Skip to content

rStar-Math

Self-evolved small models for math reasoning

Visit official site

What rStar-Math is

The study is described as showing that small language models can reach strong math reasoning through MCTS-driven deep thinking and self-evolution.

What you can do with it

  • Run test-time search with Monte Carlo tree search
  • Synthesize verified reasoning traces from code-augmented CoT data
  • Evolve the policy SLM and PPM over four rounds

Pros & cons

✓ What we like

  • A process reward model avoids step-level score labels
  • Four rounds of self-evolution use millions of synthetic solutions

! What to watch out for

  • The approach relies on test-time search
  • Both a policy SLM and a PPM must be built

FAQ

What search method is used?

Monte Carlo tree search with a process reward model.

What AIME result is listed?

An average of 53.3%, or 8 of 15 problems.

Last reviewed: 2026-11-15

More LLM API platform tools

View all →

How we review