Training Data OpenAI’s IMO Team on Why Models Are Finally Solving Elite-Level Math
195 snips
Jul 30, 2025 
Noam Brown and Alex Wei, members of OpenAI's IMO team, share their groundbreaking journey to achieve gold-level performance on International Mathematical Olympiad problems. They discuss their innovative approach using reinforcement learning techniques and admit the model's self-awareness in tackling tough challenges. The duo emphasizes the contrast between solving competition problems and genuine mathematical breakthroughs. They also explore the balance between formal and informal verification, highlighting the interplay of human oversight in developing advanced mathematical models.
AI Snips
Chapters
Transcript
Episode notes
Model Self-awareness on Hardest IMO Problem
- Problem 6 at the IMO remains unsolved by models due to its complexity and narrow solution path.
- The model showed self-awareness by acknowledging it couldn't solve this problem instead of guessing.
Internal Odds on IMO Success
- Two months before the IMO, some researchers doubted gold medal success, offering 2-to-1 odds against.
- No one wanted to bet against the team, showing cautious optimism inside OpenAI.
Math Breakthroughs Demand More Reasoning
- Despite vast progress from grade school math to IMO gold, genuine research mathematic breakthroughs require far more time.
- Models need to scale reasoning from hours to thousands of hours to tackle truly unsolved math problems.