brycewang-stanford ICCV Review Process
This models what happened to a paper inside the ICCV 2025 pipeline — the most recent completed cycle, verified 2026-07-08 — and what an author can and cannot influence at each stage. The 2027 chairs may restructure any of it (待核实 once the Hong Kong cycle posts its pages).
Pipeline at a glance (2025 anchors)
Mar 3 registration closes (abstract, authors, profiles)
Mar 7 paper + supplement due (one event, 11:59pm HST)
Mar–May assignment and reviewing; authors serve as reviewers in parallel
May 9 reviews released to authors
May 10–16 one-page rebuttal window
May–Jun reviewer/AC discussion and meta-review (authors silent)
late Jun decisions (2025 statistics: 11,239 valid submissions,
2,699 accepted, 24%)
Oct 19–23 conference, Honolulu (workshops 19–20, main program 21–23)
Between March and May the same lab members are usually on both sides of the table, because ICCV 2025 required every author to review unless exempted (exemptions covered people new to or outside the vision community). Model your own reviewers accordingly: they are authors from the same submission wave, mostly competent, uniformly overloaded, and reading your paper between their own rebuttal-week anxieties.
Enforcement, with 2025 receipts
The distinctive fact about ICCV review is that reviewer accountability is enforced against authored papers, and 2025 published the outcome rather than the threat:
- 25 reviewers were classified as highly irresponsible;
- 29 papers they authored were desk-rejected as a consequence;
- 12 of those papers had been slated for acceptance.
"Highly irresponsible" was defined with examples: one-sentence reviews, reviews generated by an LLM, reviews irrelevant to the paper, reviews ignoring substantial parts of it. Missing the review deadline outright carried the same desk-reject consequence. Author-side takeaways: pick coauthors' review weeks into the project calendar, have first-time reviewers' drafts read by someone senior, and treat a coauthor who "will get to the reviews later" as a risk to every paper you share.
The LLM line
Reviewers were prohibited from using LLMs to write reviews — locally or via API — and LLM-written reviews were named in the irresponsible-behavior list. For you as an author this cuts two ways:
| Side | Rule | Practical move |
|---|---|---|
| Your duty as reviewer-author | No LLM-drafted reviews; no feeding substantial paper content to models | Write reviews yourself; a violation risks your own submissions |
| Your paper receiving reviews | An LLM-flavored review (generic praise, hallucinated citations, wrong paper details) is a policy violation by the reviewer | Document the evidence and flag it through the official channel (confidential comment / chairs contact), never in the public rebuttal's tone |
What the decision distribution looked like
Acceptance at 24% came in tiers. Community trackers reported orals at roughly half a percent of submissions and 263 Highlight papers among the acceptances — treat tier counts as reported rather than officially re-verified. The tiering means the discussion phase is not deciding accept/reject alone; it is also ranking. A rebuttal that converts a reviewer from "weak accept" to advocate does double work.
Reading your reviews as data:
- Convergent factual objection (all reviewers name the same missing baseline): the discussion phase will not save the paper; the fix is the next cycle's experiment plan.
- Divergent scores with one detailed negative: the AC weighs argument quality; a rebuttal that surgically refutes the detailed negative has real leverage.
- Uniform shrugs ("sound, limited novelty"): a framing failure — the mechanism
was not made to look inevitable. This is writing-side work
(
iccv-writing-style), not evidence-side. - A review that misdescribes the paper: quote it against the PDF in the rebuttal's first lines; misread-driven weaknesses collapse when named calmly.
After the decision: the 2025-specific outlet
2025 introduced a Findings workshop for ICCV: papers rejected from or withdrawn during the main-conference review, judged technically sound with evidence-backed claims, could be submitted (June 30 deadline in 2025) and presented at a workshop in Honolulu. It explicitly did not take work still under review elsewhere or papers that had never been peer reviewed. If it recurs in 2027 (待核实), it changes rejection calculus: sound-but-unexciting papers gain an ICCV-week audience without waiting for ECCV. Weigh it against holding the work for a full resubmission — Findings publication may interact with later venues' dual-submission definitions, so read both CFPs first.
Where authors have zero leverage
Silence discipline matters at a venue with explicit misconduct rules: no emailing ACs or chairs to advocate, no hunting reviewer identities, no social-media campaigns about your scores while under embargo (media discussion of a paper under review is itself a violation that can remove the paper). The rebuttal PDF and the official confidential channels are the entire authorized surface.
Reverify each cycle
- Whether every-author enrollment, the exemption categories, and desk-reject enforcement carry into 2027.
- The rebuttal window's existence and length.
- Tier definitions (oral/Highlight ratios move with program capacity).
- Whether Findings runs again, and its relationship to dual-submission rules.
Output format
[Stage] submitted / in-review / rebuttal / discussion / decided
[Duty compliance] all reviewing coauthors on-time and substantive: yes/no/unknown
[Review diagnosis] per review: convergent-factual / divergent / shrug / misread
[Leverage plan] rebuttal targets ranked by winnability
[Policy flags] LLM-flavored review or misconduct evidence: <none or summary>
[If rejected] Findings-eligible? next-venue plan with dates