Google Research released the ToolGrad framework for generating tool-use training datasets for LLMs
Research publication Provisional 78% confidence first seen
Google Research introduced ToolGrad, an answer-first approach to create tool-use training data by executing a verified tool chain before annotating the corresponding user prompt. The coverage reports strong improvements in benchmark pass rates for tool-use dataset generation and highlights results such as ToolGrad-12B scoring 83.1 on the Berkeley Function Calling Leaderboard and higher pass rates on ToolBench, with claimed reductions in data generation cost.