MTDW-RAG: A Tree-Structured Medical Q&A Approach Combining Deep Progression and Wide Exploration

As a key technology in Generative Artificial Intelligence (AI), Large Language Models (LLMs) are increasingly recognized for their vast potential in medical question answering applications.However, existing retrieval-augmented generation (RAG) approaches still face significant limitations when dealing with complex reasoning tasks. This paper proposes a tree-structured medical Q&A approach combining Deep Progression and Wide Exploration (MTDW-RAG) to address these issues.MTDW-RAG employs a deep progression strategy for layer-by-layer reasoning while combining a parallelised wide exploration strategy for exploring the problem from multiple perspectives, which improves reasoning efficiency and accuracy. In addition, this paper introduces a web search mechanism that is more suitable for large language models, which further improves the model’s performance in complex medical tasks by incorporating more accurate information retrieval. Experimental results on the MMLU-Med dataset show that MTDW-RAG significantly outperforms existing benchmark methods. Our approach achieves high accuracy in medical question-and-answer tasks, especially performing well when dealing with complex questions. In addition, we provide an in-depth analysis of the key parameters affecting the model performance, which p rovides strong guidance for future model design.