OpenAI publishes 722 math manuscripts from an internal model
The public collection gives mathematicians papers and some proof artifacts to inspect, but OpenAI warns that unformalized results may contain errors.
Papers available for examination
OpenAI released a public repository of 722 mathematical manuscripts organized into 372 families of related results on October 6. Its README says an unreleased internal model produced the manuscripts and that the collection includes PDFs, source files, citation information and some Lean proof artifacts. OpenAI says it posed the model approximately 4,000 problems during the evaluation and has also published abridged reasoning summaries for selected results. These figures come from the company’s repository; they count papers, grouped results and attempted problems, not independently verified solutions.
Verification is still under way
The repository explicitly says the materials are at different stages of verification. Not every manuscript has a Lean formalization, and OpenAI warns that some unformalized arguments could contain errors. It says corrections will be released as new versions while earlier versions remain available. The Verge reported that OpenAI’s independent mathematics advisory group had urged AI labs to share results through established academic channels and disclose how they were produced. That group was announced earlier; the new event is the release of material that mathematicians can now inspect. The collection may contain important results, but its size alone does not establish that hundreds of open questions have been conclusively resolved or independently reviewed.