--- name: brycewang-stanford/acl-author-response source: https://app.decimal.ai/s/brycewang-stanford-acl-author-response@1/SKILL.md source_sha256: 6e6fca1e3fbd --- # ACL Author Response Use this when ARR reviews land. The response is written inside the review cycle, *before* the meta-review is finalized — its primary reader is the area chair who will summarize your paper for whatever conference you later commit to. Reopen aclrollingreview.org/reviewing for the current cycle's exact window and rules. ## What the response can actually move - ARR assigns at least three reviewers; after the author response there is a reviewer-discussion phase in which reviewers are expected to update reviews. A response that gives a reviewer a concrete reason to revise is worth more than one that argues. - The meta-review travels with the paper to commitment. Senior area chairs at the venue read the meta-review first, so your real goal is to shape what the AC writes, not to win a debate transcript. - Scores are not the only currency: a meta-review that says "the main weakness was resolved in discussion" can rescue a committed paper that scores alone would sink. ## Triage order 1. Factual errors in reviews (wrong dataset, misread table, missed section). 2. Misunderstandings a two-sentence clarification can dissolve. 3. Requests satisfiable from submitted material — appendix tables, checklist answers, the supplement archive. 4. Requests needing new experiments: promise only what is genuinely small, or explicitly defer to a revision cycle (ARR is built for resubmission). 5. Taste disagreements: answer once, briefly, without heat. ## NLP-reviewer objection patterns | Objection | Underlying worry | Response that works at ACL | |---|---|---| | "Only evaluated on English" | Overclaimed generality | Scope the claim explicitly; cite the multilingual appendix run if one exists | | "Gains may be contamination" | Test data in pretraining | Point to the decontamination check or concede and describe the planned audit | | "No error analysis" | Numbers without understanding | Quote the analysis section or summarize 2-3 systematic error classes from logged outputs | | "Human eval lacks agreement stats" | Unreliable judgments | Report the IAA already computed; never invent a number in the response box | | "Missing recent baseline X" | Stale comparison set | Distinguish X technically, or accept it as revision work with a stated plan | ## Rules of the response box - Stay anonymous: no names, lab hints, grant numbers, or identifying links. - Never claim results that are not in the submitted PDF, supplement, or checklist; ARR reviewers can and do check. - Keep each reviewer reply skimmable — lead with the single sentence you want quoted in the meta-review. - If a review is deficient (template text, wrong paper, LLM-generated, hostile), use the current cycle's review-issue flag to the AC rather than fighting it in the reply thread. ## Respond now or resubmit later - Respond in-cycle when the objections are misreadings or scoping problems — these are exactly what discussion can fix before the meta-review. - Prefer withdrawing the commitment plan and revising for a later cycle when reviews identify real missing experiments; a resubmission must carry a point-by-point revision note anyway, so bank the work. - Remember the asymmetry: reviews and meta-review persist and follow a resubmission. An abrasive response is permanent record. ## Skeleton reply ```text We thank Reviewer X. Main point first: [one-sentence resolution]. 1. [Factual correction] The model is trained only on the training split (Section 3.2, line refs); Table 2 reports held-out results. 2. [Clarification] The claim is scoped to the five languages tested (Section 5); we will tighten the abstract wording. 3. [Evidence pointer] Seeds, prompts, and compute are in Appendix C and checklist items C1-C3. 4. [Deferred work] A comparison to X requires retraining; we commit to it in a revision and note why it should not change the ranking. ``` ## Tone calibration for the ARR record - Write for three audiences at once: the reviewer being answered, the AC composing the meta-review, and the venue chairs who will skim the thread months later at commitment. - Concede fast and specifically; a clean concession plus a scoped fix reads as competence, while minimizing a real weakness reads as evasion. - Quantify wherever a reviewer hedged: replace "we believe the effect is robust" with the interval, the seed count, or the appendix table number. - Do not thank effusively or argue emotionally; both waste the reader's attention budget in a thread they are legally required to skim. - If two reviewers contradict each other, say so neutrally and let the AC arbitrate — pointing out the contradiction is legitimate and effective. ## A workable 5-day response window 1. **Day 1**: read all reviews twice; extract every distinct objection into a sheet with owner and evidence anchor; no drafting yet. 2. **Day 2**: run any genuinely small verification (a re-scoring, a missing statistic already computable from logged outputs). 3. **Day 3**: draft all replies; lead each with its one-sentence resolution. 4. **Day 4**: internal cold read by a co-author who did not draft; cut 30%. 5. **Day 5**: post early, not at the buzzer — reviewers who will engage in discussion tend to do it once, soon after your reply appears. ## If the real answer is "withdraw and revise" - Skipping the response entirely is a signal too — reviewers note silence in discussion, and the package records it. A short, gracious reply that acknowledges the substantive points costs little and keeps the record clean for the resubmission's mandatory revision note. - Harvest the reviews immediately into the revision plan while the team still has context; the next cycle's point-by-point response is easiest to write as a running document, not a deadline reconstruction. ## Output format ```text [Response stance] clarify-in-cycle / concede-and-revise / mixed [Meta-review target] [Per-reviewer plan] [Review-issue flags] [Deferred to resubmission] ```