Google 提出 ToolGrad:用文本梯度高效生成工具使用数据集
原标题:ToolGrad: Efficient tool-use dataset generation with textual "gradients"
AI 摘要
Google Research 在 ACL 2026 发表 ToolGrad,一种「答案优先」的工具使用数据集生成框架:先生成工具调用链,再反向标注用户查询,并借鉴 TextGrad 的「文本梯度」迭代构建复杂 API 工作流。基于 ToolBench 的 1.6 万+ API 生成 ToolGrad-500 数据集,微调 Gemma-3(1B/4B/12B)后在 BFCL 基准上超越基线,甚至匹配 SoTA 专有模型,且生成通过率接近 100%、成本更低。
正文节选
September 10, 2026 Zhongyi Zhou, Research Scientist, and Ruofei Du, Interactive Perception & Graphics Lead, Google XR ToolGrad is a data generation framework that reverses the traditional paradigm by first generating tool-use answers before user queries. We show this design enables LLMs to achieve better tool-use performance. AI agents have shown great potential in automating real-world tasks, such as conducting a Google Search, reading local computer files, or executing generated Python scripts