ToolGrad: Efficient Tool-use Dataset Generation with Textual “Gradients”
Paper · 2026 · Zhongyi Zhou, Kohei Uehara, Haoyu Zhang et al.
Summary
Training LLMs to call tools requires pairs of user requests and correct tool-call sequences. Earlier synthetic pipelines such as ToolBench start from an invented query and then search for a solution with DFS. This search is costly and often fails, and it can leave unsolvable queries or flawed traces in the data. ToolGrad flips the order. It first builds a verified tool-use workflow and only afterwards writes a matching user query and response in one LLM step.
The workflow is grown iteratively over mini-batches of APIs drawn from a large database. Four agent modules propose, execute and select APIs, then update the workflow. LLM feedback acts as textual "gradients", an idea borrowed from TextGrad. Using Gemini-2.5-flash-lite as the teacher, the authors built ToolGrad-500. It contains more complex tool use than ToolBench and was produced more cheaply with a near-perfect pass rate.
Gemma-3 models fine-tuned on it beat ToolBench-trained baselines and several proprietary models, including on unseen tools. This suggests answer-first generation lets students surpass their teacher. Code, data and models are released.
Key points
- Dataset and pass rate: ToolGrad-500 holds about 500 triplets of query, API workflow and response. Every sample starts from a valid tool call, and failed steps never enter the data, which gives an almost 100% pass rate.
- BFCL results: ToolGrad-12B places second on BFCL, only 0.1 behind Gemini 2.5 Pro. It outperforms ToolACE and its own teacher, Gemini-2.5-flash-lite. The largest gains appear on the "Multiple Parallel" subtask.
- Student beats teacher: even the 1B fine-tuned model surpasses its teacher model on the paper's benchmarks.
- Iteration scaling: pass rate levels off between 8 and 12 iterations, so 10 iterations were used to build ToolGrad-500.
- Sample scaling: Gemma-3-4B was trained on 100 to 2k samples. Accuracy rises and then falls as data grows, because generated tool uses become repetitive. The authors attribute this to the lack of a shared memory module.
Institutions
Japan96 works · 9,600 citations🎓Tohoku UniversitySendai · Japan23 works · 595 citations🔬RIKEN Center for Advanced Intelligence ProjectTokyo · Japan16 works · 957 citations🔬RIKENWako · Japan30 works · 607 citationsFields, topics and concepts
📚 Shelves 5The field of science this work belongs to (OpenAlex fields).Computer Science · 18,614 worksArtificial Intelligence · 17,965 worksArts and Humanities · 17,450 worksLanguage and Linguistics · 17,296 worksInformation Systems · 1,730 works
Cite
Cite as Zhongyi Zhou, Kohei Uehara, Haoyu Zhang, Jingtao Zhou, Lin Gu, Ruofei Du, Zheng Xu and Tatsuya Harada. 2026. ToolGrad: Efficient Tool-use Dataset Generation with Textual “Gradients”. Findings of the Association for Computational Linguistics: ACL 2026.
@inproceedings{zhou2026toolgrad,
title = {{ToolGrad: Efficient Tool-use Dataset Generation with Textual “Gradients”}},
author = {Zhongyi Zhou, Kohei Uehara, Haoyu Zhang, Jingtao Zhou, Lin Gu, Ruofei Du, Zheng Xu and Tatsuya Harada},
booktitle = {Findings of the Association for Computational Linguistics: ACL 2026},
year = {2026},
doi = {10.18653/v1/2026.findings-acl.950},
url = {https://doi.org/10.18653/v1/2026.findings-acl.950}
}References
- Xiaohe Bo, Zeyu Zhang, Quanyu Dai, Xueyang Feng, Lei Wang, Rui Li, Xu Chen, and Ji-Rong Wen. 2024. Reflective multi-agent collaboration based on large language models. Advances in Neural Information Processing Systems, 37:138595–138631. \#\# Limitations find
Updated
This page in Markdown → https://md.research.elamur.ai/free/en/papers/zhou-2026-toolgrad-efficient-tool-use-dataset/