MedJudgeRAG: Option-Wise Evidence Judgment with Dynamic Knowledge Graphs for Medical MCQA

In medical multiple-choice question answering (MCQA), Retrieval-Augmented Generation (RAG) can supplement the domain knowledge of language models (LMs). However, since vanilla RAG directly prepends retrieved documents to the input, it can degrade LM performance. To address this, we propose MedJudgeRAG. MedJudgeRAG structures retrieved documents into a dynamic knowledge graph (KG) composed of entities and relations. For each option, the model judges an evidence verdict. Based on the verdict combination, the model determines a knowledge utilization strategy to reason toward the final answer. These capabilities are trained via supervised fine-tuning using structured reasoning traces generated by a teacher LM. The training employs a weighted cross-entropy loss that differentially weights the KG and reasoning segments. Experiments on two medical MCQA benchmarks demonstrate that MedJudgeRAG consistently outperforms both vanilla RAG and parametric baselines. Furthermore, ablation analysis reveals that the dynamic KG serves more effectively as graph-conditioned supervision at training time than as an explicit output at inference time.