JOURNAL / 2026.10.12
OpenAI publishes hundreds of mathematics manuscripts: withdrawals count too
A collection generated by an internal model opens results and formal proofs to review. Three withdrawals sharing a dependency show why its value requires checking both the arguments and the scope of what has been proved.
On October 6, OpenAI published a collection of mathematical results produced by a frontier model that remains internal. The release makes manuscripts, source files and Lean formalizations available for inspection; Lean is a language for checking mathematical proofs by computer. This is a substantive change: other people can work on the arguments. The model that produced them remains unavailable to the public; downloading the proofs does not provide access to the research instrument.
OpenAI's current catalogue contains 719 manuscripts grouped into 372 families. A family may bring together a principal result, supporting arguments, consequences and alternative proofs. These units do not represent an equal number of solved open problems. The company also acknowledges different stages of verification and possible errors in unformalized results.
The scale deserves attention, but a specific correction does more to explain what has been published. The repository's history, dated October 7, records the withdrawal of three manuscripts. A sign error invalidates a construction in Algebraicity of Weil classes on split abelian eightfolds; two papers on Kuga–Satake correspondences and products of K3 surfaces depended on it. The history also records revisions to fourteen other manuscripts, including repaired arguments, clarified hypotheses and a bibliographic correction, and reference updates in thirteen more. These changes are not all equally serious.
The withdrawal notice for the first paper explains the failure without requiring a blanket judgment about AI: a step was supposed to bring a signed count to zero, but the proposed operation moves it further from zero. A hypothesis of the theorem used subsequently is therefore unmet. The notice clarifies that withdrawing the proof does not establish that its conclusion is false. What fails is the argument intended to establish it.
This case reveals a relationship that paper counts conceal. Three texts can fail because of a single shared dependency. Repairing the first text would not automatically repair the others either: how they use the construction would need to be checked again. To make use of this collection, I think following its dependencies and versions is more useful than treating every title as an independent discovery.
What a formal proof covers
The repository's checking instructions explicitly distinguish several checks of supporting results from proofs of the main theorems. For example, one surface-cone configuration checks properties of an associated ring, rather than the paper's full result. Finding Lean code alongside a manuscript requires reading which statement it verifies and under which assumptions.
A formal proof can provide a very strong guarantee about the encoded statement, within its checker's rules and dependencies. Another question remains: does that statement express exactly what the text claims? Reviewing that correspondence, the hypotheses and the imported results is part of the work. We have not compiled the library or certified any proof here; we have read the instructions and coverage notices.
There is more transparency about the search than in a selection of successes without context. According to OpenAI, the model was given approximately 4,000 problems, and most results followed a common procedure, with average compute equivalent to roughly three hours of ChatGPT Pro thinking per result. The company identifies exceptions. Dividing the published families by the problems posed would not yield a reliable solve rate: there is selection for significance, grouping and still uneven verification. Nor does that average establish the cost of repeating the work with a commercial model.
The recommendations of the independent Mathematics and AI advisory group, which OpenAI consulted, add a requirement I share: support human understanding of the results and let the community lead that process. They call for scholarly repositories outside the labs' control, attribution of prior work and traceable revisions. The group also opposes investigating advanced problems with proprietary models inaccessible to the community. Consulting it does not amount to securing its endorsement of the release as a whole.
In this journal, we already examined six papers developed with Muse Spark. Those involved identified researchers who directed the work and took responsibility for the texts. This collection presents a different scale of review and a different relationship with a model that remains closed. I welcome opening the artifacts and preserving the withdrawals: they make correction and learning possible. But scientific progress will take hold when others can verify, explain and reuse specific results. A proof with an understood scope and a citable version is worth more for building knowledge than a catalogue read as a victory tally.
Sources consulted on October 12, 2026. The figures describe the inspected catalogue, which may change with further revisions.