The audience of a response is the editor or area chair, who must decide quickly whether concerns were addressed. Make every reply easy to verify: quote the concern, answer it directly, show the change with a location.
Step 1: Triage (before writing anything)
Split every review into atomic comments and classify each in a table:
| ID | Reviewer | Comment (short) | Type | Effort | Plan |
|---|---|---|---|---|---|
| R1.1 | 1 | Missing baseline X | Experiment | High | Run X on 2 datasets |
| R2.3 | 2 | Claim in §4 overstated | Wording | Low | Soften + add limitation |
Types: misunderstanding (the text was unclear), missing analysis/experiment, wording/presentation, disagreement on substance, out of scope, factual error by reviewer.
Then:
- Identify the 2-3 concerns that drive the scores (often shared across reviewers). These get the most effort and appear first in a conference rebuttal.
- Decide what is feasible in the revision window. Never promise experiments you will not run.
- Note conflicts between reviewers so the editor can adjudicate.
Step 2: Do the work, then write
Run the experiments and edit the manuscript before drafting replies, so every reply can point to a concrete change. Keep a change log with section/page/line numbers.
Step 3: Reply patterns
Journal response letter (no strict length):
Reviewer 1, Comment 1: "<quote or faithful summary of the concern>"
Response: We thank the reviewer for ... [direct answer in the first sentence].
We have [action]. The new results (Table 3, p. 7, lines 210-224) show ...
Change in manuscript: "<quoted new/revised text>" (Section 4.2, p. 7)
Best practices:
- Answer in the first sentence ("Yes, we have added...", "We agree and have...", "We respectfully disagree because...").
- Thanks once per reviewer, not before every comment.
- Quote revised text so the reviewer need not hunt for it. Provide a tracked-changes/diff version if the journal allows.
- Number comments (R1.1, R1.2 ...) and cross-reference instead of repeating ("see our response to R2.1").
Disagreeing (it is allowed and expected when justified):
- Acknowledge the underlying concern ("We understand the concern that ...").
- Give evidence: data, citations, a small additional analysis, or a precise explanation.
- Make a concession where honest: a clarification in the text or a stated limitation. Never be sarcastic, never say the reviewer "failed to understand". Rewrite the text so the next reader won't.
Conference rebuttal (hard limits: characters/words per reviewer or one page; check the venue):
- Lead with a short general response covering shared concerns and new results.
- Then per reviewer, only the comments that affect the decision, 2-5 sentences each.
- Put new numbers in compact tables. State exactly what will change in the camera-ready.
- Correct factual misreadings politely with pointers ("Section 3.2, Eq. 4 defines ...").
- Respect the venue's rules on new experiments, links and supplementary material. Many prohibit external links or new PDFs during rebuttal.
- On OpenReview discussions: reply promptly and concisely, and summarize the resolved points for the AC near the end.
Step 4: Cover note to the editor (journals)
Summarize the main changes in 3-6 bullets, flag any reviewer disagreements you could not resolve and why, and confirm all co-authors approved the revision.
Quality checklist
- Every comment has a response. Count them against the triage table.
- Every "we have added/changed" is actually in the manuscript at the cited location.
- New numbers match between response, manuscript, tables and abstract.
- Tone: professional, specific, never defensive. Read it as the reviewer would.
- No new claims beyond what the revised evidence supports.
- Length within the venue limit (count it).
Related skills
peer-review (reviewing others' work: useful to anticipate criticism), scientific-writing,
ml-paper-writing, scientific-critical-thinking, statistical-analysis.