Reproduce “Grounded Chess Reasoning in Language Models via Master Distillation” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Zhenwei Tang, Qianfeng Wen, Seth Grief-Albert, Yahya Elgabra, Blair Yang, Honghua Dong, Ashton Anderson

The "Master Distillation" framework enables compact language models to perform grounded chess reasoning by synthesizing explainable chain-of-thought data from expert systems and frontier LLMs. This approach allows a 4B parameter model to achieve 48.1% accuracy in chess puzzle solving, surpassing its larger teacher model and significantly reducing token usage compared to baselines.

Key insights

  • Master Distillation effectively bridges symbolic expert knowledge with neural language generation, providing a method to inject domain expertise into LLMs for explainable reasoning.
  • Student models (C1-4B) can exceed the performance of their teacher LLMs (Gemini-3-Flash) through this distillation process combined with RLVR, suggesting that the synthesized data facilitates deeper learning.
  • Full fine-tuning, rather than parameter-efficient methods like LoRA, and theme-balanced data sampling are crucial for acquiring novel domain knowledge and ensuring robust performance across diverse tactical concepts.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.