Multi-hop GraphRAG-QA for evidence-grounded biomedical reasoning

Abstract

Large language models achieve strong results in biomedical question answering but often lack interpretability and reliable grounding. We present Multi-hop GraphRAG-QA, a framework that integrates structured biomedical knowledge graphs with LLMs for evidence-based reasoning. The method retrieves and serializes multi-hop relational subgraphs to guide answer generation. Experiments on MedMCQA, MedQA, and BioASQ show that graph-based retrieval improves fidelity and reduces hallucinations for concise questions, but deeper multi-hop expansion can degrade performance on complex clinical cases due to evidence dilution. Qualitative analysis highlights the benefits and limitations of structured multi-hop retrieval for trustworthy biomedical QA.