OpenAI has withdrawn three mathematical manuscripts after finding a sign error in an argument that two other papers depended on. The October 7 entry in its repository history also records revisions to 14 other manuscripts. The update follows the 722-manuscript release we covered, produced with an unreleased internal model.
The withdrawals show how one failed argument can affect several published results. The size of the release alone gives readers too little evidence to judge the reliability of the batch.
One error affects three manuscripts
OpenAI says the sign error invalidates a stabilization-trace cancellation argument in “Algebraicity of Weil classes on split abelian eightfolds” and the construction used by two dependent papers. It withdrew:
- “Algebraicity of Weil classes on split abelian eightfolds”
- “Algebraicity of Kuga-Satake Correspondences for K3 Surfaces”
- “The rational Hodge conjecture for products of K3 surfaces”
The papers’ README files now explain the gap and link to archived manuscripts. OpenAI’s repository README lists 719 manuscripts in the current catalogue and says results are at different stages of verification.
Revisions cover more than the withdrawals
The 14 revisions include proof repairs, corrected statements, clearer hypotheses and dependencies, and a correction to an obsolete citation. Among them, OpenAI corrected a cone-equality claim in “Taming implies compatibility” and added an example of strict inclusion. Six manuscripts on Kähler minimal model programs and abundance received expanded arguments and clearer statements of the results they rely on.
OpenAI also updated 13 additional manuscripts to cite revised companion papers. The history describes those changes as reference and version-date updates. They belong in a different category from repairs to a paper’s own argument.
The same entry reports six additional formalizations and five additions covering supporting results. OpenAI puts the total share of top-line results formalized at 300 out of 719, about 42%.
