Reproduce “ReGraM: Region-First Knowledge Graph Reasoning for Medical Question Answering” on OpenResearch

About this paper

Chaerin Lee, Sohee Park, Hyunsik Na, Daseon Choi

Soongsil University researchers developed ReGraM, a framework for medical question answering that enforces a 'region-first' approach to knowledge graph reasoning, constructing query-aligned subgraphs to constrain Large Language Model inference. This method achieves an average 8.04% absolute accuracy gain on multiple-choice questions and a 42.9% relative reduction in factual hallucination in short-answer questions compared to prior baselines.

Key insights

• Effective LLM-KG integration for medical QA requires not just access to knowledge, but explicit structural bounding and guidance within KGs to achieve reliable and factually accurate responses.
• Domain-specific priors and multi-hop query decomposition are crucial for building accurate KG regions and structuring reasoning, demonstrating substantial performance degradation when absent.
• A dedicated verification module is essential for hallucination suppression, acting as a conservative generation constraint, though it can sometimes lead to under-answering in cases of sparse KG evidence.

Summary by alphaXiv.