Google Research Releases ToolGrad: Answer-First Framework Hits 99.8% Pass Rate for Tool-Use Data Generation

Google Research and collaborators introduced ToolGrad, a novel method that generates verified tool-use chains before queries, achieving a 99.8% pass rate on tool-use data generation. Their Gemma-3 models fine-tuned on this data perform comparably to leading proprietary models on the Berkeley Function Calling Leaderboard. ToolGrad's code, dataset, and models are openly available on Hugging Face and PyPI.