ToolGrad 是一种数据生成框架,它逆转了传统范式,先生成工具使用答案,再生成用户查询。我们证明该设计可提升大语言模型(LLM)的工具使用性能。与先前通过搜索低通过率用户查询的解来构建工具使用数据集的方法不同,ToolGrad 先生成成功的工具使用链,再据此生成提示,从而获得高通过率。
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. While prior art generates tool-use datasets by searching solutions of user queries with low pass rate, ToolGrad generates successful tool-use chains before generating prompts, yielding high pass rate.
ToolGrad 是一种数据生成框架,它颠覆了传统范式,率先生成工具调用答案,再生成用户查询。我们证明该设计可提升大语言模型(LLM)的工具使用性能。此前的方法通过搜索用户查询的解决方案来构建工具使用数据集,但因通过率较低而受限;而 ToolGrad 则先生成成功的工具调用链,再生成提示(prompt),从而实现高通过率。ToolGrad 包含四个核心模块,按顺序执行“提出(propose)、执行(execute)、选择(select)、更新(update)”操作。重复这一迭代过程,最终生成一个数据样本,其中包含用户查询、经验证的 API 工作流以及最终的 AI 响应。ToolGrad 在生成提示之前即生成成功的工具调用链,因而获得高通过率。图中对比了“查询优先”方法(基线)与“答案优先”方法(本工作)在生成效率上的差异。
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. While prior art generates tool-use datasets by searching solutions of user queries with low pass rate, ToolGrad generates successful tool-use chains before generating prompts, yielding high pass rate. ToolGrad features four core modules that sequentially propose, execute, select, and update. Repeating this iterative process results in a data sample consisting of a user query, a verified API workflow, and the final AI response. ToolGrad generates successful tool-use chains before generating prompts, yielding a high pass rate. Generation efficiency comparison between the query-first approach (baseline) and the answer-first approach (ours).