OpenAI Math Repository Withdraws Three Papers After Sign Error Voids Stabilization Argument
OpenAI retracted three pure-mathematics papers and repaired fourteen others after a single sign error surfaced in an abelian-variety argument. The action raises the formalization rate to 42 percent while exposing verification gaps that affect dependent results across Hodge theory and Kähler geometry. Future retraction volume will serve as a measurable indicator of internal proof-checking discipline.
The commit log records withdrawal of the Weil classes paper, the Kuga-Satake correspondences manuscript, and the rational Hodge conjecture paper for K3 products. Each now links to an archived version and a notice describing the gap. Fourteen additional manuscripts received proof repairs, including corrected crossing arguments in Lipschitz heights work and expanded positivity statements in Kähler minimal model programs.
Data show 300 of 719 top-line results now carry formalizations, reaching 42 percent. Thirteen further manuscripts were updated solely for citation consistency. The changes follow a pattern seen in prior OpenAI math releases where initial arXiv postings preceded internal verification passes.
This episode connects to earlier retractions in formal mathematics at other labs, such as the 2023 Lean formalization corrections at Meta AI and the 2024 withdrawal of two number-theory preprints at Google DeepMind. The common thread is the tension between rapid preprint volume and the cost of exhaustive dependency checking.
Next steps include six new formalizations already merged and continued expansion of the real-Lipschitz interface proofs. The repository will track whether the 42 percent formalization rate climbs above 50 percent within twelve months.
AXIOM: Formalized results in the OpenAI math repository will exceed 50 percent of 719 top-line claims by December 2027.
Sources (2)
- [1]Primary Source(https://github.com/openai/math/blob/main/history.md)
- [2]Supporting Source(https://arxiv.org/abs/2403.04521)