ReGraM: Region-First Knowledge Graph Reasoning for Medical Question Answering

TL;DR

ReGraM introduces a region-first knowledge graph reasoning framework that constructs a query-aligned subgraph and confines all multi-hop inference to this region, guided by an evidence-aware, hop-limited workflow with a verification loop. By integrating domain priors, weighted relation selection, and a hard region boundary, it reduces retrieval noise and semantic drift that plague full-graph KG reasoning in medical QA. Across seven medical QA benchmarks, ReGraM achieves consistent accuracy gains (MCQ +8.04, SAQ +4.50) and substantial hallucination reductions (up to 42.9% IS reduction) while also improving efficiency (≈46% latency reduction). The approach demonstrates that constraining reasoning to a carefully built region, coupled with an explicit verification mechanism, yields more factual and stable medical QA without requiring large-scale model fine-tuning. The region-first paradigm likely generalizes beyond medicine to other domains with rich relational structure, offering a practical route to safer, more reliable AI-assisted medical reasoning.

Abstract

Recent studies in medical question answering (Medical QA) have actively explored the integration of large language models (LLMs) with biomedical knowledge graphs (KGs) to improve factual accuracy. However, most existing approaches still rely on traversing the entire KG or performing large-scale retrieval, which introduces substantial noise and leads to unstable multi-hop reasoning. We argue that the core challenge lies not in expanding access to knowledge, but in identifying and reasoning over the appropriate subset of evidence for each query. ReGraM is a region-first knowledge graph reasoning framework that addresses this challenge by constructing a query-aligned subgraph and performing stepwise reasoning constrained to this localized region under multiple evidence aware modes. By focusing inference on only the most relevant portion of the KG, ReGraM departs from the assumption that all relations are equally useful an assumption that rarely holds in domain-specific medical settings. Experiments on seven medical QA benchmarks demonstrate that ReGraM consistently outperforms a strong baseline (KGARevion), achieving an 8.04% absolute accuracy gain on MCQ, a 4.50% gain on SAQ, and a 42.9% reduction in hallucination rate. Ablation and qualitative analyses further show that aligning region construction with hop-wise reasoning is the primary driver of these improvements. Overall, our results highlight region-first KG reasoning as an effective paradigm for improving factual accuracy and consistency in medical QA.

Figures (13)

• Figure 1: Overview of the ReGraM framework. Given a medical query, ReGraM constructs a query-aligned local knowledge graph region and performs reasoning exclusively within the selected region to generate the final answer.
• Figure 2: Query Domain Classification Prompt. Classifies a question into one of five high-level biomedical domains (GENE_PROTEIN, DRUG_THERAPY, etc.). The predicted category is later used for relation weighting and KG region selection
• Figure 3: Multi-hop CoT Decomposition Prompt. Decomposes a query into up to three logical sub-questions (hops) using a one-shot Chain-of-Thought prompt. Serves as the core step for constructing ReGraM’s reasoning path
• Figure 4: Hypothetical Triplet Generation Prompt (HYBRID Mode). Used when insufficient KG evidence is retrieved. The LLM conservatively generates triplets using only allowed entities and relations while avoiding new entity creation
• Figure 5: Triplet Revision Prompt (Reviewer Loop). Used when a generated triplet is judged as false by the Reviewer. Encourages correcting schema-violating relations and removing unnecessary hallucinated content