DeepResearcher is the first comprehensive framework for end-to-end training of LLM-based deep research agents through scaling reinforcement learning (RL) in real-world environments with authentic web search interactions. Our qualitative analysis reveals emergent cognitive behaviors from end-to-end RL training, including the ability to formulate plans, cross-validate information from multiple sources, engage in self-reflection to redirect research, and maintain honesty when unable to find definitive answers.
Model Details
License: Apache 2.0
Model type: Reinforcement learning-based LLM (Large Language Model).
Language(s): The model is designed for tasks in English.
Finetuned from model: The model is built using the Qwen2.5-7B-Instruct architecture .
To get started, you can visit the DeepResearcher repository on GitHub, where the model's code and setup instructions are provided .
Training Details
Training Data
The model was trained on open-domain question-answering datasets, including:
NaturalQuestions (NQ)
TriviaQA (TQ)
HotpotQA
2Wiki MultiHopQA
Training Procedure
DeepResearcher was trained using reinforcement learning (RL) with the Group Relative Policy Optimization (GRPO) algorithm. It was tested in both in-domain (NQ, TQ, HotpotQA) and out-of-domain (Musique, Bamboogle, PopQA) settings .
Evaluation
Testing Data
The model was evaluated on several datasets, including:
NQ (Natural Questions)
TQ (TriviaQA)
HotpotQA
2Wiki
Musique
Bamboogle
PopQA .
Results
DeepResearcher outperforms all baseline models, achieving a substantial improvement in task completion across the datasets, particularly in out-of-domain scenarios.
Citation
@misc{zheng2025deepresearcherscalingdeepresearch,
title={DeepResearcher: Scaling Deep Research via Reinforcement Learning in Real-world Environments},
author={Yuxiang Zheng and Dayuan Fu and Xiangkun Hu and Xiaojie Cai and Lyumanshan Ye and Pengrui Lu and Pengfei Liu},
year={2025},
eprint={2504.03160},
archivePrefix={arXiv},
primaryClass={cs.AI},
url={https://arxiv.org/abs/2504.03160},
}
