TsuGO: Probing Search Efficiency in LLM Reasoning via Go Life-and-Death Problems

2026-08-13

AI Summary

Current evaluations of large language model (LLM) reasoning struggle to capture their ability to organize search during inference. This work proposes the first process-level reasoning benchmark based on Go life-and-death problems, leveraging their closed, verifiable, and adversarial solution space to compel models to generate candidates, backtrack, and compare branches—thereby decomposing reasoning chains into structured search trees. The approach uniquely disentangles domain knowledge from search organization capability, introducing a novel evaluation dimension centered on search efficiency, complemented by token-efficiency metrics and visual diagnostic tools. Experiments reveal that existing models still fall short in reliably solving life-and-death problems; while more capable models tend to focus on promising branches early, their overall behavior resembles unguided search rather than efficient, neurally guided strategies.

Abstract

The evaluation of LLM reasoning is moving from final-answer accuracy to process-level assessment, yet existing methods still fail to capture how models plan reasoning paths and allocate reasoning resources--that is, how they organize search. Prior process-level methods focus on the coherence and redundancy of chain-of-thought (CoT), and most benchmark tasks have a single objective solvable by static capabilities such as derivation and tool use, leaving search organization unmeasured. We introduce TsuGO, a process-level reasoning benchmark for evaluating Search Efficiency in LLM reasoning through Go life-and-death problems. These problems provide closed and verifiable solution spaces with an inherent adversarial structure, making candidate generation, response checking, branch comparison, and backtracking necessary parts of reasoning rather than incidental trace patterns. By constraining the solution space, TsuGO disentangles domain knowledge from search organization, parses CoT into a structured search tree, and reports Search Efficiency together with Token Efficiency and other diagnostic metrics and visualizations. Experiments show that current LLMs remain far from stable tsumego solving: stronger models succeed by finding the correct candidate earlier and sustaining effort on productive branches, but most models still behave much closer to unguided search algorithms than to neural-guided KataGo. Longer CoT or higher Token Efficiency does not necessarily imply better search. Our results identify search organization and reasoning-resource allocation as missing dimensions in LLM reasoning evaluation.

Problem

• search efficiency
• LLM reasoning
• reasoning evaluation
• Go life-and-death problems
• search organization

Innovation

• Search Efficiency
• Go Life-and-Death Problems
• Chain-of-Thought
• Reasoning Evaluation
• Structured Search Tree