ToolGrad reaches 99.8% pass rate for tool-use data generation
Researchers from Google, the University of Tokyo, RIKEN AIP and Tohoku University have introduced ToolGrad, a framework that reverses the usual process for creating tool-use training data. Instead of generating a user request first and searching for a suitable API sequence, it executes APIs to build a verified chain, then generates a matching query and response.
ToolGrad uses four modules in a repeated loop: an API Proposer identifies possible next calls, Executors run them, a Selector chooses the strongest result, and an LLM Updater revises the synthetic query and answer. The default setup performs 10 iterations over 50 sampled APIs. On the ToolBench database of more than 16,000 APIs, the reported pass rate increased from 63.8% with a depth-first-search method to 99.8%. Chains averaged 3.4 ground-truth tool uses instead of 2.1, while tool-use steps fell from 34.3 to 20.0 per sample.
The team also created the 500-example ToolGrad-500 dataset with Gemini 2.5 Flash-Lite and used it to fine-tune Gemma-3 models with 1B, 4B and 12B parameters. On the Berkeley Function Calling Leaderboard, ToolGrad-12B scored 83.1, close to Gemini 2.5 Pro’s 83.2 and above the reported scores for Claude 4.5 Opus and GPT-5. The code, dataset and models are released under the stated distribution arrangements.
Written by Sakhanda Wire AI from reporting by MarkTechPost. Facts, quotes and full context are in the original article below.