Title: Graph-R1: An Agentic GraphRAG Framework Driven by End-to-End Reinforcement Learning

Abstract: While Retrieval-Augmented Generation (RAG) reduces hallucinations in Large Language Models (LLMs) by integrating external knowledge, it is limited by chunk-based retrieval methods that fail to capture structural semantics. Although GraphRAG approaches enhance RAG by representing knowledge as entity-relation graphs, they continue to struggle with high construction costs, static one-time retrieval mechanisms, and a heavy dependence on long-context reasoning and prompt engineering. To overcome these limitations, we introduce Graph-R1, the pioneering agentic GraphRAG framework utilizing end-to-end reinforcement learning (RL). This framework features a lightweight knowledge hypergraph construction process, treats retrieval as a multi-turn interaction between an agent and its environment, and refines the agent’s workflow through a comprehensive end-to-end reward system. Evaluations on standard RAG benchmarks demonstrate that Graph-R1 surpasses conventional GraphRAG and RL-enhanced RAG techniques in terms of reasoning accuracy, retrieval efficiency, and output quality. The associated code and data are publicly accessible at https://github.com/LHRLAB/Graph-R1.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC