Researchers from Google and several Japanese universities have introduced ToolGrad, a framework that inverts the traditional pipeline for generating tool-use datasets by constructing verified API chains first and then writing matching user queries. This approach aims to eliminate the inefficiency of query-first methods that often fail during agent exploration.

  • The system uses four modules in a loop: API Proposer, API Executors, API Selector, and LLM Updater.
  • On the ToolBench database, ToolGrad achieved a 99.8% pass rate compared to 63.8% for DFS-based methods.
  • It reduced tool-use steps per sample from 34.3 to 20.0 while increasing ground-truth chain length.
  • Fine-tuning Gemma-3 models on the resulting ToolGrad-500 dataset improved performance across all parameter sizes.
  • The 12B model reached 83.1 on the Berkeley Function Calling Leaderboard, surpassing Gemini 2.5 Flash-Lite and competing with frontier proprietary models.

The framework provides open-source code, datasets, and models to enable more efficient and reliable training of LLMs for tool use.