R3-RAG: Learning Step-by-Step Reasoning and Retrieval for LLMs via Reinforcement Learning

📖 Overview

R3-RAG is a novel framework that uses Reinforcement learning to teach LLMs how to Reason and Retrieve step by step. Unlike traditional RAG methods that rely on human-designed workflows, R3-RAG enables models to autonomously learn optimal reasoning-retrieval strategies through reinforcement learning with both outcome and process rewards.

Key Features

• 🔥 Autonomous Learning: Uses RL to learn reasoning-retrieval strategies instead of relying on fixed human-designed workflows
• 🎯 Dual Reward System: Combines outcome rewards (answer correctness) with process rewards (document relevance)
• 🚀 Strong Performance: Achieves significant improvements over state-of-the-art iterative RAG methods
• 🔄 Transferable: Works across different retrievers (E5, BGE, BM25) with consistent performance
• 📊 Comprehensive: Evaluated on multiple multi-hop QA datasets (HotpotQA, 2WikiMultiHopQA, MuSiQue)

📊 Main Results

Our method significantly outperforms existing baselines across three multi-hop QA datasets:

🚀 Quick Start

Environment Setup

We recommend setting up three separate conda environments to avoid dependency conflicts:

1. FlashRAG Environment (for retrieval tools): Please refer to FlashRAG to set up the environment.

2. LLaMA-Factory Environment (for cold start training): Please refer to LLaMA-Factory to set up the environment.

3. OpenRLHF Environment (for RL training): Please refer to OpenRLHF to set up the environment, then install our modified openrlhf code in this repository.

Model Download

Download our pre-trained models from Hugging Face:

# Cold start models
git clone https://huggingface.co/Yuan-Li-FNLP/R3-RAG-CS-Llama
git clone https://huggingface.co/Yuan-Li-FNLP/R3-RAG-CS-Qwen

# Full R3-RAG models
git clone https://huggingface.co/Yuan-Li-FNLP/R3-RAG-Llama
git clone https://huggingface.co/Yuan-Li-FNLP/R3-RAG-Qwen

Quick Demo

Experience R3-RAG with our visualization interface:

# First, start the server
cd startup
bash server.sh

# Then, start the visualization interface
bash startup_visualize.sh

Make sure to configure the model paths and parameters in the startup scripts before running.

📁 Repository Structure

R3-RAG/
├── benchmark/ # Evaluation scripts and benchmarks
│ ├── evaluate.py # Main evaluation script
│ ├── metrics/ # Evaluation metrics implementation
│ └── datasets/ # Dataset loading and processing
├── data/ # Cold start data construction
│ ├── build_coldstart_data.py # Generate high-quality cold start trajectories
│ ├── data_processing/ # Data preprocessing utilities
│ └── templates/ # Prompt templates for data generation
├── startup/ # Demo and visualization scripts
│ ├── server.sh # Start the server
│ ├── startup_visualize.sh # Start visualization interface
│ └── demo_config.py # Configuration for demo
├── tool/ # Retrieval tools and services
│ ├── retrieval/ # Retrieval tool implementations
│ ├── vllm_service/ # VLLM service code
│ └── utils/ # Utility functions
├── train/ # Training frameworks
│ ├── llamafactory/ # SFT training with LLaMA-Factory
│ │ ├── sft_training.py # Cold start SFT training script
│ │ └── configs/ # Training configurations
│ └── openrlhf/ # RLHF training with OpenRLHF
│ ├── rl_training.py # Reinforcement learning training
│ ├── reward_models/ # Reward model implementations
│ └── configs/ # RL training configurations
├── README.md # This file
└── LICENSE # License file

🤗 Available Models and Data

Models

• R3-RAG-CS-Llama: Cold start model based on Llama-3.1-8B
• R3-RAG-CS-Qwen: Cold start model based on Qwen2.5-7B
• R3-RAG-Llama: Full R3-RAG model based on Llama-3.1-8B
• R3-RAG-Llama-ORM: R3-RAG model with outcome reward only
• R3-RAG-Qwen: Full R3-RAG model based on Qwen2.5-7B

Datasets

• R3-RAG-ColdStartTrainingData: High-quality cold start training trajectories
• R3-RAG-RLTrainingData: Reinforcement learning training data

All models and datasets are available on Hugging Face.

📄 Citation

If you find our work helpful, please consider citing:

@misc{li2025r3raglearningstepbystepreasoning,
      title={R3-RAG: Learning Step-by-Step Reasoning and Retrieval for LLMs via Reinforcement Learning}, 
      author={Yuan Li and Qi Luo and Xiaonan Li and Bufan Li and Qinyuan Cheng and Bo Wang and Yining Zheng and Yuxin Wang and Zhangyue Yin and Xipeng Qiu},
      year={2025},
      eprint={2505.23794},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2505.23794}, 
}

🤝 Contributing

We welcome contributions! Please feel free to submit issues and pull requests.

📝 License

This project is licensed under the GNU General Public License v3.0 - see the LICENSE file for details.

🙏 Acknowledgments

• FlashRAG for retrieval framework
• LLaMA-Factory for training framework
• OpenRLHF for reinforcement learning framework

📞 Contact

For questions or collaborations, please contact: - Yuan Li: liyuan24@m.fudan.edu.cn - Xipeng Qiu: xpqiu@fudan.edu.cn

Made with ❤️ by the Fudan NLP Group