MiniMax details MaxProof math-proof pipeline

MiniMax has detailed the methods behind MaxProof, the framework it credits for lifting its M3 model above the human gold-medal threshold on IMO 2025 and USAMO 2026. The company first reported those results in the M3 release post. The new blog post covers how they were achieved.

The approach has two parts. To raise best@K, MiniMax trained three expert models. A Proof Expert came out of long-horizon reinforcement learning on proof generation. A Verifier Expert was aligned with RL to find errors, trained on "golden" verification analyses gathered during the proof stage. A "Fixed Expert" was fine-tuned with rejection sampling to repair flawed proofs from those diagnoses. All three went into M3's final merged training. MaxProof, an evolutionary-search-style test-time scaling framework, then has the merged model refine its answers over multiple rounds to turn best@K into a more reliable pass@1.

The main departure from standard RLVR is that rewards come mostly from a generative verifier. MiniMax says it did not copy DeepSeek-Math-V2's self-verify and meta-verify setup, because its earlier models lacked enough verification ability to grade themselves. It used external frontier models as verifiers instead. To limit false positives and reward hacking, the verifier assigns zero to outputs that match known failure modes, normalizes proofs written in different styles, and takes the lowest score when several verifiers disagree. Scores are continuous in [0, 1].

The available excerpt ends partway through the verifier section. It does not include details of MaxProof's search procedure or per-problem scores.