ToolGrad: Efficient tool-use dataset generation with textual "gradients"

Google Research introduces ToolGrad, a framework for generating tool-use datasets by reversing the standard workflow. It creates tool-use chains first, then annotates corresponding user prompts. This answer-first approach reduces cost and complexity compared to prior search-based methods. The paper, presented at ACL 2026, claims improved performance on out-of-distribution tasks. No code or weights are released.