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
- Factual errors in reviews (wrong dataset, misread table, missed section).
- Misunderstandings a two-sentence clarification can dissolve.
- Requests satisfiable from submitted material — appendix tables, checklist
answers, the supplement archive.
- Requests needing new experiments: promise only what is genuinely small, or
explicitly defer to a revision cycle (ARR is built for resubmission).
- 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
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
-
Day 1: read all reviews twice; extract every distinct objection into
a sheet with owner and evidence anchor; no drafting yet.
-
Day 2: run any genuinely small verification (a re-scoring, a missing
statistic already computable from logged outputs).
-
Day 3: draft all replies; lead each with its one-sentence resolution.
-
Day 4: internal cold read by a co-author who did not draft; cut 30%.
-
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
[Response stance] clarify-in-cycle / concede-and-revise / mixed
[Meta-review target] <sentence you want the AC to write>
[Per-reviewer plan] <R1/R2/R3: main point + evidence anchor>
[Review-issue flags] <deficient-review reports to AC, if any>
[Deferred to resubmission] <experiments promised for a later cycle>