cvpr-review-process
GitHub解析CVPR大规模评审流程，涵盖OpenReview时间线、AC讨论机制及25%录取率预期。重点说明作者应对策略：针对新手审稿人写作、利用清晰反驳争取关注，并强调严禁LLM撰写评审的合规红线及违规举报渠道。
Trigger Scenarios
Install
npx skills add brycewang-stanford/Awesome-Journal-Skills --skill cvpr-review-process -g -y
SKILL.md
Frontmatter
{
"name": "cvpr-review-process",
"description": "Use when reasoning about how CVPR review actually works at 16,000-submission scale, covering the OpenReview pipeline and timeline, reviewer-duty enforcement and desk rejects, the reviewer LLM ban, AC and discussion dynamics, oral\/highlight\/poster decision tiers, and calibrating expectations to a ~25% acceptance rate."
}
CVPR Review Process
Use this to model what happens to a CVPR paper between upload and decision, and where author leverage exists. Every mechanism below is the 2026 cycle as verified 2026-07-08; scale numbers come from official CVPR program announcements.
The machine, by the numbers
The 2026 cycle put 16,092 submissions through full review (withdrawals and desk rejects excluded from that count) and accepted 4,090 — 25.42%. Doing that required roughly doubling the reviewer pool to 25,149 reviewers across 97 countries, coordinated by 909 area chairs. Two consequences follow for authors:
- Your reviewers are statistically likely to include first-time CVPR reviewers (the pool doubled in one year; official training material exists for a reason). Write for a competent generalist in vision, not the five people in your subfield.
- No individual paper gets much AC deliberation time by default. Papers that make the AC's job easy — clear claims, clean rebuttal, consistent reviews — get the benefit of scarce attention.
Timeline anatomy (2026 cycle)
| Phase | 2026 dates | What authors can do |
|---|---|---|
| Profile/abstract registration | Nov 6, 2025 | Fix profiles, finalize author list |
| Paper + CRF deadline | Nov 13, 2025 | Submit; after this, nothing |
| Supplement deadline | Nov 20, 2025 | Upload depth material |
| Reviewing | Dec – Jan 12, 2026 | Serve your own reviewer duties well |
| Reviews released | Jan 22, 2026 | Read cold, then triage |
| Rebuttal | → Jan 29, 2026 | One page (see cvpr-author-response) |
| Reviewer–AC discussion | Jan 30 – Feb 5, 2026 | Nothing — your rebuttal speaks |
| Decisions | Feb 20, 2026 | Plan camera-ready or resubmission |
Enforcement culture: the distinctive part
CVPR's response to scale is to make authorship and reviewing one social contract, with teeth that surprised people in recent cycles:
- Enrollment: submitting authors are expected to review and are added to the pool unless exempted; complete OpenReview profiles are mandatory for everyone.
- The nuclear clause: papers authored by a reviewer who fails to deliver assigned reviews on time, or delivers highly irresponsible ones, can be desk-rejected — all of that person's submissions — at PC discretion. A coauthor's negligence in January can kill your November submission.
- Reviewer LLM ban: reviews and meta-reviews may not be written by LLMs (local or API), and sharing substantial paper content with an LLM is prohibited; grammar checkers and background research are the carve-outs. As an author, this means a suspiciously LLM-flavored review is reportable to the AC, not just annoying.
# Confidential-channel escalation, in order
1. Rebuttal PDF → factual corrections all reviewers see
2. Confidential comment to AC (where the form allows)
→ reviewer misconduct: LLM-written review, review of the wrong paper,
demonstrably unread paper. Cite evidence, stay unemotional.
3. Never → contacting reviewers or ACs outside OpenReview; that
is an integrity violation on YOUR side.
What a CVPR review contains
The review form's exact fields shift by cycle (待核实 each year), but the durable skeleton a CVPR review argues through is: summary of the paper in the reviewer's own words (misreadings here predict everything downstream), claimed strengths, weaknesses with the score-driving one usually listed first, and a recommendation with confidence. Read your reviews against that skeleton: a wrong summary is rebuttal priority one, because every weakness derived from it inherits the error.
Decision tiers
Acceptance is not binary. The 2026 program sorted acceptances into poster, highlight (program-flagged posters, 578 papers per program trackers), and oral (141 papers in four parallel tracks), plus 74 award candidates — treat those tier counts as reported, and the tiering itself as the durable pattern. Tier decisions ride on review scores plus AC/SAC advocacy, which is one more reason the rebuttal's real audience is the AC.
After a rejection: mining reviews for signal
Roughly three of four reviewed papers exit here, so treat rejection processing as part of the process, not an aftermath:
- Separate objections into evidence gaps (missing baseline, unmatched backbone —
fix before any resubmission; CVF reviewer pools overlap and repeat objections
compound), communication failures (a reviewer missed something the paper does
say — a writing-style problem, see
cvpr-writing-style), and taste (novelty judgments that a different draw of three reviewers may not share). - Score patterns carry information: uniform lukewarm reviews usually mean a framing problem; one champion plus one detractor means the discussion phase decided it, and the rebuttal is where to look for what failed.
- Do not shop an unchanged paper to the next deadline. The overlap between CVPR, ICCV, and ECCV reviewer pools makes "same paper, new lottery ticket" a strategy with memory.
Calibration for authors
- A ~25% rate at 16k submissions means ~12,000 rejections of mostly competent papers. Rejection is the modal outcome for good work; plan the ICCV/ECCV/next-CVPR path at submission time, not in grief.
- Borderline papers live or die in discussion week. What moves them: a rebuttal that kills a factual objection, and one reviewer willing to champion. What doesn't: eloquence about importance.
- Reviews disagreeing 2-vs-1 is normal at this scale; the AC weighs arguments, not averages, more than authors assume.
Reverify each cycle
- Whether decisions remain single-round (no conditional-accept revision cycle existed in 2026).
- Reviewer-duty enforcement wording, LLM-policy scope, and review-form structure.
- Tier definitions (highlight/oral percentages shift with program capacity).
Output format
[Stage] pre-submission / in-review / rebuttal / discussion / decided
[Process risks] duty-compliance · profile validity · policy exposure
[Review read] per-reviewer: score-driver + fixability in one line
[AC theory] what the meta-reviewer needs in order to advocate
[Next actions] <dated, owner-assigned>
Version History
- 9f86f09 Current 2026-07-19 14:42
