Graph-R1 applies end-to-end reinforcement learning to improve LLMs’ graph reasoning capabilities. It constructs a knowledge hypergraph from n-ary relation extraction and uses an explicit reward mechanism to train the model through an iterative cycle of thought, query generation, subgraph retrieval, and reconsideration. This approach bridges the gap between graph-structured knowledge and language modalities, enabling more effective utilization of structured knowledge for knowledge-intensive tasks.