KG-QUEST: Knowledge Graph–Enhanced Question Answering and Reasoning in Large Language Models

Udiptaman Das, Krishnasai Atmakuri, Duy Hoang Ho, Yugyung Lee

Keywords: Knowledge Graphs, Medical QA, RAG, LLMs, EAV/ERE, Query-Graph Matching, Dynamic KG, Graph Reasoning

TL;DR: \paragraph*{TL;DR} KG-QUEST enhances medical QA by dynamically constructing knowledge graphs from EAV/ERE triples, achieving state-of-the-art results on MedQA and MMLU with interpretable, verifiable reasoning.

Abstract: Large Language Models (LLMs) achieve strong results on medical and open-domain QA but remain limited by static retrieval and parametric memory, which hinder adaptation to evolving ontologies and multi-hop reasoning. We present KG-QUEST, a framework for Knowledge Graph–Enhanced QA that grounds questions in Entity–Attribute–Value (EAV) and Entity–Relation–Entity (ERE) triples and dynamically constructs an answer-specific knowledge graph during inference. A query graph is softly matched to a global biomedical KG and expanded via hop-limited frontier search with predicate weights, synonym and inverse alignment, and negation-aware pruning to form a minimal, high-support subgraph. Phase I (KG generation) fine-tunes LLaMA 3.1 (8B) with ensemble refinement to produce ontology-aligned triples; answers are then selected by dual grounding—scoring KG paths (and optional text evidence) with hop decay and abstention—yielding explicit evidence chains. On MedQA (USMLE) and MMLU medical subsets, KG-QUEST achieves new state-of-the-art results (93.7% and 92.0% accuracy, respectively), surpassing GPT-4 and Med-PaLM 2 while maintaining verifiability. Beyond medical QA, KG-QUEST demonstrates how LLMs can not only retrieve but also construct and navigate structured knowledge graphs for complex reasoning.

Primary Area: foundation or frontier models, including LLMs

Submission Number: 10431