PARM: Passage-Anchored Relation-Aware Multi-Hop Retrieval for Evidence-Grounded Question Answering

Multi-hop question answering requires discovering interconnected facts from distributed text passages and entity relations, and constructing complete evidence chains to support the final answer. Existing graph-based retrieval-augmented generation approaches often rely on limited initial retrieval seeds or semantic anchors for graph expansion, making evidence discovery sensitive to anchor coverage, candidate generation strategies, and early filtering decisions. To address these challenges, we propose PARM, a graph retrieval framework that introduces passage-level evidence anchors before graph reasoning. Specifically, PARM first identifies relevant passages through lexical retrieval and semantic reranking, and then combines question entities with passage-linked entities to construct complementary graph search anchors. A bounded multi-hop graph search is subsequently performed to generate explicit relation paths under predefined hop and path budgets. The generated paths are further ranked by jointly modeling semantic relevance, question–relation compatibility, and graph structural information. Finally, the selected passages, relation paths, and provenance information are organized into a traceable evidence package for answer generation. Extensive experiments are conducted on the MuSiQue and 2WikiMultiHopQA benchmarks. Across three fixed 1000-question subsets of 2WikiMultiHopQA, PARM increases the mean EM score from 0.3403 to 0.3963 and the mean token-level F1 score from 0.3879 to 0.4676 relative to QASA.