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

2026-06-08 · Source: cs.CL updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, extended

Summary

HKVM-RAG is a novel key-value-separated evidence-organization layer designed for multi-hop Retrieval-Augmented Generation (RAG) systems. It addresses the challenge of organizing retrieved text into explicit answer chains under fixed retrieval budgets. HKVM-RAG constructs answer-path hyperedges from cached passage-level LLM evidence tuples, using these as higher-order retrieval keys while retaining passage text as values. A controlled protocol isolates key-space design, showing that weighted hypergraph key-value retrieval improves over KG-PPR by +3.426 F1 on 2WikiMultiHopQA and +3.592 F1 on MuSiQue. While HotpotQA showed no standalone F1 gains, a dense-aware controller combining frozen ColBERTv2 and HKVM features achieved 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. Code and data are available on GitHub and Hugging Face.

Key takeaway

For Machine Learning Engineers building multi-hop RAG systems, consider integrating HKVM-RAG's hypergraph evidence organization. Its key-value-separated approach significantly improves answer F1 on benchmarks like 2WikiMultiHopQA and MuSiQue when combined with dense retrievers like ColBERTv2. You should explore using its rank/score outputs as a control signal to reorder retrieved passages, enhancing overall system performance rather than replacing dense retrieval.

Key insights

HKVM-RAG uses hypergraph keys to organize multi-hop evidence, improving RAG performance by complementing dense retrievers.

Principles

• Multi-hop RAG requires explicit evidence organization.
• Hypergraph keys enhance multi-hop evidence structuring.
• Structured signals effectively complement dense retrieval.

Method

HKVM-RAG assembles answer-path hyperedges from LLM-extracted evidence tuples, weights them with confidence, diffuses query-seeded scores, and projects to passage-value scores. A dense-aware controller then reorders passages using combined rank/score features.

In practice

• Use DeepSeek V4 Flash for evidence tuple extraction.
• Combine hypergraph signals with ColBERTv2 for re-ranking.
• Isolate key-space design for controlled RAG experiments.

Topics

• Multi-hop RAG
• Hypergraph Neural Networks
• Evidence Organization
• Retrieval-Augmented Generation
• ColBERTv2
• Knowledge Graphs

Code references

• Mingyu-Zh/HKVM-RAG

Best for: Research Scientist, AI Architect, AI Engineer, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• HKVM-RAG: Key-Value-Separated Hypergraph Evidence Organization for Multi-Hop RAG — HKVM-RAG uses key-value-separated hypergraphs to organize multi-hop evidence, improving RAG performance by explicitly linking answer paths.
• Self-Augmenting Retrieval for Diffusion Language Models — Low-confidence tokens in discrete diffusion models provide a useful lookahead signal for dynamic retrieval-augmented generation.
• A Better, Cheaper RAG (Neuro-Sym Multi-hop Reasoning) — TGS RAG fuses vector and graph search to overcome semantic drift and broken reasoning paths, reducing costs and improving accuracy.
• RARE: Redundancy-Aware Retrieval Evaluation Framework for High-Similarity Corpora — Existing RAG benchmarks misrepresent real-world performance due to high document redundancy and similarity in enterprise corpora.
• How to integrate a graph database into your RAG pipeline — Graph databases enhance RAG by providing explicit relational understanding for complex, multi-hop reasoning beyond semantic similarity.
• NeocorRAG: Less Irrelevant Information, More Explicit Evidence, and More Effective Recall via Evidence Chains — Optimizing retrieval quality through "Evidence Chains" significantly improves RAG reasoning accuracy and efficiency.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CL updates on arXiv.org.