OpenAI Withdraws Three Machine-Written Math Papers After a Sign Error

OpenAI has withdrawn three of the mathematics manuscripts it published in a public code repository, after a sign error in one of them invalidated an argument the other two relied on. The repository's own history file, dated Oct. 7, 2026, records those withdrawals and 14 further manuscripts revised to repair proofs, correct statements or spell out what each result assumes.

The collection holds papers and supporting proof files produced by an unreleased internal OpenAI model, which the company says it ran on open research problems as part of developing its models. The material sits in a public repository under an Apache-2.0 license, so anyone can download it, reuse it and check it. Withdrawn papers keep a notice explaining the gap and a link to the archived manuscript, and earlier editions of the revised papers stay reachable through their version notes.

By OpenAI's own count, the catalog lists 719 manuscripts grouped into 372 families, and 300 of the 719 top-line results now carry a proof written in Lean, a language a computer can check line by line. The remaining 419 do not. OpenAI states that the collection includes results at different stages of verification, that some of the unformalized results could have issues, and that it will try to fix any such issues quickly. No independent check of the catalog by mathematicians outside the company has been published.

The repository says the vast majority of the results came from one procedure using that unreleased model, which was posed about 4,000 problems over the course of the evaluation. What reached the catalog is what remained after the output was grouped into families and filtered for significance.

OpenAI says further corrections and revisions will be recorded as new versions, with previously released versions staying accessible.