Knowledge-enhanced MCTS for LLM-based Medical Diagnosis Reasoning

ICLR 2026 Conference Submission7349 Authors

Keywords: large language models, medical diagnosis, knowledge-enhanced reasoning

Abstract: Medical diagnosis is a high-stakes, knowledge-intensive task that requires precise reasoning over complex patient information. While Large Language Models (LLMs) have shown promise across a range of medical applications, their ability to perform accurate and interpretable diagnostic reasoning remains limited. Existing LLM-based approaches often rely on shallow, single-step inferences and lack mechanisms to systematically evaluate multiple diagnostic hypotheses. To address these challenges, we propose Med-MCTS, a knowledge-enhanced diagnostic reasoning framework that integrates Monte Carlo Tree Search (MCTS) with external medical knowledge. Med-MCTS formulates diagnosis as a sequential decision-making process and introduces domain-specific state and action representations that align with clinical reasoning practices. During MCTS tree expansion, the model traverses structured medical knowledge graphs to enrich reasoning trajectories with relevant contextual information. To select high-quality paths, Med-MCTS employs a multi-dimensional scoring mechanism that evaluates self-consistency, factual accuracy, and diversity of reasoning. Experiments on multiple benchmark datasets demonstrate that Med-MCTS significantly improves diagnostic accuracy, enabling open-source LLMs to outperform domain-specific medical models and approach the performance of advanced proprietary systems such as GPT-4o.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 7349