SPIRAL: LLM Planning with Grounded Search

This paper introduces SPIRAL, a novel framework for LLM planning that integrates a cognitive architecture within a Monte Carlo Tree Search (MCTS) loop. It addresses the limitations of LLMs in complex planning tasks by incorporating a Planner, Simulator, and Critic to guide the search process. The key contribution is the synergy between these agents, transforming MCTS into a guided, self-correcting reasoning process. The paper demonstrates significant performance improvements over existing methods on benchmark datasets, highlighting the effectiveness of the proposed approach.