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

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.

💡 Research Summary

The paper introduces PRoH, a novel Retrieval‑Augmented Generation (RAG) framework that operates directly on Knowledge Hypergraphs (KHs). Unlike traditional graph‑based RAG systems that model only binary relations, KHs capture n‑ary facts by allowing hyperedges to connect multiple entities simultaneously. Existing KH‑RAG approaches (e.g., HyperGraphRAG, HGRAG) suffer from three key drawbacks: (1) static retrieval planning that does not adapt to the query or graph context, (2) non‑adaptive, one‑shot retrieval execution that cannot incorporate intermediate reasoning results, and (3) superficial use of hypergraph structure, treating hyperedges merely as routing mechanisms without exploiting their rich semantics.

PRoH addresses these limitations through three core innovations. First, a context‑aware planning module sketches the local neighborhood of the topic entities in the hypergraph before any question decomposition. By feeding this subgraph summary to a large language model (LLM), the system generates a feasible reasoning plan that aligns with the actual graph topology, reducing the mismatch between linguistic cues and available structured knowledge.

Second, PRoH employs a structured question decomposition that organizes sub‑questions into a dynamically evolving Directed Acyclic Graph (DAG). Each node represents a sub‑question, and edges encode logical precedence. As reasoning proceeds, the DAG is iteratively refined: new sub‑questions may be added, existing ones updated, and dependencies re‑wired based on intermediate answers. This dynamic DAG enables multi‑trajectory exploration, allowing the system to maintain several candidate reasoning paths simultaneously—a crucial capability for handling the ambiguity inherent in n‑ary relations.

Third, the framework introduces an Entity‑Weighted Overlap (EWO)‑guided retrieval algorithm. When traversing from one hyperedge to its neighbors, PRoH computes an overlap score that weights each shared entity by its semantic similarity to the current sub‑question (using entity embeddings). The EWO score prioritizes hyperedges that contribute meaningfully to the query, rather than relying solely on structural overlap. Consequently, the retrieved reasoning paths are semantically coherent and better aligned with the question’s intent.

The overall pipeline consists of four stages: (1) Graph construction and indexing, where documents are parsed into entities and hyperedges, and synonym hyperedges are added to improve connectivity; (2) Graph anchoring, which extracts a question‑specific subgraph; (3) Planning, which generates the initial DAG‑based reasoning plan using the context‑aware module; (4) Reasoning with graph retrieval, where a state‑space search explores multiple DAG branches, guided by EWO, and iteratively refines the DAG; and finally (5) Answer generation, where the completed DAG and its associated text chunks are fed to the LLM to produce the final answer.

Experiments were conducted on several multi‑domain benchmarks, including open‑domain QA (HotpotQA), medical QA, and legal QA. PRoH consistently outperformed the previous state‑of‑the‑art HyperGraphRAG, achieving an average 19.73 % absolute gain in F1 and 8.41 % absolute gain in Generation Evaluation (G‑E) score. The advantage is especially pronounced on long‑range multi‑hop questions (four or more hops), where PRoH’s dynamic planning and EWO retrieval reduce error propagation and improve recall of relevant facts. Ablation studies confirm the contribution of each component: removing context‑aware planning drops F1 by ~12 %; replacing the DAG with a linear sequence reduces F1 by ~9 %; and using plain entity overlap instead of EWO lowers F1 by ~9 %. Adding synonym hyperedges improves graph connectivity and yields a modest 4 % performance boost.

The authors also discuss limitations. The planning stage incurs additional token cost because the LLM must process a subgraph sketch, and the dynamic DAG can grow large for highly complex queries, potentially increasing search space. Moreover, the quality of EWO depends on the underlying entity embeddings, which may be suboptimal for niche domains. Future work aims to compress planning prompts, introduce cost‑benefit based pruning for DAG expansion, and learn domain‑specific entity representations to further enhance robustness.

In summary, PRoH presents a comprehensive solution that fully leverages the expressive power of knowledge hypergraphs for retrieval‑augmented generation. By integrating context‑aware planning, dynamically refined DAG‑based question decomposition, and semantically weighted hypergraph traversal, it achieves state‑of‑the‑art performance on multi‑hop QA while offering greater interpretability and adaptability than prior static pipelines. The framework opens new avenues for applying hypergraph‑centric reasoning in large‑scale, real‑world AI systems.