README.md

---
language:
- en
tags:
- GraphRAG
- retrieval-augmented-generation
- reinforcement-learning
- multi-hop-reasoning
license: apache-2.0
datasets:
- 2Wiki
- MuSiQue
- HotpotQA
---

# GraphRAG-R1

**This is the official repository for the paper “[GraphRAG-R1: Graph Retrieval-Augmented Generation with Process-Constrained Reinforcement Learning](https://arxiv.org/abs/2507.23581)” (Accepted to WWW '26).**

| **Item** | **Details** |
|:---|:---|
| **📄 Paper** | [GraphRAG-R1: Graph Retrieval-Augmented Generation with Process-Constrained Reinforcement Learning](https://arxiv.org/abs/2507.23581) |
| **🏆 Conference** | **The ACM Web Conference 2026 (WWW '26)** <br> *April 13–17, 2026, Dubai, United Arab Emirates* |
| **💻 Full Code** | **GitHub:** (https://github.com/ycygit/GraphRAG-R1) |
| **🤗 HF Models** | This repository provides the LoRA adapters. See [Model Weights & Usage](#model-weights--usage). |

---

## 🔍 Quick Links
- [Abstract](#abstract)
- [Model Weights & Usage](#model-weights--usage)
- [Citation](#citation)
- [Contact](#contact)

---

## 📝 Abstract

GraphRAG-R1 is a Graph Retrieval-Augmented Generation framework enhanced with **Process-Constrained Reinforcement Learning**. It is designed to significantly improve the reasoning capabilities of large language models (LLMs) on complex, multi-hop question answering tasks by integrating structured knowledge graph retrieval with constrained reinforcement learning over the reasoning process.

**Official BibTeX Citation:**
```bibtex
@inproceedings{yu2025graphrag,
  title={GraphRAG-R1: Graph Retrieval-Augmented Generation with Process-Constrained Reinforcement Learning},
  author={Yu, Chuanyue and Zhao, Kuo and Li, Yuhan and Chang, Heng and Feng, Mingjian and Jiang, Xiangzhe and Sun, Yufei and Li, Jia and Zhang, Yuzhi and Li, Jianxin and others},
  booktitle = {Proceedings of the ACM Web Conference 2026 (WWW '26)},
  year={2026}
}
````

---

## 🤖 Model Weights & Usage

This repository provides **LoRA adapters** for the GraphRAG-R1 framework. To use them, you must load them onto the corresponding base model.

**Available Adapters:**

1. **For `Qwen2.5-7B`**: Adapter for the base model.
2. **For `Qwen2.5-7B-Instruct`**: Adapter for the instruction-tuned variant.

---

## 🙏 Contact

For questions regarding the model or paper, please open an issue in the future GitHub repository or contact the authors.

---