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

AI summary

Overview

Research area: Natural language processing / large language model training, applied to the domain of the board game Go.

Technical level: Advanced (mixed supervised fine-tuning, long chain-of-thought distillation data, and GRPO reinforcement learning), though the paper's central question is stated in plain terms.

Scope: One-sentence scope: the paper argues that general LLMs can reach expert-level performance in a data-scarce specialized domain (Go) by combining heuristically constructed structured expert data with general long chain-of-thought reasoning data, and demonstrates this with a model named LoGos.

Paper identity: arXiv:2601.16447v1 [cs.CL], 23 Jan 2026, by Yichuan Ma, Linyang Li, Yongkang Chen, Peiji Li, Jiasheng Ye, Qipeng Guo, Dahua Lin, and Kai Chen, with affiliations including Shanghai Artificial Intelligence Laboratory, Fudan University, and The Chinese University of Hong Kong.

What This Paper Is About

Large language models solve mathematics and coding at or above human-expert level, but they perform "far below even beginner level" at Go, where AlphaGo established a very high ceiling. The problem is data scarcity: unlike math or code, Go offers little natural-language human reasoning, mostly structured professional data, so pretraining and distillation alone do not teach the model the game.

The goal is to close the gap between a general model's broad reasoning ability and the expert knowledge needed in a specialized domain, producing a single LLM that both plays Go at a professional-comparable level and keeps its general reasoning skills.

Key Contributions

1. LoGos, a general LLM with expert-level Go ability. Both a 7B and a 32B version are trained to play Go in natural language with strategic reasoning and next-move prediction, while retaining strong general reasoning performance.
2. A training strategy for domains with only structured data. The method uses large-scale, heuristically constructed expert-level synthetic data for a cold start, then reinforcement learning to align that injected knowledge with the model's existing reasoning behavior; the paper validates the strategy on Go.
3. New datasets and a benchmark. The authors build and state they will release the first large-scale Go dataset for LLM training (a 10-million-scale next-step prediction set plus a 100K-scale commentary set) and the first LLM Go evaluation benchmark, KataGo-Bench-1K.
4. Analysis of training dynamics and input representation. The paper studies self-exploration during RL, cold start with and without heuristic structure, reward design, data mixing ratios, and a "context curse" that is mitigated by rendering the board as a 2-D array.

Main Findings

• Go performance far exceeds all existing general LLMs. On KataGo-Bench-1K, LoGos(7B) scores 88.1 and LoGos(32B) scores 88.6. The strongest general model apart from LoGos is Claude3.7-Sonnet at 34.3, followed by o1-mini at 27.3 and DeepSeek-R1 at 17.6. Base models score 1.4 (Qwen2.5-7B-Base), 1.5 (Qwen2.5-32B-Base), and 4.7 (DeepSeek-R1-Distill-Qwen-32B). The paper states this is nearly 2.6 times the accuracy of Claude3.7-Sonnet.

• LoGos matches specialized Go models at the top of their range. KataGo-HumanSL-18k scores 67.4, KataGo-HumanSL-1d 79.7, KataGo-HumanSL-5d 85.5, and KataGo-HumanSL-9d 87.8. LoGos(32B) at 88.6 and LoGos(7B) at 88.1 exceed KataGo-HumanSL-9d, which is described as simulating top amateur and professional players. The KataGo-HumanSL series covers 27 distinct proficiency levels from 18k to 9d.

• General reasoning is largely preserved. LoGos(32B) scores 63.6 on GPQA-Diamond, 34.1 on BBEH, 74.8 on KOR-Bench, 56.7 on AIME, 96.5 on MATH, and 50.9 on LiveCodeBench. LoGos(7B) scores 37.9, 22.1, 65.7, 40.0, 93.2, and 23.4 respectively. The paper reports the 7B and 32B LoGos models outperform the Qwen2.5 Instruct counterparts on all general tasks evaluated and surpass the DeepSeek-R1-Distill-Qwen models on most benchmarks.

• Cold start is necessary. Applying GRPO directly to instruction models without cold start improves prediction accuracy but never reaches beginner-level performance, cited as 67.36%; the authors note the self-exploration ceiling for DeepSeek's distilled models was lower than for Qwen's instruction models.

• Heuristic structure in the cold-start data matters. Replacing heuristic rules with naive direct prediction for cold start yields a capability ceiling below 50% on the 7B model, compared with 88% when heuristic rules are used.

• RL produces rapid self-exploration gains. Starting from mixed fine-tuned models with initial performance between 60% and 70%, both model sizes quickly find a stable reasoning strategy, surpass KataGo-HumanSL-1d, and eventually exceed KataGo-HumanSL-9d.

• Reward shaping has a measurable effect. Ablations with rewards only for the top-1 move, only for the top-3 moves without distinction, and tiered rewards with intra-level terms set to zero (α1 and α2 zero) all enabled self-exploration, but the sparse top-1 and top-3 designs produced lower performance than the tiered and original reward functions; the authors' original design reached a higher performance ceiling.

• A "context curse" limits long games, and 2-D rendering mitigates it. Because move sequences grow long and any swap of two moves can change the board, accuracy drops as games lengthen for models reading move lists, while engine-style models that represent the board directly do not degrade. Providing a 19×19 2-D board array (1 = black stone, -1 = white stone, 0 = empty) kept prediction accuracy high even for sequences exceeding 200 moves.

• Go data volume does not measurably harm general reasoning. Table 3 reports MATH (97.0 without Go, 97.1 at 500K, 96.2 at 2M, 97.4 at 4M, 97.4 at 10M), AIME (73.2, 72.3, 72.5, 67.9, 72.9), OmniMath (65.0, 65.1, 66.4, 66.3, 66.3), and GPQA-Diamond (61.9, 61.2, 60.1, 59.6, 60.6).

• More Go epochs do not help the final ceiling. Under one epoch of general reasoning data, models initialized with 2 epochs of Go data started stronger but did not learn better during RL and ended lower than the single-epoch model; the one-epoch setting was adopted, also reducing compute.

• Human evaluation shows a gap between move quality and explanation quality. Top amateur Go players rated responses: accurate move predictions in 96.5% of cases, with 55.6% of explanations judged correctly given. The table's counts span correct, ambiguous, and incorrect categories for both moves and explanations. The paper attributes explanation errors partly to misuse of Go's more than 600 specialized terms and to vague opening-phase comments such as "establishing strong positional influence."

• Win-rate comparisons against specific opponents are presented in Figure 1. Opponents include DeepSeek-R1, o1-mini, Claude3.7-Sonnet, the KataGo-HumanSL series, and Golaxy-GiantElephant, described as comparable to mid-level professional players. The specific win-rate numbers are not reported in the provided text.

Methodology in Plain English

The authors treat a Go game as a text sequence: moves are written as a letter-number coordinate with X for black and O for white, and the model's job is to predict the next move given the preceding list. They note Go's complexity as roughly b^d with b ≈ 250 and d ≈ 150, far above chess at b ≈ 35 and d ≈ 80.

To get expert data without human annotation, they collect over 5 million game records from top amateur and professional players, uniformly sample over 10 million game states, and label each with the open-source engine KataGo, which uses Monte Carlo Tree Search and returns the top 10 most probable next moves, following variations, and numerical analysis. These annotations are turned into a four-part template: confirm whose turn it is, analyze several candidate moves, summarize and pick the best one, and emit a structured result (next player, next position, win rate). Separately, 100K commentary cases are collected and processed so the model learns to comment on a position with correct terminology.

The cold start mixes the Go data with long chain-of-thought data from code, math, and general reasoning, so one fine-tuning stage both injects Go knowledge and teaches long-form reasoning. Base models are Qwen2.5-7B-Base and Qwen2.5-32B

Authors’ 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.