Google's ToolGrad research creates a successful tool-use workflow before generating the user request that would require it.
Earlier synthetic-data methods often invented a task first, then searched for a working chain of API calls. ToolGrad starts with a proposed API sequence and tests the calls. A selector reviews execution reports and chooses useful steps; a language model then adjusts the request and response to match the verified workflow. The researchers call the iterative feedback "textual gradients" because a language-model critique guides changes instead of a numerical training gradient.
Google reports that this answer-first process produces more complex tool-use examples at lower generation cost than the methods it compared. Models trained on the resulting data performed better in its tool-use evaluations, including tests with previously unseen tools. The work was presented at ACL 2026.