Research map / Large language models / Retrieval and knowledge
Graph-R1: Towards Agentic GraphRAG Framework via End-to-end Reinforcement Learning
Haoran Luo, Haihong E, Guanting Chen, Qika Lin, Yikai Guo, Fangzhi Xu, Zemin Kuang, Meina Song, Xiaobao Wu, Yifan Zhu, Anh Tuan Luu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add knowledge construction (K.C.) of a hypergraph (ablation) | Graph-R1 · Ablation (Figure) | F1 | Improved |
| add multi-turn interaction (M.I.) with the graph (ablation) | Graph-R1 · Ablation (Figure) | F1 | Improved |
| add end-to-end reinforcement learning (R.L.) (ablation) | Graph-R1 · Ablation (Figure) | F1 | Improved |
| add RL training with hypergraph retrieval to graph-based RAG | HyperGraphRAG · Average over 6 QA datasets, 7B | F1: 29.40 -> 57.82 | Improved |
| use hypergraph knowledge with RL instead of chunk-based Search-R1 | Search-R1 · Average over 6 QA datasets, 7B | F1: 46.19 -> 57.82 | Improved |
| scale base model from 1.5B to 3B to 7B | Graph-R1 · Average over 6 QA datasets | F1: 40.09 -> 51.26 -> 57.82 | Improved |
| use graph structure with prompt-only (no RL) instead of StandardRAG | StandardRAG · Average over 6 QA datasets, GPT-4o-mini | F1: 32.05 -> 24.87 | Hurt |
| train Graph-R1 on Qwen3-4B (already RL-trained) instead of Qwen2.5 | Graph-R1 · Base model comparison | F1 | Hurt |
| use GRPO instead of REINFORCE++ or PPO (ablation) | Graph-R1 · RL algorithm comparison | F1 | Improved |
| use hypergraph knowledge instead of binary-relation graph or chunks, with RL | Graph-R1 · Knowledge representation comparison | F1 | Improved |
| simplified hypergraph construction versus GraphRAG | GraphRAG · Construction cost | seconds per 1K tokens: 8.04 -> 5.69 | Improved |
| use Graph-R1 localized retrieval instead of HyperGraphRAG | HyperGraphRAG · Query time | seconds per query: 9.6 -> 7.0 | Improved |
Later papers that compared with this method
2 later accepted papers report Graph-R1 as a baseline in their results tables; 2 of them report a better number on most of the results they share with it.
- GRASP: Graph Reasoning via Agentic Solving and Probing of LLMs (ICML 2026 · better on 1 of 1 shared result)
- PatternBloom: Empowering Agentic RAG with Externalized RL-Distilled Graph Patterns (NeurIPS 2026 · better on 1 of 1 shared result)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Process vs. Outcome Reward: Which is Better for Agentic RAG Reinforcement Learning (NeurIPS 2025 · similarity 0.88)
- MAGE: Multi-Agent Self-Evolution with Co-Evolutionary Knowledge Graphs (NeurIPS 2026 · similarity 0.88)
- Memory is Reconstructed, Not Retrieved: Graph Memory for LLM Agents (ICML 2026 · similarity 0.87)
- C-3PO: Compact Plug-and-Play Proxy Optimization to Achieve Human-like Retrieval-Augmented Generation (ICML 2025 · similarity 0.87)
- MC-Search: Evaluating and Enhancing Multimodal Agentic Search with Structured Long Reasoning Chains (ICLR 2026 · similarity 0.87)
- Improving Retrieval-Augmented Generation through Multi-Agent Reinforcement Learning (NeurIPS 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
