MaxProof: Scaling Mathematical Proof with Generative-Verifier RL and Evolutionary Search - MiniMax Research
MiniMax's MaxProof framework lets M3 exceed the human gold-medal threshold on IMO 2025 and USAMO 2026.
- MaxProof samples 32 candidate solutions, verifies each 4 times, runs up to 10 refinement rounds, and selects 4 parents per round.
- The training pipeline produces three expert models—Proof Expert, Verifier Expert, and Fixed Expert—that are merged into M3.
- Proof RL uses external frontier models as generative verifiers with layered defenses and pessimistic min aggregation across multiple judges.
- Verifier Expert training uses reward R = 0.7 * R_error + 0.3 * R_verdict for joint error finding and classification.
- The M2 reward-hacking case study showed false-positive rate rising from 2.9% to 8.0% and visible solution length tripling from 3.5K to 10K characters.
Math-reasoning teams now have a published MaxProof design for using generative verifiers in RL and evolutionary test-time scaling to convert best@K into stable pass@1.
Sources
Read this as text
Back to the AI news