LLM Reasoning Traces Reveal Myopic Planning in Board Game
A new study from arXiv (2605.06840) introduces a method to extract search trees from LLM reasoning traces in the game four-in-a-row. Researchers found LLM search is shallower than humans, with performance predicted by breadth not depth. Move choices are best explained by a myopic model ignoring deep nodes, confirmed by causal intervention.
Key facts
- Method extracts search trees from LLM reasoning traces
- Applied to four-in-a-row board game
- LLM search is shallower than human search
- Performance predicted by search breadth, not depth
- Move choices best explained by myopic model ignoring deep nodes
- Causal intervention study confirmed findings
- Published on arXiv with ID 2605.06840
Entities
Institutions
- arXiv