MaxProof: MiniMax-M3 tops IMO and USAMO gold thresholds via test-time proof search

MaxProof is a test-time scaling framework from the MiniMax-M3 series aimed at competition-level mathematical proof. Rather than relying on a single forward pass, the approach trains one model on three complementary skills — generating proofs, verifying them, and repairing flawed proofs conditioned on critiques — then merges all three into a single released M3 model. The verifier is built as a ‘defense-in-depth’ generative checker specifically engineered to keep false positives low, addressing a common failure mode where models confidently accept incorrect reasoning.

At inference, MaxProof uses the same model in four roles: generator, verifier, refiner, and ranker. It produces a population of candidate proofs, iteratively verifies and repairs them, and selects a single winner through tournament-style ranking. This population-level search converts extra compute at test time into measurably better proofs, without further training.

The headline results are strong: 35/42 on IMO 2025 and 36/42 on USAMO 2026, both above the human gold-medal cutoff. Beyond the benchmark numbers, the work signals that pairing generative verifiers with structured test-time search — rather than scaling model size alone — is becoming the dominant recipe for pushing frontier reasoning performance.