OpenAI Retracts Three Math Manuscripts After Sign Error Breaks Proof Chains
OpenAI has withdrawn three algebraic geometry manuscripts and patched 14 others after uncovering a sign error in a core trace cancellation proof. The incident reveals that only 42% of the project's 719 top-line results are formally verified in Lean, demonstrating why automated reasoning pipelines cannot rely purely on natural-language mathematical arguments.
Why it matters
Audit synthetic mathematical data and LLM reasoning pipelines with formal theorem provers like Lean 4 instead of trusting unverified natural language chains of thought.
TL;DR
- 01Natural language reasoning traces can conceal sign errors that silently invalidate downstream dependencies.
- 02Machine-checked proofs in interactive theorem provers like Lean 4 remain the gold standard against reasoning hallucinations.
- 03Only 42% (300 of 719) of top-line results in OpenAI's math repository have reached complete Lean formalization.
Key facts
- Withdrawn manuscripts
- 3
- Revised manuscripts
- 14
- Citation updates
- 13
- Top-line formalization rate
- 300 / 719 (~42%)
- New Lean formalizations
- 6
Cascading Retractions Across Dependent Manuscripts
OpenAI updated its public research repository to document the retraction of three mathematical manuscripts: *Algebraicity of Weil classes on split abelian eightfolds*, *Algebraicity of Kuga–Satake Correspondences for K3 Surfaces*, and *The rational Hodge conjecture for products of K3 surfaces*. The root cause was a sign error in the first paper that invalidated a critical stabilization-trace cancellation argument, breaking the mathematical foundation relied upon by the two dependent works.
Fourteen Papers Repaired Across Five Domains
Beyond the three retractions, OpenAI issued corrections across 14 companion manuscripts to repair proofs, tighten hypotheses, and fix dependencies:
- Lipschitz heights and Ashkin–Teller currents (4 manuscripts): Repaired crossing, boundary-attachment, conditioning, and convergence arguments, adding proofs for the real-Lipschitz interface.
- Kähler minimal model programs and abundance (6 manuscripts): Expanded positivity and contraction arguments while clarifying required input hypotheses.
- Taming and hypersymplectic deformation (2 manuscripts): Corrected a cone-equality claim, added a strict-inclusion example, and removed an obsolete dependency.
- Box Transport and Birch–Swinnerton-Dyer: Revised torus-projection estimates and purged an obsolete citation.
An additional 13 manuscripts were updated to reference the revised companion editions.
The Lean Formalization Deficit
The incident underscores the peril of unformalized proofs in automated research. OpenAI added 6 formalizations and 5 supporting additions, bringing total formalized top-line results to 300 out of 719 (~42%). Without machine-checkable Lean proofs, natural language mathematical arguments remain prone to latent human and synthetic errors.
Try it in 2 minutes
git clone https://github.com/openai/math.git && cd mathbash
✓ When to use
- When constructing automated theorem proving pipelines or synthetic data generation for reasoning models.
- When evaluating whether LLM reasoning chains require interactive theorem prover guarantees like Lean 4.
✕ When NOT to use
- When working on standard application engineering where conventional unit and integration tests suffice.
- When generating informal developer documentation where approximate summaries are acceptable.
What to do today
- Enforce formal verification via Lean 4 in automated reasoning pipelines instead of trusting natural language proofs.
- Audit downstream dependencies whenever synthetic mathematical training sets receive version updates.
What the community says
“The write-ups are one thing and more of a cherry on top but I would say the more pressing matter is the lack of Lean formalization which means you can't really say it has been (dis)proven or not.”
“flipping a sign in a draft paper that your buddy catches in a pre-read is SUPER common. People don't do math in a vacuum. Well, many do not. Some do.”