MaxProof: Scaling Mathematical Proof with Generative-Verifier RL and Population-Level Test-Time Scaling

The buried detail is cost structure: tournament selection over candidate populations at inference time is computationally expensive in ways that parameter scaling is not, and the paper does not appear to address what this means for deployment economics outside of competition benchmarks.

Recent Modelwire coverage has been tracking domain-specific reasoning evaluation, most directly the SupraBench work from June 11 on chemistry benchmarks. That paper and MaxProof are converging on the same underlying question from opposite directions: SupraBench asks whether LLMs can reason reliably in constrained scientific domains, while MaxProof demonstrates a compute-intensive method for pushing reasoning quality higher on formal tasks. Together they sketch a pattern where evaluation rigor and inference-time orchestration are developing in parallel, each raising the bar the other must clear. The broader context is that MaxProof belongs to a cluster of inference-scaling approaches (alongside chain-of-thought sampling and self-consistency methods) that labs have been quietly investing in as parameter scaling returns compress.

Watch whether MiniMax or a competing lab publishes per-proof inference cost figures alongside accuracy results in the next six months. If nobody does, that omission will tell you something important about whether this approach is viable outside of benchmark conditions.