PRoH: Dynamic Planning and Reasoning over Knowledge Hypergraphs for Retrieval-Augmented Generation

TL;DR

This work tackles the limitations of static and shallow KH-based RAG by introducing PRoH, a dynamic framework that integrates context-aware planning, structured DAG-based question decomposition, and EWO-guided reasoning path retrieval over Knowledge Hypergraphs. By sketching a plan-sensitive KH neighborhood, evolving reasoning DAGs, and semantically informed traversal, PRoH enables adaptive, multi-hop reasoning that better preserves higher-order relations. Empirical results across five domains show substantial improvements over the prior SOTA HyperGraphRAG in F1 (average +19.73%) and G-E (average +8.41%), with strong performance on long-range QA and efficient variants like PRoH-L that reduce token usage. Overall, PRoH advances KH-based RAG by enhancing plan feasibility, reasoning flexibility, and retrieval relevance, improving accuracy and interpretability in retrieval-augmented generation systems.

Abstract

Knowledge Hypergraphs (KHs) have recently emerged as a knowledge representation for retrieval-augmented generation (RAG), offering a paradigm to model multi-entity relations into a structured form. However, existing KH-based RAG methods suffer from three major limitations: static retrieval planning, non-adaptive retrieval execution, and superficial use of KH structure and semantics, which constrain their ability to perform effective multi-hop question answering. To overcome these limitations, we propose PRoH, a dynamic Planning and Reasoning over Knowledge Hypergraphs framework. PRoH incorporates three core innovations: (i) a context-aware planning module that sketches the local KH neighborhood to guide structurally grounded reasoning plan generation; (ii) a structured question decomposition process that organizes subquestions as a dynamically evolving Directed Acyclic Graph (DAG) to enable adaptive, multi-trajectory exploration; and (iii) an Entity-Weighted Overlap (EWO)-guided reasoning path retrieval algorithm that prioritizes semantically coherent hyperedge traversals. Experiments across multiple domains demonstrate that PRoH achieves state-of-the-art performance, surpassing the prior SOTA model HyperGraphRAG by an average of 19.73% in F1 and 8.41% in Generation Evaluation (G-E) score, while maintaining strong robustness in long-range multi-hop reasoning tasks.

Figures (9)

• Figure 1: Illustration of Knowledge Hypergraph.
• Figure 2: An overview architecture of PRoH. Given the input question and KH, the workflow begins with Graph Anchoring (green) and constructs a question subgraph. The following Plan Initialization Module (purple) sketches the question subgraph and obtains context to generate initial reasoning plans and constructs reasoning DAGs for the question. Those DAGs serve as roots of the state search trees in the Reasoning stage (gray). At each state within the State Search Tree, the model retrieves answer-path pairs from the KH and then iteratively completes and refines the DAGs to transit into the next state until one or more DAGs are completed. These completed DAGs are then passed to the final answer generation stage (yellow) as the retrieved knowledge for producing the final answer.
• Figure 3: Impact of depth limit $d_{\max}$ on performance.
• Figure 4: F1 score vs. length of ground-truth context.
• Figure 5: F1 score vs. #entities in ground-truth context.

Theorems & Definitions (7)

• Definition 1: Hypergraph-based RAG
• Definition 2: Reasoning Path
• Definition 3: Reasoning Plan
• Definition 4: Plan Context Graph
• Definition 5: Reasoning DAG
• Definition 6: Completed Reasoning DAG
• Definition 7: Reasoning State