Google's ToolGrad hits 83.1 on tool use

Google Research reports ToolGrad-12B scored 83.1 on the Berkeley Function Calling Leaderboard after Gemma-3 fine-tuning on ToolGrad-500, generated from ToolBench’s 16,000+ APIs, according to the research release. The answer-first pipeline reports almost 100% pass rate in data generation and lower generation cost than depth-first search; teams building tool-using agents should test this data-centric recipe on unseen tools, not read one benchmark as production reliability, alongside the archive’s harness-cost analysis.