Dispatch
Hugging Face put agents on 2,226 ICML papers and falsified claims in 23% of them
A community hackathon run July 15–August 2 and written up Aug 13 had 1,221 participants use Claude Code, Codex and Cursor to reproduce 2,226 of ICML 2026's 6,352 accepted papers, producing 6,816 logbooks and 2,962 cloud jobs. 51% of examined papers had at least one claim verified and 23% had claims falsified, with 49 papers falsified outright and 242 showing conflicting verdicts across independent teams. The organizers' line is the one worth keeping: 'Checking a paper carefully used to cost a reviewer a weekend; an agent can attempt it in an afternoon' — though agent-only runs hit limits on scale-dependent behavior and human-guided workflows were the most reliable.
↳ Follow the thread