BASIN Prevents Reasoning Basin Collapse in Inference-Time LLM Search

BASIN is a training-free method that prevents LLM search from concentrating on redundant reasoning paths by grouping states into basins and penalizing repeated strategies. It outperforms Tree of Thoughts by up to 22 percentage points on Game of 24 by reallocating compute to distinct reasoning trajectories.