TsuGO is a process-level reasoning benchmark designed to evaluate how Large Language Models (LLMs) organize search and allocate reasoning resources. While existing benchmarks focus on final-answer accuracy or the coherence of a single chain-of-thought, TsuGO uses Go life-and-death problems (tsumego) to force models to navigate complex, adversarial decision spaces where they must generate, verify, and backtrack through multiple potential reasoning paths.
Evaluating Search Organization
The authors, Shunwen Bai et al., argue that current evaluation methods fail to measure "Search Efficiency" (SearchE)—the ability to organize reasoning toward effective solutions across competing trajectories. Because tsumego problems require models to account for an opponent's responses, they create a closed, verifiable environment where branching is a functional necessity rather than a stylistic choice. TsuGO parses a model's free-form chain-of-thought into a structured "process search tree," allowing researchers to track how models move through the solution space, where they waste effort on incorrect branches, and how they prioritize productive paths.
How the Benchmark Works
TsuGO consists of 1,500 problems categorized by difficulty, ranging from shallow kyu-level tasks to complex dan-level problems. The benchmark evaluates models under two conditions: a bounded-candidate setting (selecting from four moves) and an open-generation setting (proposing moves independently).
To analyze performance, the researchers use an extraction pipeline that identifies four atomic reasoning steps: exploration, reading, evaluation, and backtracking. This data is used to calculate:
Search Efficiency (SearchE): A composite metric that weighs the model's ability to avoid wrong branches, identify the correct move early, and rank the correct candidate highly.
Token Efficiency (TokenE): A measure of accuracy relative to the number of tokens consumed during the reasoning process.
Diagnostic Metrics: These include trajectory depth, fan-out, and node counts, which provide a granular view of how a model allocates its "thinking" resources.
Key Findings
Experiments conducted by the authors reveal that current LLMs are not yet proficient at stable tsumego solving. While stronger models tend to identify the correct candidate earlier and sustain effort on productive branches, most models perform more like unguided search algorithms than the neural-guided KataGo engine.
A significant finding is that longer chains-of-thought or higher token efficiency do not guarantee better search organization. The results indicate that even frontier models struggle to effectively allocate resources across adversarial branches, suggesting that search organization remains an underdeveloped dimension in current LLM reasoning.
Limitations and Considerations
The authors note that their evaluation of proprietary models is limited to observable summaries, which may not fully reflect the internal computation of those systems. Additionally, while the benchmark uses Go problems to create a controlled environment, the researchers emphasize that the goal is not to test Go-playing strength, but to use the game's adversarial structure as a proxy for measuring how models manage complex, multi-path reasoning tasks. The benchmark is designed to remain challenging as model capabilities evolve, with harder subsets of the dataset reserved for future, more advanced systems.
Comments (0)
to join the discussion
No comments yet
Be the first to share your thoughts!