TsuGO: New Benchmark Measures LLM Search Efficiency via Go Life-and-Death Problems
A new research paper introduces TsuGO, a process-level reasoning benchmark designed to evaluate search efficiency in large language models (LLMs) using Go life-and-death problems. The paper, available on arXiv (2608.13221), argues that current evaluation methods focus on final-answer accuracy or coherence of chain-of-thought, but fail to measure how models organize search—planning reasoning paths and allocating resources. TsuGO addresses this by presenting problems with closed, verifiable solution spaces and an adversarial structure, requiring candidate generation, response checking, branch comparison, and backtracking. By constraining the solution space, the benchmark disentangles domain knowledge from search organization, offering a more precise assessment of reasoning processes. The work is significant for AI research, particularly in evaluating LLM reasoning capabilities beyond simple answer accuracy.
Key facts
- TsuGO is a process-level reasoning benchmark for evaluating search efficiency in LLM reasoning.
- It uses Go life-and-death problems, which have closed and verifiable solution spaces.
- The benchmark requires candidate generation, response checking, branch comparison, and backtracking.
- TsuGO disentangles domain knowledge from search organization.
- The paper is available on arXiv with identifier 2608.13221.
- It was announced as a new paper on arXiv.
- The evaluation moves from final-answer accuracy to process-level assessment.
- Existing methods fail to capture how models plan reasoning paths and allocate resources.
Entities
Institutions
- arXiv