Wsdm supplementary

WSDM Supplementary Material

Allocate content across WSDM's three containers. The defining constraint, verified for the 2026 edition: appendices are inside the 9-page cap, and only references plus the required ethical-considerations section sit outside it (2025: 8 pages + at most 2 for references-plus-ethics). WSDM has kept a famously lean budget since its early editions, and no separate supplementary-upload channel was verified for current cycles (待核实 - if the current CFP adds one, re-plan around it). So "put it in the supplement" is not an available sentence here; the real containers are:

Container	Counted?	Reviewer obligation	Right contents
Body (≤9 pp incl. appendices, 2026)	Yes	Read	Claims, method, all decisive evidence
References	No	Consulted	Complete citations - never compressed
Ethics section	No	Read as ethics	Societal impact, privacy, misuse, mitigations
Cited anonymous repository	No	Voluntary	Code, configs, raw results, full prompts, extra ablations

Triage rule: decisive vs supporting vs archival

• Decisive evidence - anything a reviewer needs to believe a headline claim (main comparisons, the ablation that isolates the mechanism, the bias-control analysis) - must be in the body. A decisive table living only in the repo is invisible: consulting it is voluntary and many reviewers will not.
• Supporting material - per-dataset breakdowns, sensitivity sweeps, secondary metrics - goes to a compact in-budget appendix if space remains, else to the repository with a one-line pointer stating what is there.
• Archival material - raw logs of runs, full hyperparameter grids, prompt transcripts - always repository.

The failure mode to catch in review-of-the-draft: a 9th page spent on a supporting sensitivity sweep while the decisive ablation is squeezed into a footnote. Rank content by claim-criticality, not by writing order.

Compression tactics that keep meaning

1. Merge, don't shrink, tables. Two 6-column tables over the same systems become one 8-column table; dropping font size below the class minimum is a format violation, merging is craft.
2. Plot the sweep, table the winner. A hyperparameter study is one small figure; the chosen values go in one row of a settings table; the full grid goes to the repo.
3. State theorems, sketch proofs. WSDM papers with formal results give the statement plus a 3-5 line proof sketch in the body and the full proof in an in-budget appendix only when it is short; otherwise the repo carries a PDF note - and the claim sentence must not overreach what the sketch shows.
4. Prompts by exemplar. For foundation-model components, show one representative prompt inline (verbatim, boxed) and point to the repo for the full set; paraphrased prompts are unreproducible.
5. Ethics section carries ethics. Resist parking experimental overflow there; the SPC reads it and the move is transparent.

The pointer pattern

Every externalized item gets an explicit, specific in-body pointer so reviewers know evidence exists and was not hidden:

% Good: specific, verifiable, sets expectations
Full per-locale results (12 markets), the complete prompt set, and the
seed-level metric dumps behind Tables~2--4 are in the anonymous
repository \cite{anonrepo}, directory \texttt{results/}.

% Bad: vague hand-wave reviewers rightly distrust
Additional experiments are available in our repository.

And the repository must deliver: a pointer to a directory that does not exist, or to results that mismatch the paper's numbers, is worse than no pointer.

Budget worksheet

Run this a week before the deadline, not the night of:

page budget (2026 anchor)        9.00
- title/abstract/intro           1.25
- related work                   0.75
- method                         2.50
- experimental setup             1.25
- results + ablations            2.50
- discussion/limitations/concl.  0.50
- in-budget appendix             0.25
                                 ----
head-room                        0.00   <- if negative, cut supporting
references                       uncounted
ethics section                   uncounted (but write it properly)

If head-room is negative, cut in this order: supporting appendix -> secondary metric columns -> related-work prose (compress to positioning, see wsdm-related-work) -> introduction repetition. Never cut: setup details that gate reproducibility, limitations, or the decisive ablation.

Edge cases

• Short-track papers (track introduced in 2026) compress the same triage into a smaller budget: there is no room for any in-budget appendix at all, so the split is binary - body or repository. Verify the short-paper cap in the current call; the pack could not pin the 2026 number (待核实).
• Demonstrations were capped at 4 pages inclusive of references in 2026
  • a different counting rule than the long track in the same year. Never assume counting rules transfer across tracks of the same edition.
• Theory-heavy papers: when full proofs genuinely cannot fit in budget, the honest structures are (a) a shortened conference claim whose proof does fit, or (b) a companion technical-report/preprint cited openly - checking first that the current anonymity rules permit citing it without breaking blinding (an anonymized preprint pointer in the repo is the safe default).
• Cross-edition drift: the 2025→2026 change moved references and ethics from a shared 2-page allowance to unrestricted space while pulling appendices inside the cap. Assume the containers change again; re-derive this skill's table from the current CFP before applying it.

Output format

[Container map] body / references / ethics / repo allocation per artifact
[Decisive-in-body check] all headline-claim evidence in the 9 pages: yes / no
[Budget] worksheet head-room: +/- N pages; cut list if negative
[Pointers] each external item has a specific in-body pointer: yes / gaps
[Repo delivery] pointers resolve and numbers match: verified / pending