Reproduce “Mixing Expert Knowledge: Bring Human Thoughts Back To the Game of Go” on OpenResearch

About this paper

Yichuan Ma, Linyang Li, Yongkang Chen, Peiji Li, Jiasheng Ye, Qipeng Guo, Dahua Lin, Kai Chen

Researchers from Shanghai AI Lab and Fudan University developed LoGos, an AI framework that enables general Large Language Models to achieve human professional-level proficiency in the game of Go while retaining broad general reasoning capabilities. The 32B LoGos model demonstrated 88.6% prediction accuracy on KataGo-Bench-1K, surpassing specialized Go AIs and maintaining high performance across various reasoning benchmarks.

Key insights

• Combining heuristically constructed, reasoning-rich domain data with general CoT reasoning and RL effectively injects expert knowledge into LLMs without sacrificing their general intelligence.
• A multi-tiered, piecewise reward function in RL is crucial for guiding LLMs to develop stable and high-performing strategic reasoning in complex domains.
• Integrating 2D board state representations into LLM inputs effectively mitigates the 'context curse,' allowing models to maintain high accuracy over long game sequences by better comprehending complex spatial information.

Summary by alphaXiv.