Scholarlink

<p>Retrieval-Augmented Generation (RAG) is commonly used to mitigate factual hallucinations in large language models, but it remains constrained by flat retrieval, Euclidean hierarchical distortion, and candidate-level isolated decision-making in complex multi-hop reasoning, making it difficult to consistently preserve hierarchical topology, higher-order relations, and stable chains of evidence. To address this, this paper proposes Hyperbolic Hierarchical Structure-aware Retrieval-Augmented Generation (HHS-RAG), a unified framework that integrates hierarchical hypergraph modeling, hyperbolic structure representation learning, and subgraph-level structural decision-making. Specifically, we first construct a hierarchical hypergraph that integrates document hierarchies, topic abstractions, and high-order co-occurrences of multiple entities. We then design Hyperbolic Hierarchical Graph Contrastive Learning (HHGCL) to align macro-level topic hyperedges with micro-level fact hyperedges in a shared hyperbolic space, thereby preserving hierarchical depth and structural consistency. During the query phase, by combining dual-path structure-aware retrieval to enhance entity expansion and path-level recall, we propose Query-driven Structural Decision Subgraph (QSDS), which elevates the final answer selection from candidate-level ranking to subgraph-level structural decision-making. Experimental results demonstrate that HHS-RAG consistently outperforms baselines on PCS and AER, achieving strong results on multiple complex question-answering benchmarks. These findings further show that unified hierarchical structure modeling, dual-path retrieval, and subgraph-level decision collectively enhance structured reasoning capabilities in complex multi-hop question-answering tasks.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

HHS-RAG: Hierarchical hypergraph retrieval-augmented generation with hyperbolic contrastive learning and subgraph-level decision

  • 期刊论文

  • 发表日期：

    2026年7月29日

查看全文
  • Tianci Zhang,
  • Jianbin Wu,
  • Yueguang Kang

摘要

Retrieval-Augmented Generation (RAG) is commonly used to mitigate factual hallucinations in large language models, but it remains constrained by flat retrieval, Euclidean hierarchical distortion, and candidate-level isolated decision-making in complex multi-hop reasoning, making it difficult to consistently preserve hierarchical topology, higher-order relations, and stable chains of evidence. To address this, this paper proposes Hyperbolic Hierarchical Structure-aware Retrieval-Augmented Generation (HHS-RAG), a unified framework that integrates hierarchical hypergraph modeling, hyperbolic structure representation learning, and subgraph-level structural decision-making. Specifically, we first construct a hierarchical hypergraph that integrates document hierarchies, topic abstractions, and high-order co-occurrences of multiple entities. We then design Hyperbolic Hierarchical Graph Contrastive Learning (HHGCL) to align macro-level topic hyperedges with micro-level fact hyperedges in a shared hyperbolic space, thereby preserving hierarchical depth and structural consistency. During the query phase, by combining dual-path structure-aware retrieval to enhance entity expansion and path-level recall, we propose Query-driven Structural Decision Subgraph (QSDS), which elevates the final answer selection from candidate-level ranking to subgraph-level structural decision-making. Experimental results demonstrate that HHS-RAG consistently outperforms baselines on PCS and AER, achieving strong results on multiple complex question-answering benchmarks. These findings further show that unified hierarchical structure modeling, dual-path retrieval, and subgraph-level decision collectively enhance structured reasoning capabilities in complex multi-hop question-answering tasks.