AutoTIR: Autonomous Tools Integrated Reasoning via Reinforcement Learning

Inspired by the human ability to adaptively select tools, we introduce AutoTIR, a reinforcement learning framework that enables LLMs to autonomously decide whether and which tool to invoke during the reasoning process, rather than following static tool-use strategies. AutoTIR leverages a hybrid reward mechanism that jointly optimizes for task-specific answer correctness, structured output adherence, and penalization of incorrect tool usage, thereby encouraging both precise reasoning and efficient tool integration

📰 News

• [2025-08-12] 🤗 We release our trained models on Hugging Face, please check it out!
• [2025-07-28] 🎉 We release the paper, update the code and open-source the models.
  • 📝 The paper is released on arXiv, more details and evaluation results can be found in our paper.
  • 🛠️ The repository is updated with the new implementation, especially the rollout with multiple action strategies during RL training. This version of implementation is based on the latest release of verl.

🤝 Acknowledge

This training implementation is based on verl and the evaluation is based on FlashRAG. The serving of retriever is based on FastAPI. The model serving is based on SGLang. AutoTIR,models are trained based on Qwen2.5 and are modifications of the code from ReSearch. We sincerely appreciate their contributions to the open-source community.

📚 Citation

If you find this work useful, please cite it as follows:

@article{wei2025autotir,
  title={AutoTIR: Autonomous Tools Integrated Reasoning via Reinforcement Learning},
  author={Wei, Yifan and Yu, Xiaoyan and Weng, Yixuan and Pan, Tengfei and Li, Angsheng and Du, Li},
  journal={arXiv preprint arXiv:2507.21836},
  year={2025}
}