Manifold AIs Win Math Olympiad Gold: Prof. Lin Yang (UCLA) – #97
31 snips
Oct 23, 2025 Lin Yang, a computer science professor at UCLA with dual PhDs in physics and computer science, discusses his groundbreaking work with AI in solving International Math Olympiad problems. He explains the innovative verifier-refiner pipeline that allows models like ChatGPT to reach gold medal levels. The conversation dives into industry applications, the efficiency of token budgeting, and the future of AI in scientific research and legal analysis. Yang also shares insights into the limitations of current models and the potential path toward AGI.
AI Snips
Chapters
Transcript
Episode notes
Gold Medal Using Off-The-Shelf Models
- Lin Yang and collaborator used off-the-shelf models (Gemini, Grok, GPT-5) with their pipeline to solve five of six IMO problems.
- They achieved gold-medal level performance without training new models.
Verifier-Refiner Loop Amplifies Reasoning
- Lin Yang built a verifier-refiner loop to overcome LLM context and reasoning limits for hard math proofs.
- The pipeline iterates solution → verifier finds gaps → refiner fixes until repeated verifications pass.
Verify Multiple Times Before Accepting
- Re-verify candidate solutions multiple times to reduce random errors from temperature and sampling.
- Accept a solution only after repeated independent verifications (they used five passes) to increase reliability.