Definition
Plain language
A system that wraps a non-frontier model in a paranoid grading-and-search loop to reach gold-medal-level math olympiad performance.
As stated in the literature
MiniMax's framework combining a four-layer generative verifier, RL-trained proof/verifier/fixer experts, and a population-level evolutionary test-time search that proposes, critiques, patches, and selects proofs; reported to cross gold-medal thresholds on IMO 2025 and USAMO 2026.
Why it matters: It shows that wrapping a non-frontier model in a rigorous grade-and-search loop can reach top competition-level proof performance without a stronger base model.
For example, it proposes a math proof, has several judges critique it, patches the flaws they find, and only then picks the survivor that holds up.
Heard on the show
“… The paper is called "MaxProof: Scaling Mathematical Proof with Generative-Verifier RL and Population-Level Test-Time Scaling," …”Episode 133 — How MiniMax Turned a Reward-Hacking Disaster Into Olympiad Gold