What most flagships allow
Venues like NeurIPS, ICML, ICLR, and CVPR typically allow arXiv (and similar preprints) and do not treat a reviewer stumbling on one as an automatic conflict.
The hard rule remains: the review PDF, supplements, and anonymous code must not identify you; cite your own preprint in the third person.
Where people cross the line
Writing “our arXiv version” in the review PDF, linking a named GitHub, or posting “under review at Double-Blind Venue X” with the preprint—those are riskier than merely hosting on arXiv.
A few venues are stricter; the live CFP wins. Do not assume English conference norms transfer to a Chinese double-blind journal.
- Keep the review PDF anonymous; do not link the non-anonymous full text
- Self-cite preprints in the third person without saying you are the submitter
- During review, avoid naming the venue you are under review at
Whether and when to post
Priority, collaborator sync, and a venue that allows it while you can stay blind are common reasons to post. If your institution or coauthors need secrecy—or the target forbids it—wait.
Many groups post around the deadline; what matters is isolating identity between the review packet and the public version, not the exact day.
FAQ
- A reviewer found my arXiv—am I doomed?
- At most ML/vision flagships that alone is not an automatic conflict; your review PDF must still be anonymous. Follow that year’s policy.
- Update arXiv before acceptance?
- Technical updates are fine; avoid “under review at …” on the public page during review. Syncing after camera-ready is cleaner.