Skip to content

rStar

Boost small language model reasoning with self-play

Visit official site
Category AI models

What rStar is

The approach is a self-play mutual reasoning method that raises the reasoning ability of small language models. It separates solution generation from mutual verification.

What you can do with it

  • Generate reasoning trajectories with MCTS
  • Verify trajectories with a discriminator SLM
  • Improve QA and inference task results

What to watch out for

  • The environment lists Python 3.10 and CUDA 12
  • PyTorch, transformers and vllm are named
  • Generation uses the run_gsm8k_generator.sh script

Pros & cons

✓ What we like

  • No fine-tuning or stronger model is required
  • Experiments cover multiple SLMs

! What to watch out for

  • The setup names Python 3.10, CUDA 12, PyTorch, transformers and vllm
  • Generation runs through the run_gsm8k_generator.sh script

FAQ

Does it need fine-tuning?

The material states no fine-tuning or more advanced model is needed.

How are trajectories checked?

A second SLM of similar capability verifies them as a discriminator.

Last reviewed: 2026-11-15

More LLM API platform tools

View all →

How we review