ToolGrad: Google's Revolutionary Tool-Use Data Generation System

Google Research has unveiled a novel approach to generating tool-use data called ToolGrad. The system flips the traditional pipeline by first constructing a verified tool-use chain and then writing a user query to match it.

This innovative method has shown impressive results, with a pass rate of 99.8% on the ToolBench API database compared to 63.8% for the traditional DFS-based approach.

The researchers also fine-tuned the Gemma-3 model on the ToolGrad-500 dataset and achieved remarkable scores on the Berkeley Function Calling Leaderboard, surpassing even its teacher model Gemini 2.5 Flash-Lite.