brycewang-stanford CVPR Author Response
Use this when CVPR reviews land. In the 2026 cycle, reviews reached authors on January 22 and the rebuttal closed January 29, with reviewer–AC discussion running January 30 – February 5 and decisions on February 20. One week, one page — the format is the strategy.
What the one page legally is
The rebuttal is an optional, single-page PDF built on the rebuttal.tex template in
the official author kit. Two rules from the Author Guidelines carry teeth:
- Responses longer than one page are simply not read, and that includes pages that technically fit because margins or fonts were squeezed — altered formatting counts as over-length.
- The rebuttal exists to correct factual errors and supply requested information. It is explicitly not for new contributions or unrequested new experiments, reviewers are told not to demand significant new experiments, and it must stay anonymous with no links to external material (no anonymous repos, no video URLs, no project pages).
That last clause surprises people arriving from venues with revision uploads: at CVPR you cannot attach anything. If a number matters, it goes in the page as a sentence or a five-row table.
Triage before typing
With ~16,000 papers in the 2026 pool, your reviewers wrote many reviews and your AC is tracking dozens of papers. The rebuttal's real reader is the AC during the discussion window. Triage accordingly:
| Review signal | Rebuttal move | Budget |
|---|---|---|
| Factual error ("no ablation on X" when Table 4 has it) | Quote the reviewer, cite the exact table/line, one sentence | 2–3 lines each |
| Misread contribution / scope | Restate the claim in one crisp sentence, point to where the paper says it | 4–6 lines |
| Requested clarification (protocol, metric, split) | Answer directly; give the number | 3–5 lines |
| Requested small experiment | Run it if it fits the week; report as a mini-table | 6–10 lines |
| Demand for a full new benchmark or method variant | Decline politely; explain why out of scope for a rebuttal | 2–3 lines |
| Subjective "not novel enough" | Delta-list against the 2–3 closest citations, no adjectives | 4–6 lines |
Answer every reviewer, but not every sentence of every reviewer. Ignoring one review entirely reads as concession in the discussion phase.
A page layout that ACs can skim
% rebuttal.tex skeleton — headings are navigation for the AC
\section*{Shared concerns (R1, R3): fairness of the baseline comparison}
% one paragraph + mini-table with matched-backbone numbers
\section*{R1: reported speed}
% cite Sec. 4.2 line refs; give ms/frame on the named GPU from the CRF hardware
\section*{R2: missing citation [Foo 2025]}
% concede, state the delta in one sentence, promise camera-ready citation
\section*{R3: qualitative failures}
% point to supplement Sec. C; summarize the two failure modes in text
Group shared objections once instead of three times; name reviewers in headings; lead with the concern you can kill with a fact, because the first inches of the page set the AC's prior about who is being careful.
Tone calibration
- Quote reviewers verbatim before answering; paraphrases invite "that's not what I said."
- Concede real errors fast and specifically — a clean concession buys credibility for the contested points.
- No sarcasm, no appeals to arXiv fame, no "as any expert knows." The discussion phase happens without you in the room; give allies quotable sentences.
- Numbers over adjectives. "mIoU 54.1 → 55.8 with matched training schedule" survives discussion; "significantly better" does not.
A paragraph, before and after
Draft rebuttal prose tends to argue; shipping rebuttal prose demonstrates. Compare:
Before: "We respectfully disagree with R2's claim that our comparison is unfair. As explained in the paper, our method uses standard settings, and it is common practice in the field to compare this way. We believe the improvement is clearly significant."
After: "R2: 'the baseline uses a weaker backbone.' Both rows of Table 2 use ResNet-50 with the schedule from [24] (Sec. 4.1, L412–418). Under R2's suggested ViT-B setting, the gap is +1.9 AP (mini-table below), consistent with the paper's +2.1."
The rewrite quotes the objection, cites the paper by line, gives a number, and never characterizes its own results. Every paragraph on the page should survive the same transformation: objection verbatim → paper location → number.
First-hours protocol when reviews land
- Read all reviews once without responding, then close the laptop for a few hours; the first-read emotional register writes bad rebuttals.
- Build a flat list of every distinct objection with reviewer IDs and a fixability tag: factual / clarification / small-run / out-of-scope.
- Check scores against the venue's realities (25% acceptance) before deciding how much of the week the paper deserves versus the resubmission draft.
- Assign the roles from
cvpr-workflowand set the internal freeze 24 hours before the OpenReview deadline.
What you cannot fix in January
The rebuttal cannot rescue a paper whose evidence is missing, and reviewers are instructed not to expect new experiments. If every review converges on the same absent comparison, the plan is the next cycle (or ICCV/ECCV), not one page of protest. Spend the week where movement is possible: factual corrections, misreadings, and small requested numbers.
Reverify each cycle
- Rebuttal existence, length cap, and template — chairs have restructured response formats at CVF venues before.
- Exact window dates and whether reviewer–author discussion (beyond one PDF) exists.
- Whether revision uploads or supplement updates are permitted (they were not in 2026).
Output format
[Rebuttal plan] file / skip (reason)
[Shared concerns] <grouped items with reviewer IDs>
[Factual corrections] <claim → paper location>
[Numbers to include] <experiment → mini-table budget>
[Declined requests] <item → one-line rationale>
[Page budget] <lines allocated per section, total ≤ 1 page>