Agent Skills
› brycewang-stanford/Awesome-Journal-Skills
› recsys-submission
recsys-submission
GitHub用于在截稿前审计ACM RecSys投稿，涵盖多赛道路由、8页含附录格式检查、推荐系统特有的双盲匿名化及身份泄露风险排查，防止因赛道不符或违规被直接拒稿。
触发场景
准备提交ACM RecSys会议论文
检查论文是否符合特定赛道要求
验证双盲匿名化是否彻底
确认页数和格式合规性
安装
npx skills add brycewang-stanford/Awesome-Journal-Skills --skill recsys-submission -g -y
SKILL.md
Frontmatter
{
"name": "recsys-submission",
"description": "Use when auditing an ACM RecSys submission before the deadline — routing between the Main, Reproducibility, Industry and Resource tracks, the ACM two-column content-page budgets with appendices counted inside, double-blind anonymization for recommender papers, the anonymous-repository expectation, and desk-reject triage."
}
RecSys Submission
Run this audit before anything is uploaded to the RecSys submission system. RecSys is a single-community but multi-track venue: the same manuscript can be legal in one track and desk-rejectable in another, so the audit begins with "which track is this PDF for?" and only then checks format. Reopen the current edition's Call for Contributions — every number below is a 2026-cycle anchor (verified 2026-07-09), not a permanent rule.
Pick the track before the format
For 2026 RecSys reorganized its lineup: it dropped the Late-Breaking Results (LBR) poster track and added Research & Practice (R&P) Notes. Route the contribution first.
| Track | 2026 budget | What belongs here |
|---|---|---|
| Main - long paper | 8 content pages | A rounded recommendation contribution with offline (and ideally online) evidence |
| Main - short paper | 4 content pages | One focused, complete finding |
| Main - Past, Present & Future | 4 content pages | Reflective/forward-looking position on a recommendation topic |
| Reproducibility | see current CFP | Repeating, refuting, or re-scoping prior results; datasets/frameworks for future reproduction |
| Industry | see current CFP | Deployed systems, production constraints, live A/B evidence |
| Resource / Dataset | see current CFP | A community dataset or software resource with build methodology |
| R&P Notes (posters) | see current CFP | Short findings/experiences; replaces LBR for 2026 |
References are excluded from the page count; appendices count inside it. Use the ACM two-column template unmodified — no margin, font, or spacing surgery.
Anonymization for recommender papers specifically
Generic double-blind advice misses the leaks recommender papers actually carry:
- Proprietary logs and platform. "Interaction logs from our video platform in region X" can name the company. Keep the statistics, genericize the source.
- Deployed system names. If the ranker already runs in production under a public name, the name de-anonymizes; rename for review or argue from the current policy on preprints.
- The anonymous repository. RecSys expects a code/data link inside the paper rather than a supplement upload — mirror it to an anonymized host and strip the owner, commit authors, and cluster paths from history.
- Self-citations in third person; acknowledgements and funding removed; PDF metadata stripped.
Blocking risks
- Wrong track for the contribution (a deployed-system paper squeezed into a Main long paper, or a reproduction study submitted as a novel-method paper).
- Over budget once appendices are counted inside the limit.
- Identity leak via logs, platform, deployed system name, or repository owner.
- Untuned baselines or a random split on sequential data — not a format desk reject, but the
fastest route to a low score (see
recsys-experiments). - Parallel submission of the same work to another venue.
Desk-reject and triage table
| Trigger | Severity at RecSys | Repair window |
|---|---|---|
| Over budget (appendix counted in) | Desk reject | None after the deadline |
| Template tampering | Desk reject or chair flag | None |
| Identity leak in PDF, logs, or repository | Desk reject | None |
| Wrong track's submission form | Desk reject or forced move | Usually none |
| Parallel/dual submission | Desk reject at any time | None |
| Baselines untuned; random split on session data | Review-stage damage | Only before the deadline |
Final-week sequence for an offline-evaluation paper
- Freeze the split: regenerate every ranking table from the versioned dataset and a temporal split so the PDF and the anonymous repository agree.
- Re-run baseline tuning last under an equal budget; a stale, under-tuned baseline is the classic reviewer catch when the numbers move.
- Compress into the content-page budget by demoting per-dataset analyses to the appendix or repository, never by shrinking floats.
- Anonymity sweep: PDF metadata, acknowledgements, log provenance, deployed-system name, repository owner.
- Confirm the submission-form abstract matches the PDF abstract and the track is correct.
# Pre-upload checks that catch the silent killers
pdfinfo paper.pdf | grep -iE "author|creator" # metadata identity leak
grep -inE "acknowledg|our (platform|service|company)" *.tex # textual / platform leaks
pdftotext paper.pdf - | grep -icE "tuned|grid search|equal budget" # baseline tuning stated?
pdftotext paper.pdf - | grep -icE "temporal split|leave-one-last" # leakage-aware split?
Format anchors
- RecSys uses the ACM two-column proceedings template; dense ranking tables and multi-panel offline/online figures overflow columns fast, so compress early rather than on deadline night.
- The page counts above describe the 2026 cycle; treat every number as provisional and recheck the current Call for Contributions before relying on it.
Output format
[RecSys readiness] Ready / Needs fixes / Wrong track / Not ready
[Track] main-long / main-short / past-present-future / reproducibility / industry / resource / R&P-note
[Budget check] <pages used>/<budget>, appendix-inside-budget confirmed yes-no
[Anonymity risks] <logs / platform / deployed name / repository owner / metadata>
[Evidence risk] <baseline tuning / split leakage / offline-only claim>
[Fix order] <ordered blocking fixes>
版本历史
- 9f86f09 当前 2026-07-19 17:29
