HKVM-RAG: Key-Value-Separated Hypergraph Evidence Organization for Multi-Hop RAG

HKVM-RAG introduces a key-value-separated evidence-organization layer designed to enhance multi-hop retrieval-augmented generation (RAG) tasks by organizing retrieved text into hypergraph structures that reveal answer chains. The model demonstrates significant improvements, achieving +3.426 F1 on 2WikiMultiHopQA and +3.592 F1 on MuSiQue compared to previous methods, while a dense-aware controller combining ColBERTv2 features and HKVM predictions achieves F1 scores of 88.846, 65.073, and 85.810 across three benchmarks. This approach offers a novel evidence-control mechanism that could optimize retrieval processes for practitioners developing multi-hop reasoning systems.