OpenAI releases 722 AI math papers amid criticism from mathematicians

OpenAI has published 722 mathematical manuscripts from an unreleased AI model, reporting solutions to hundreds of open problems amid criticism of how the company shares its research.

The collection on GitHub, released on October 6, contains 372 families of related results. Each family can include a main proof, supporting arguments, consequences or alternative proofs, so the manuscript count does not represent 722 separate problems solved.

OpenAI says it gave the internal frontier model approximately 4,000 research problems after performance on its existing math evaluations reached a ceiling. The company expanded testing to open questions, where producing a correct answer would contribute new mathematical knowledge.

The vast majority of results came from the same evaluation procedure. Each used computing resources equivalent to roughly three hours of ChatGPT Pro thinking on average, according to OpenAI. This describes computational effort, not the time needed to independently verify a proof.

Alongside the manuscripts, OpenAI released shortened reasoning summaries for 10 results and formalizations of many proofs in Lean, a language that allows computers to check mathematical arguments. The collection is at different stages of verification, and the company acknowledges that some results without formalizations could contain errors.

Ahead of publication, academics told WIRED that competition between OpenAI and Anthropic was turning mathematics into a showcase for their models, sidelining established practices for publishing research and crediting earlier contributions. Northwestern University mathematician Bryna Kra said researchers wanted papers that explained the work so they could understand and build on it.

Access to the technology is another point of contention. The independent Advisory Group on Mathematics and Artificial Intelligence, which has advised OpenAI, previously called on labs to stop testing advanced mathematical problems on proprietary models unavailable to the wider scientific community. It warned of a two-tier system in which AI companies generate discoveries using tools other researchers cannot access.

In its response to the release, the group called the publication an important event but said its involvement was not an endorsement of OpenAI’s methods or an assessment of the results. It argued that mathematicians must be able to pursue their own research, beyond interpreting work produced by AI labs.

OpenAI says the group’s advice informed the release and has pledged funding for workshops, conferences and programs to help researchers understand major AI-generated results. It also committed to improving the papers’ citations and explanations in future releases.