This paper proposes Dynamically Adaptive MCTS-based Reasoning (DAMR), a novel framework for Knowledge Graph Question Answering (KGQA). To overcome the limitations of existing retrieve-then-reason and LLM-based dynamic path generation methods, DAMR integrates Monte Carlo Tree Search (MCTS)-based symbolic search with adaptive path evaluation. An LLM-based planner selects the top k relevant relations at each step to reduce the search space, and a lightweight transformer-based scorer performs context-aware likelihood estimation by co-encoding question and relation sequences. Furthermore, a dynamic pseudopath refinement mechanism mitigates the lack of high-quality supervised data. Experimental results demonstrate that DAMR significantly outperforms existing state-of-the-art methods.
