AI Achieves Gold-Medal Level on IMO 2025 Through Population-Level Proof Scaling

MaxProof demonstrates that AI can now match human mathematicians on competition-level proof problems by using a novel test-time scaling approach that treats a single model as a generator, verifier, and refiner across a population of candidate proofs. This breakthrough reveals that scaling inference-time resources—rather than training parameters—may be a critical path for advancing AI reasoning on complex formal problems.