ICML 2026 Desk-Rejects 497 Papers After Watermark Sting Catches 398 LLM Reviewers
Summary: ICML 2026 opened in Seoul on July 6 with a record 23,918 paper submissions — more than double last year's total — and immediately announced it had caught 398 reviewers using LLMs to write reviews, desk-rejecting all 497 papers they were responsible for.
Key Facts
- 23,918 submissions this year, roughly 2× last year's count, reflecting the surge in AI research output
- Program chairs embedded hidden "honeypot" prompts inside submitted PDFs, designed to trigger specific phrase pairs if an LLM processed the document
- 398 reviewers who violated Policy A (no LLM use permitted) were flagged; their reviews were removed and all 497 papers they covered were rejected outright
- Detection has a known blind spot: reviewers who paraphrase AI output or use LLMs only for drafting would not be caught
Why It Matters
This marks the first large-scale enforcement action against LLM misuse in top-tier AI peer review, and it sets a concrete precedent. The same methods may spread to other venues, reshaping how conferences verify reviewer authenticity. The episode also highlights a paradox: the AI community's own publication infrastructure is now vulnerable to the very technology it studies.
Read More
- ICML 2026 opens in Seoul — TechTimes
- ICML desk-rejects 497 papers — AI Weekly