ToolGrad Flips Tool-Use Data Generation Script, Achieves Near-Perfect Accuracy

A team of researchers from Google and several universities has developed ToolGrad, a new approach to generating tool-use data. Unlike previous methods that follow a query-first recipe, ToolGrad inverts this pipeline by first building a verified tool chain and then writing the user query.

This method significantly improves the efficiency and accuracy of data generation. In tests on the ToolBench API database, ToolGrad achieved a pass rate of 99.8%, compared to 63.8% for traditional methods.

The team also fine-tuned the Gemma-3 model on the ToolGrad dataset and evaluated it on the Berkeley Function Calling Leaderboard. The results showed that fine-tuning on ToolGrad-500 improved tool-use scores at every parameter size, with the 12B student outperforming its teacher model.