Agent Skills
› brycewang-stanford/Awesome-Journal-Skills
› recsys-review-process
recsys-review-process
GitHub解析ACM RecSys双盲评审流程，涵盖三轮审稿、Senior PC统筹及反驳阶段。分析可复现性危机对评估的影响，强调离线指标与在线部署的关联性及基线公平性，指导作者优化策略以获录用。
Trigger Scenarios
解释RecSys同行评审机制
制定论文反驳策略
评估可复现性与基线公平性
理解Senior PC决策逻辑
Install
npx skills add brycewang-stanford/Awesome-Journal-Skills --skill recsys-review-process -g -y
SKILL.md
Frontmatter
{
"name": "recsys-review-process",
"description": "Use when explaining or planning around ACM RecSys peer review — the mutually-anonymous model with at least three PC members and a Senior PC overseer, the rebuttal phase, how the reproducibility-crisis culture shapes reviewer priorities, the offline-versus-online evaluation lens, and how acceptance leads to ACM Digital Library publication."
}
RecSys Review Process
Use this to reason about review-stage strategy. Reopen the current Call for Contributions, the committees page, and any reviewer guidelines before making process claims — mechanics are cycle-specific.
Process model
- RecSys review is mutually anonymous (double-blind). Each submission is read by at least three PC members and overseen by a Senior PC member who synthesizes the recommendation.
- There is an author rebuttal phase (2026: June 4-9) for a short clarifying narrative.
- Reviewers weigh recommendation novelty, evaluation validity, reproducibility, clarity, and relevance to the recommender community — not raw metric wins alone.
- The most useful reply is a decision-focused clarification that gives the Senior PC a clean rationale for acceptance or rejection.
- Accepted papers are published in the ACM Digital Library, so camera-ready compliance and metadata matter as much as the initial decision.
Who reviews here, and what they distrust
- The pool is a single-domain recommender community: expect at least one reviewer who has internalized the field's reproducibility debate and will probe baseline tuning line by line.
- Because RecSys is topically tight, matches are close and an under-tuned comparison or a leaky split gets caught rather than skimmed past.
- Borderline offline-evaluation papers usually fall on one of three edges: baselines tuned less hard than the proposed method, a random split where a temporal one was needed, or an offline metric asserted to imply a deployment win with no bridge.
Scoring leverage table
| Review dimension | What raises it | What sinks it |
|---|---|---|
| Recommendation novelty | A named user/item modeling or evaluation idea | An architecture swap with no recommendation-specific insight |
| Evaluation validity | Equal-budget baselines, temporal split, full-ranking metrics, variance | Untuned baselines, random split, sampled metrics reported as full |
| Deployment relevance | Off-policy estimate, simulator, or A/B result | "Offline nDCG rose, therefore users benefit" |
| Reproducibility | A runnable anonymous repository regenerating the tables | A promise to release code "upon acceptance" only |
| Clarity | One notation source, honest limitations | Buried assumptions; a random-split protocol left implicit |
Stage-by-stage realism
- Initial reviews: triage by what the Senior PC would weigh, not by reviewer tone.
- Rebuttal: the window is short; an early, precise narrative on the central evaluation objection
beats a late line-by-line reply (see
recsys-author-response). - Decision: the Senior PC synthesizes; one unresolved evaluation-validity objection outweighs several resolved clarity complaints.
- Post-decision: ACM Digital Library publication means the rights form and metadata become the final gate.
Vignette: reading a split decision
Two reviewers like the method; one flags that baselines were tuned only on defaults. At RecSys that single objection is decisive because it maps onto the community's reproducibility anxiety — so the rebuttal must resolve that thread, with the equal-budget grid, before polishing anything the other reviewers raised.
Output format
[Current stage] submitted / reviews / rebuttal / decision / camera-ready
[Decision actors] <PC reviewers / Senior PC>
[Likely leverage] <novelty / evaluation validity / deployment relevance / reproducibility / clarity>
[Forbidden moves] <identity leak / unseen new results / unsupported deployment claims>
[Next response move] <one action>
Version History
- 9f86f09 Current 2026-07-19 17:28
