Agent Skills
› brycewang-stanford/Awesome-Journal-Skills
› recsys-author-response
recsys-author-response
GitHub用于在ACM RecSys论文作者回复阶段起草反驳信。指导如何基于已提交证据澄清误解，应对基线调优、数据泄露及离线在线差距等常见质疑，强调匿名性、引用具体附录/仓库位置，并遵循先纠正事实后解释影响的写作模式。
Trigger Scenarios
需要撰写ACM RecSys会议论文的反驳信
收到关于基线未充分调优的审稿意见
面临数据泄露或离线指标与线上效果脱节的质疑
Install
npx skills add brycewang-stanford/Awesome-Journal-Skills --skill recsys-author-response -g -y
SKILL.md
Frontmatter
{
"name": "recsys-author-response",
"description": "Use when drafting an ACM RecSys rebuttal during the author-response window, covering the short narrative RecSys allows, anonymity, anchoring answers to submitted evidence and the anonymous repository, the recurring objections about untuned baselines, leakage, and offline-online gaps, and Senior-PC-focused clarification."
}
RecSys Author Response
Use this after RecSys reviews are released and the rebuttal window opens. RecSys runs a brief author-response phase (2026: June 4-9) in which authors add a short narrative to clarify misconceptions. Reopen the current instructions before drafting — the length and mechanics are cycle-specific and the phase is short.
Triage
- Answer concerns that move the decision: correctness, recommendation novelty, evaluation validity, reproducibility, and fit.
- Use already-submitted evidence — paper sections, tables, ablations, the appendix, and the anonymous repository. The rebuttal clarifies; it does not add a new experiment the reviewers cannot see.
- Keep the reply anonymous: no institution, platform, deployed-system name, or repository owner.
- Correct factual misreadings first, then address requested comparisons, tuning details, or uncertainty that the paper already supports.
- Do not paste new result tables the reviewers had no chance to vet; point to what exists.
Drafting pattern
- State the decision-critical correction or concession in the first line.
- Point to the exact submitted evidence (table, figure, appendix section, repository path).
- Explain the recommendation consequence — what it means for the ranking claim or deployment.
- Promise a camera-ready wording fix only where it adds no unsupported new claim.
Recommender-reviewer pushback patterns
| Pushback | What it signals | RecSys-ready fix |
|---|---|---|
| "Were the baselines tuned as hard as your method?" | The reviewer read the 2019 reproducibility literature | Point to the equal-budget tuning grid in the appendix and the per-baseline selected settings; never wave it off as "standard config" |
| "This is a random split on sequential data" | Suspected temporal leakage | Confirm the temporal/leave-one-last protocol and where it is stated, or concede and scope the claim |
| "Offline nDCG gains may not deploy" | The offline-online gap objection | Anchor to the off-policy estimate, the simulator, or an A/B result; if you have none, scope the claim to offline |
| "Metrics look sampled, not full-ranking" | Sampled-metric distortion suspected | State whether ranking is over the full item set and the cutoff, or report the full-ranking number |
| "Missing the obvious matrix-factorization baseline" | A weak-comparison gap | Anchor to the appendix comparison, or concede a camera-ready addition |
Response micro-example
Reviewer objection: the Recall@20 gain probably vanishes against a properly tuned baseline. Reply skeleton:
- Concede the concern is the right one to raise.
- Point to Appendix C, where every baseline is tuned on the same grid with the same budget, and the selected hyperparameters are listed.
- Note the gain persists after tuning (Table 1) and the ablation (Table 2) isolates the mechanism.
- Offer one camera-ready sentence making the equal-budget protocol explicit in the main text.
Rebuttal calibration
- One decision-critical point per reviewer beats an exhaustive line-by-line reply; the Senior PC reads for whether the central evaluation objection was actually resolved.
- Reply early — RecSys rebuttal windows are short and a precise early narrative outweighs a late comprehensive one.
- Length and formatting norms vary by cycle; recheck the current author-response instructions before sending.
Output format
[Priority issue] <reviewer concern>
[Decision dimension] correctness / novelty / evaluation validity / reproducibility / fit
[Draft response] <RecSys-ready anonymous narrative>
[Evidence anchor] <table / figure / appendix / repository path>
[Forbidden content removed] <identity leaks, unseen new results, unsupported claims>
Version History
- 9f86f09 Current 2026-07-19 17:28
