brycewang-stanford SIGIR Supplementary
SIGIR has no appendix escape hatch. The 2025 full-paper wording (the clearest recent statement of the standing rule) caps papers at 9 pages including figures, tables, proofs, appendixes, and acknowledgments — everything except references; the 2026 pages restate the budgets as 9/4 pages excluding references. There is no separate supplementary upload in the NeurIPS sense on the verified track pages. Overflow therefore has exactly three destinations: compression, the cited repository, or the bin. This skill is the triage protocol.
The three-container model
| Container | What belongs there | What must never be there |
|---|---|---|
| Paper body (9 or 4 pp) | Every claim a reviewer must verify to accept: method, main tables, significance results, tuning protocol summary, limitations | Filler related work, per-query dumps, prompt walls |
| References (free) | Complete citations — never trimmed for space | Content smuggled into bibnotes |
| Cited repository | Run files, full configs, per-query analyses, extra collections, prompts, proofs of secondary lemmas, negative results | Anything acceptance depends on |
The governing rule: reviewers must be able to accept the paper without opening the repository. Track pages do not oblige reviewers to consult external material, so evidence that lives only in the repo is evidence that may not exist for review.
What earns its place in the budget
Rank candidate content by review-criticality, not by effort spent producing it:
- The comparison table that carries the main claim, with significance markers.
- The tuning-protocol paragraph (symmetric budgets, dev split) — its absence is a standing SIGIR objection.
- One ablation table that isolates the mechanism.
- The evaluation-setup block: collections, metrics with cutoffs, qrels versions.
- Error/qualitative analysis — one distilled example pattern, not a gallery.
- Efficiency numbers if the claim mentions cost, latency, or scale at all.
What loses its seat first: exhaustive metric grids (report the pre-registered ones; put the grid in the repo), second and third qualitative examples, restated definitions of standard metrics, and architecture recaps of well-known baselines.
Compression before amputation
% Honest space wins that don't touch \vspace (which risks desk rejection):
% 1. Merge setup prose into the table caption.
\caption{nDCG@10 on \colA{} and \colB{}. Paired t-test vs BM25,
Bonferroni-corrected; $\dagger$: $p<0.05$. 3 seeds, mean$\pm$sd.}
% 2. One combined results table with grouped columns beats three tables.
% 3. booktabs + \small in tables (allowed by class) before cutting content.
% 4. Inline enumerations for contribution lists; drop the roadmap paragraph.
Never buy space by deleting the significance-testing sentence, the seed count, or the collection versions — those cuts convert a space problem into a soundness objection, which is a worse trade at this venue than a missing extra ablation.
Repository-as-appendix discipline
If the paper says "details in the repository," the repository must behave like an appendix, not a junk drawer:
- A
PAPER_MAP.mdthat mirrors the paper's section numbering: "§5.2 ablation →eval/ablation_table.md+runs/ablate_*.trec." - Per-query result files for every table, so a skeptical reviewer can drill down.
- The full hyperparameter grid with the selected cell marked.
- Anonymized for double-blind tracks (fresh export, no usernames); the Resources track's single-anonymous regime is the exception, not the rule.
Cross-references must be one-directional: the paper points to the repo; the repo never contains claims absent from the paper (reviewers cannot credit them, and policy-wise unreviewed claims must not ride into the version of record later).
Worked micro-triage
A full-paper draft sits at 10.5 pages against the 9-page budget. Candidate cuts:
- A 0.6-page per-query case-study gallery (three annotated examples) → keep one example that names the diagnosed pattern; the other two move to the repo with their run-file coordinates. (-0.4 pp)
- A 0.5-page secondary-metric grid (five metrics × four collections) → keep the two
pre-committed metrics; grid to
eval/full_grid.md. (-0.4 pp) - A 0.4-page derivation of a variant loss → state the result, cite the repo notebook. (-0.35 pp)
- The tuning-protocol paragraph — never; it pre-answers the venue's most common fatal objection.
- Caption-merge the setup prose and drop the roadmap paragraph. (-0.35 pp)
Net: on budget, with zero soundness loss — every cut demoted redundancy or depth, not protocol.
Short papers: the 4-page special case
A short paper has no room for a supplementary strategy — it is the strategy. Standard shape that fits: one focused question, one main table, one analysis figure, setup compressed into captions, everything else in the repo. If the triage above still leaves review-critical content homeless, the work is a full paper submitted to the wrong track; re-route rather than compress soundness away.
Pre-submission checklist
- Page count within budget with acknowledgments space reserved (camera-ready restores them inside the same budget unless the track says otherwise — 待核实).
- No review-critical evidence lives only in the repository.
- Repository PAPER_MAP matches the submitted section numbering.
- References complete and untrimmed; no content in bibliography notes.
- No spacing/margin/font tampering; only class-sanctioned sizes used.
Output format
[Budget state] <pages>/<limit> with acks reserve y/n
[Cut list] <content demoted to repo, in order>
[Kept-by-criticality] main table / tuning protocol / ablation / setup / analysis
[Repo appendix] PAPER_MAP present y/n; per-query files y/n; anonymized y/n
[Soundness guard] significance sentence, seeds, versions all still in-paper y/n
[Re-route flag] none / "this is a full paper in short clothing"