lixiaoxi45/DeepAgent-QwQ-32B

DeepAgent: A General Reasoning Agent with Scalable Toolsets

DeepAgent is an end-to-end deep reasoning agent that performs autonomous thinking, tool discovery, and action execution within a single, coherent reasoning process. It is designed to overcome the limitations of traditional, predefined workflows by maintaining a global perspective on tasks and dynamically discovering tools.

• Paper: DeepAgent: A General Reasoning Agent with Scalable Toolsets
• Repository: GitHub - RUC-NLPIR/DeepAgent

Key Features

• Unified Agentic Reasoning: DeepAgent operates in a single stream of thought, autonomously reasoning about the task and discoverying necessary tools.
• Autonomous Memory Folding: To handle long-horizon interactions, DeepAgent introduces a mechanism that compresses past interactions into structured episodic, working, and tool memories, reducing context explosion while preserving critical information.
• ToolPO Strategy: An end-to-end reinforcement learning strategy tailored for general tool use, utilizing LLM-simulated APIs and fine-grained credit assignment for tool invocation.

Performance

Extensive experiments on eight benchmarks, including general tool-use tasks (ToolBench, API-Bank, TMDB, Spotify, ToolHop) and downstream applications (ALFWorld, WebShop, GAIA, HLE), demonstrate that DeepAgent consistently outperforms baselines across both labeled-tool and open-set tool retrieval scenarios.

Citation

If you find this work helpful, please cite the paper:

@misc{deepagent,
      title={DeepAgent: A General Reasoning Agent with Scalable Toolsets}, 
      author={Xiaoxi Li and Wenxiang Jiao and Jiarui Jin and Guanting Dong and Jiajie Jin and Yinuo Wang and Hao Wang and Yutao Zhu and Ji-Rong Wen and Yuan Lu and Zhicheng Dou},
      year={2025},
      eprint={2510.21618},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2510.21618}, 
}