
• ACM authorship policy applies: the author list is expected to be stable and human.

Desk-reject triage

Final-week sequence for a retrieval paper

1. Freeze runs: regenerate every table from versioned run files (TREC-format runs/*.trec) so the PDF, repository, and any leaderboard entry agree.
2. Re-run the significance script last — copy-pasted p-values from stale runs are a classic reviewer catch when the numbers moved.
3. Compress into the page budget by demoting per-query analyses to the repository, not by shrinking floats or tampering with sigconf spacing.
4. Anonymity sweep: PDF metadata, acknowledgments, repo URL, system name, log provenance.
5. Upload to the correct track group, complete the AI-disclosure and PC-nomination fields, and verify the OpenReview abstract matches the PDF abstract.

# Pre-upload checks that catch the silent killers
pdfinfo paper.pdf | grep -iE "author|creator"        # metadata identity leak
grep -inE "acknowledg|our (system|engine|company)" *.tex  # textual leaks
grep -c "documentclass\[sigconf" main.tex             # right ACM class
pdftotext paper.pdf - | grep -icE "p ?[<=] ?0\.05|t-test|significan"  # stats present?

Abstract-deadline gotchas

The abstract deadline (a week before the PDF in recent cycles) is a harder gate than teams expect:

• Title, author list, and track are effectively frozen there; ACM authorship policy makes later author changes painful and cross-track moves are usually impossible.
• The abstract text should be the PDF's abstract, not a placeholder — reviewers bid on it, and a bait-and-switch abstract draws mismatched expertise.
• Conflicts-of-interest declarations and the PC-member nominee belong to this stage; chasing a nominee's consent during PDF week is a self-inflicted crisis.

What this audit does not cover

Evidence quality lives in sigir-experiments; page-budget writing strategy in sigir-writing-style and sigir-supplementary; track choice in sigir-topic-selection. If the paper is a dataset or a replication, stop here and re-route: Resources and Reproducibility have different budgets, different anonymity regimes (Resources is single-anonymous), and different dates.

Output format

[SIGIR readiness] Ready / Needs fixes / Wrong track / Not ready
[Track + group] full / short + OpenReview group id
[Budget check] <pages used>/<budget>, appendix-inside-budget confirmed yes-no
[Anonymity risks] <system name / leaderboard / logs / metadata / links>
[Policy gates] parallel-submission, cross-track, AI-disclosure, PC-nomination status
[Fix order] <ordered blocking fixes>
[Reopen first] <current track page URL to re-verify dates and budgets>

Source: brycewang-stanford/Awesome-Journal-Skills → SIGIR-Skills/skills/sigir-submission/SKILL.md