Skip to content

OpenAI

One sign error pulls three OpenAI math papers

2 min read

Written by an AI model from the linked sources and checked against them by a second model. How we work

All about: OpenAI models and products →
Image: github.com

OpenAI on October 7 withdrew three papers from its math repository after a sign error broke a construction in one of them that the other two relied on, so the repository’s claim on one case of the Hodge conjecture, a Millennium Prize Problem, is off the table. The same entry in its history log on GitHub puts Lean formalizations at 300 of 719 top-line results, about 42%, which leaves more than half of the collection’s headline claims without a machine-checked proof.

The error sits in “Algebraicity of Weil classes on split abelian eightfolds”, where it invalidates a stabilization-trace cancellation argument and the construction that “Algebraicity of Kuga–Satake Correspondences for K3 Surfaces” and “The rational Hodge conjecture for products of K3 surfaces” build on. Each of the three now carries a notice that explains the gap and links to the archived manuscript. A separate paper in the repository, “The rational Hodge conjecture for CM abelian varieties”, dated September 30, is not on the withdrawal list; it, too, stops short of the full conjecture.

Fourteen other manuscripts got proof repairs, corrected statements, clearer hypotheses and dependencies, and in one case a fix to an obsolete citation. Most of them fall into two groups: six papers on Kähler minimal model programs and abundance, and four on Lipschitz heights and Ashkin–Teller currents. In “Taming implies compatibility”, the cone-equality claim was corrected and a strict-inclusion example added. Thirteen more manuscripts were updated to cite the revised editions of companion papers, which changes their references and version dates. OpenAI’s own announcement tallies the update as 6 new Lean formalizations, 19 modifications and 3 withdrawals.

The repository launched with protocols for paper revisions and citations. A withdrawn paper’s README explains the gap and links to the retracted manuscript, while an updated paper keeps its earlier editions reachable through the version notes in its README, so anyone who cited an earlier edition can still find it.

At launch, Decrypt counted 162 of 722 papers with a computer-checked main result, going by the repository’s formalization catalog, and the README warned that “some of the unformalized results could have issues”. A Lean check is also a narrower guarantee than the word “formalized” suggests: it confirms that a proof follows from the statement as written in Lean, not that the statement matches the open problem or that the result is new, so a mathematician still has to confirm that each formal statement says what the problem actually asks.

Announcing the update, OpenAI said it will “continue to update the repo with new formalizations and with any errata we notice.”

No comments yet. Be the first.

Join the conversation

Sign in to leave a comment. It appears after moderation.