A Query-Driven Graph Retrieval Framework with Adaptive Pruning for Multi-Hop Question Answering

Multi-hop question answering (MHQA) requires models to retrieve and reason over evidence distributed across multiple documents, which remains challenging for conventional retrieval-augmented generation (RAG) approaches. Although RAG improves factual grounding by incorporating external knowledge, flat retrieval strategies often struggle to maintain coherent reasoning chains when implicit dependencies among entities and documents are involved. This paper presents a query-driven dual-layer graph retrieval framework for MHQA. The framework operates on a unified heterogeneous graph integrating entities, relations, and supporting texts, and dynamically constructs candidate subgraphs through joint retrieval over entities and relations, complemented by lexical retrieval signals. Reasoning paths are refined by combining structural strength modeling with contrastive learning-based path scoring, and an adaptive pruning strategy is employed to regulate evidence scale according to query complexity and path score distributions. Experiments on HotpotQA and 2WikiMultihopQA show that the proposed framework achieves higher EM and F1 scores than existing RAG and graph-based retrieval methods, particularly in complex multi-hop scenarios. These results indicate the importance of structured and query-adaptive evidence organization for multi-hop reasoning.