Pickmath: 722 Math Papers Proven by OpenAI Internal Models Released

This is a collection of 722 math papers and Lean formalization evidence written by OpenAI's internal models solving open research problems. As existing evaluation benchmarks reached saturation, the collection contains results from new, more difficult problems that the models solved autonomously.

Generated after an average of 3 hours of inference computation, significant results from over 4,000 problems are categorized into 372 families. It includes approaches to classical open problems such as the zero-free region of the Riemann zeta function and the Hodge conjecture, with some formally verified in Lean.

Reasoning Traces summarizing the models' thought processes are also provided, allowing observation of how AI develops complex mathematical logic. Released under the Apache-2.0 license, it can be used by researchers for verification or follow-up studies.