MedJudgeRAG: Dynamic Knowledge Graphs for Medical MCQA

A new framework called MedJudgeRAG improves medical multiple-choice question answering (MCQA) by dynamically constructing knowledge graphs from retrieved documents. Unlike standard Retrieval-Augmented Generation (RAG), which indiscriminately uses all retrieved text, MedJudgeRAG builds a dynamic knowledge graph of entities and relations. For each answer option, the model evaluates evidence from both the documents and the graph, then selects a knowledge utilization strategy to reason toward the correct answer. The system is trained via supervised fine-tuning on structured reasoning traces generated by a teacher language model, using a weighted cross-entropy loss that differentially weights the knowledge graph and reasoning segments. Experiments on two medical MCQA benchmarks show consistent outperformance over vanilla RAG. The paper is published on arXiv with ID 2607.24838.