Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Mixing Expert Knowledge: Bring Human Thoughts Back To the Game of Go

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

TL;DR

LoGos demonstrates that a general LLM can reach human professional-level Go performance while preserving strong general reasoning by mixing expert Go data with long chain-of-thought data and applying reinforcement learning (GRPO) for self-exploration. The approach relies on two large Go datasets (Next Step Prediction and Commentary), a cold-start fine-tuning phase, and a reward-driven RL stage aligned to KataGo annotations. Results show LoGos outperforms all existing LLMs on Go benchmarks and maintains competitiveness on broad reasoning tasks, suggesting a viable path to domain-specific mastery without sacrificing generality. The work provides a practical framework and released resources to enable broader adoption of mixing-domain expertise with general reasoning in LLMs.

Abstract

Large language models (LLMs) have demonstrated exceptional performance in reasoning tasks such as mathematics and coding, matching or surpassing human capabilities. However, these impressive reasoning abilities face significant challenges in specialized domains. Taking Go as an example, although AlphaGo has established the high performance ceiling of AI systems in Go, mainstream LLMs still struggle to reach even beginner-level proficiency, let alone perform natural language reasoning. This performance gap between general-purpose LLMs and domain experts is significantly limiting the application of LLMs on a wider range of domain-specific tasks. In this work, we aim to bridge the divide between LLMs' general reasoning capabilities and expert knowledge in domain-specific tasks. We perform mixed fine-tuning with structured Go expertise and general long Chain-of-Thought (CoT) reasoning data as a cold start, followed by reinforcement learning to integrate expert knowledge in Go with general reasoning capabilities. Through this methodology, we present \textbf{LoGos}, a powerful LLM that not only maintains outstanding general reasoning abilities, but also conducts Go gameplay in natural language, demonstrating effective strategic reasoning and accurate next-move prediction. LoGos achieves performance comparable to human professional players, substantially surpassing all existing LLMs. Through this work, we aim to contribute insights on applying general LLM reasoning capabilities to specialized domains. We will release the first large-scale Go dataset for LLM training, the first LLM Go evaluation benchmark, and the first general LLM that reaches human professional-level performance in Go at: https://github.com/Entarochuan/LoGos.

Mixing Expert Knowledge: Bring Human Thoughts Back To the Game of Go

TL;DR

LoGos demonstrates that a general LLM can reach human professional-level Go performance while preserving strong general reasoning by mixing expert Go data with long chain-of-thought data and applying reinforcement learning (GRPO) for self-exploration. The approach relies on two large Go datasets (Next Step Prediction and Commentary), a cold-start fine-tuning phase, and a reward-driven RL stage aligned to KataGo annotations. Results show LoGos outperforms all existing LLMs on Go benchmarks and maintains competitiveness on broad reasoning tasks, suggesting a viable path to domain-specific mastery without sacrificing generality. The work provides a practical framework and released resources to enable broader adoption of mixing-domain expertise with general reasoning in LLMs.

Abstract

Large language models (LLMs) have demonstrated exceptional performance in reasoning tasks such as mathematics and coding, matching or surpassing human capabilities. However, these impressive reasoning abilities face significant challenges in specialized domains. Taking Go as an example, although AlphaGo has established the high performance ceiling of AI systems in Go, mainstream LLMs still struggle to reach even beginner-level proficiency, let alone perform natural language reasoning. This performance gap between general-purpose LLMs and domain experts is significantly limiting the application of LLMs on a wider range of domain-specific tasks. In this work, we aim to bridge the divide between LLMs' general reasoning capabilities and expert knowledge in domain-specific tasks. We perform mixed fine-tuning with structured Go expertise and general long Chain-of-Thought (CoT) reasoning data as a cold start, followed by reinforcement learning to integrate expert knowledge in Go with general reasoning capabilities. Through this methodology, we present \textbf{LoGos}, a powerful LLM that not only maintains outstanding general reasoning abilities, but also conducts Go gameplay in natural language, demonstrating effective strategic reasoning and accurate next-move prediction. LoGos achieves performance comparable to human professional players, substantially surpassing all existing LLMs. Through this work, we aim to contribute insights on applying general LLM reasoning capabilities to specialized domains. We will release the first large-scale Go dataset for LLM training, the first LLM Go evaluation benchmark, and the first general LLM that reaches human professional-level performance in Go at: https://github.com/Entarochuan/LoGos.
Paper Structure (52 sections, 5 equations, 33 figures, 3 tables)

This paper contains 52 sections, 5 equations, 33 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Mixed Learning of Go
  3. Modeling the Game of Go
  4. Expert-Level Go Dataset Collection
  5. Next Step Prediction Dataset
  6. Commentary Dataset
  7. Mixed Cold Start
  8. Self-Exploration with Reinforcement Learning
  9. RL Optimization with GRPO
  10. Reward Modeling
  11. Experiments
  12. Experimental Setup
  13. Datasets
  14. Benchmarks
  15. Models
  16. ...and 37 more sections

Figures (33)

  • Figure 1: Win rates of LoGos against various models. For general LLMs, we select DeepSeek-R1, o1-mini, and Claude3.7-Sonnet for gameplay. Additionally, we play LoGos against several specialized Go models. The KataGo-HumanSL model series katago is designed to simulate human players at different skill levels. KataGo-HumanSL-1k emulates intermediate amateur players, while KataGo-HumanSL-9d mimics top amateur and professional players. Golaxy-GiantElephant is a well-known Go AI model that achieves performance comparable to mid-level professional players.
  • Figure 2: Human evaluation results on the generated responses, analyzing whether the predicted move is correct, and whether the corresponding explanation is correct, incorrect, or ambiguous.
  • Figure 3: Modeling the game of Go.
  • Figure 4: Examples of the next step prediction dataset. The heuristic template consists of four parts: (i) confirming whether the next player is black or white; (ii) analyzing several possible next moves; (iii) summarizing and selecting the optimal next move; and (iv) structured output.
  • Figure 5: Our methodology for integrating Go professional capabilities with LLMs' long COT reasoning abilities. After mixed cold start and GRPO training, our model ultimately successfully transfers the reasoning capabilities acquired from long CoT data to Go tasks. For a given query, the model correctly performs analysis, thinking, reasoning, and summarization, ultimately selecting a reasonable next move.
  • ...and 28 more figures