Foreverskyou/ToolBench-X

Beyond Function Calling: Benchmarking Tool-Using Agents under Tool-Environment Unreliability

📄 Arxiv Report | 📦 Github

ToolBench-X is the first benchmark for evaluating tool-using agents under recoverable tool-environment unreliability, which features executable multi-step tool-use tasks, structured reliability hazards, and automatic evaluation of hazard diagnosis and recovery capabilities.

• ToolBench-X contains 1,106 tasks, including 378 sequential, 358 parallel, and 370 mixture tasks.
• It includes 7 main topics, 84 unique subtopics, and five categories of reliability hazards: Specification Drift, Invocation Error, Execution Failure, Output Drift, and Cross-source Conflict.

License

ToolBench-X is only used for academic research. Commercial use in any form is prohibited.
Without prior approval, you cannot distribute, publish, copy, disseminate, or modify ToolBench-X in whole or in part. 
You must strictly comply with the above restrictions.

Citation

If you find our work helpful for your research, please consider citing our work.

@article{tian2026beyond,
  title={Beyond Function Calling: Benchmarking Tool-Using Agents under Tool-Environment Unreliability},
  author={Tian, Yang and Shi, Zhengpeng and Zhao, Bo},
  journal={arXiv preprint arXiv:2606.25819},
  year={2026}
}