OpenAI has released a broad set of new mathematical results produced by an internal frontier AI model, alongside details on how the results were generated and protocols aimed at improving transparency and reproducibility in AI-assisted mathematical research.
The results are being published through a GitHub repository, following consultations with the independent Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study. OpenAI said the group’s recommendations have helped shape its approach to releasing AI-generated mathematical research, including protocols for paper revisions and citations.
The repository also includes formalised versions of several proofs written in Lean, a programming language that enables mathematical proofs to be checked computationally. OpenAI said additional formalisations will be added as they become available.
The company is also providing information on the process used to obtain the results. This includes 10 summaries of the model’s reasoning, estimates of compute expenditure expressed in terms of ChatGPT Pro usage, and statistics on the number of problems attempted. According to OpenAI, the average result required compute equivalent to roughly three hours of ChatGPT Pro thinking.
The release reflects a broader effort to establish practices for communicating major scientific results generated with AI. OpenAI said it plans to improve future releases through stronger citations, clearer mathematical exposition and improved presentation of results.
The company is also planning a series of workshops, conferences and special programmes focused on understanding major mathematical results produced by AI. It said these initiatives are intended to encourage further engagement between AI researchers and the mathematics community.
OpenAI said it is continuing to evaluate its internal frontier models on mathematics and other scientific disciplines, while working towards responsibly releasing the model that produced the results. The company said feedback from the scientific community will inform future standards for disclosure of major AI-generated scientific advances.
