KGReasoner: Knowledge Graph-Enhanced Retrieval and Reasoning with Large Language Models

Knowledge-intensive question answering requires both accurate retrieval of relevant evidence and coherent multi-hop reasoning over interconnected facts. While retrieval-augmented generation (RAG) has emerged as a promising paradigm for grounding large language models (LLMs) with external knowledge, conventional RAG methods rely primarily on semantic similarity for retrieving isolated text chunks, which often results in disconnected fragments that lack the relational structure necessary for complex reasoning. Knowledge graphs (KGs) provide structured representations of entities and their relationships, offering a natural complement to text-based retrieval. In this paper, we propose KGReasoner, a unified framework that integrates knowledge graph-enhanced retrieval with structured reasoning for knowledge-intensive question answering. KGReasoner comprises three key components: (1) a KG-guided subgraph retrieval module that efficiently identifies query-relevant subgraphs using a lightweight MLPbased triple scorer with directional structural distance encoding and dynamic subgraph size adjustment; (2) a structured context organization module that transforms retrieved subgraphs into coherent reasoning paths by extracting relation chains from query entities to answer candidates, and augmenting these paths with associated text chunks to create hybrid structured-unstructured contexts; and (3) a weak-to-strong reasoning alignment module that improves LLM reasoning capability on structured contexts through debate-augmented training and ensemble verification. Extensive experiments on three benchmark datasets-WebQSP, ComplexWebQuestions, and HotpotQA-demonstrate that KGReasoner achieves state-of-theart performance, outperforming existing methods including SubgraphRAG with GPT-4o by 1.7 Hit@1 points on WebQSP, 2.3 Hit@1 points on CWQ, and 2.3 EM points on HotpotQA. Ablation studies confirm the contribution of each component, and analysis shows that the improvements are particularly significant for complex multi-hop reasoning questions.