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

A new evidence-organization layer called HKVM-RAG treats multi-hop retrieval-augmented generation as a data-engineering problem, using answer-path hyperedges as retrieval keys to improve evidence assembly over standard pairwise graph methods, according to a paper submitted June 5, 2026 [1]. The system constructs answer-path hyperedges from cached passage-level LLM evidence tuples and uses them as retrieval keys, while retaining passage text as answer values [1]. The design separates the key space from the value space: extracted entities, pairwise edges, answer-path hyperedges, and confidence weights form the retrieval-key side, while passage text remains the value side used for answer extraction [2]. This key-value-separated memory abstraction is operational rather than biological, naming a structured evidence-indexing mechanism [3]. To isolate the effect of key-space design, the researchers employed a fixed-substrate protocol that holds the tuple cache, candidate passages, reader, and evaluation budget constant across pairwise graph and hypergraph variants [1]. Under this protocol, weighted hypergraph key-value retrieval improved over KG-PPR by +3.426 F1 on 2WikiMultiHopQA and +3.592 F1 on MuSiQue [1]. On HotpotQA, however, higher structured support coverage did not translate into standalone answer-F1 gains, leading the authors to study the hypergraph signal as an evidence-control mechanism rather than a dense-retrieval replacement [2]. Oracle decomposition revealed large support-selection headroom, and train-to-dev support scoring recovered part of it [1]. A dense-aware controller then combined frozen ColBERTv2 and HKVM rank and score features using out-of-fold HKVM predictions [1]. The controller reached 88.846 F1 on 2WikiMultiHopQA, 65.073 F1 on MuSiQue, and 85.810 F1 on HotpotQA, improving over ColBERTv2 by +11.084, +6.763, and +5.966 F1 respectively [1]. Source-level ablations showed that matched non-WHG structured signals did not match the WHG-KV gains, indicating the improvement is specific to the hypergraph key-value design under the tested protocol [2]. The paper frames multi-hop RAG as a fixed-budget evidence-organization problem rather than only passage matching [3]. Dense retrievers score passages independently, while graph-based memories make associations explicit but often rely on pairwise or entity-centered keys that fragment multi-hop evidence [1]. HKVM-RAG instead uses a single answer-path hyperedge key that maps back to multiple passage values, contrasting with query-seeded pairwise KG keys that link to passages through provenance edges [4]. The contribution is bounded but useful: fixed-substrate evidence that answer-path hypergraph keys can improve evidence organization over pairwise KG keys, and cached train-to-dev diagnostics showing WHG-KV can act as a dense-aware evidence-control signal [5].