ICML 2026 desk-rejects 497 papers for LLM review policy violations

The International Conference on Machine Learning 2026 program chairs desk-rejected 497 papers. The rejections corresponded to submissions by 398 reciprocal reviewers who violated rules on large language model use that they had explicitly agreed to follow. Reviewers chose between a conservative policy that banned any LLM use and a permissive policy that allowed limited assistance for understanding papers and polishing text. Reviewers assigned to the no-LLM policy produced 795 reviews that were later found to contain LLM-generated content. The detection method relied on watermarked submission PDFs that contained hidden instructions visible only to LLMs. The instructions directed any LLM to include one of two randomly selected phrases from a dictionary of 170,000 phrases in the review text. Every flagged review was manually inspected by a human to confirm the violation. All detected LLM-generated reviews were removed from the system. Fifty-one reviewers who produced LLM content in more than half their reviews were removed from the reviewer pool along with their remaining reviews. The chairs said they regretted the resulting disruptions and had been in direct contact with senior area chairs and area chairs to offer support. They reported that the probability of incorrectly flagging even a single review from a reviewer required to follow the no-LLM policy was 0.0001.