OpenAI's IMO Team on Why Models Are Finally Solving Elite-Level Math
OpenAI researchers Alex Wei, Sheryl Hsu, and Noam Brown join the Training Data podcast to discuss their model achieving gold-medal performance at the International Mathematical Olympiad (IMO). The three-person team built on general-purpose techniques for scaling test-time compute and tackling hard-to-verify tasks, rather than IMO-specific optimizations. They describe a rapid final sprint of only a couple of months, with strong early evidence coming from improvements on problems without verifiable rewards. The model reasoned for roughly 100 minutes per problem, up from about a tenth of a minute a year ago, and used natural-language proofs rather than Lean formal verification, which the team deemed too narrow for their general-reasoning goals. The team discusses why the model declined to attempt Problem 6 (a notoriously hard combinatorics problem) rather than hallucinating an answer, calling that self-awareness a meaningful sign of progress. Proofs were graded by three external former IMO medalists who reached unanimous consensus on correctness, and the raw proofs were published on GitHub despite being intentionally hard for humans to read. Noam Brown frames the work in terms of scaling reasoning from minutes toward the thousands of hours that would be needed for research-level mathematics and potentially Millennium Prize problems, while emphasizing that the underlying techniques—parallel compute via multi-agent systems, long-horizon RL on hard-to-verify tasks—are general-purpose and intended to feed back into broader OpenAI model capabilities. Looking ahead, the team plans to expose the system to mathematicians for evaluation rather than ship it directly as a product, and sees problem generation and long-horizon research reasoning as the next frontiers beyond time-boxed competition math.
Original sources
Podcast Training Data
Episode:OpenAI’s IMO Team on Why Models Are Finally Solving Elite-Level Math ↗
YouTube Training Data
Video:OpenAI’s IMO Team on Why Models Are Finally Solving Elite-Level Math ↗
AI Radar wrote this summary from the publisher’s notes with AI, not from a transcript. How it works.