Google’s ToolGrad generates tool-use data with 99.8% pass rate
single source · 1 articles · research · confidence: high · first seen 2026-09-11 06:01 UTC
Google Research released ToolGrad, a framework for generating tool-use training data, at ACL 2026 Findings. It reverses the usual order: first it builds a verified chain of API calls, then writes the matching user query, guided by textual feedback from a four-module propose-execute-select-update loop. On ToolBench, a benchmark for tool-use data generation, it passes 99.8% of generated examples, against 63.8% for DFS search. A Gemma-3-12B fine-tuned on 500 ToolGrad samples scores 83.1 on BFCL, one point below Gemini 2.5 Pro. Code, dataset, and models are public under Apache-2.0.
What this means for you
If you train models to call APIs, ToolGrad is public under Apache-2.0 and a 500-sample fine-tune of Gemma-3-12B came within one point of Gemini 2.5 Pro on BFCL. Worth evaluating for cheaper tool-use training data.
Key facts
- ·ToolGrad is an ACL 2026 Findings framework from Google Research. source
- ·It generates tool-use data by building a verified API chain first, then writing the matching user query. source
- ·ToolGrad reaches a 99.8% pass rate on ToolBench, versus 63.8% for DFS search. source
- ·Gemma-3-12B fine-tuned on 500 ToolGrad samples scores 83.1 on BFCL, next to Gemini 2.5 Pro at 83.2. source
- ·Code, dataset, and models are released under Apache-2.0. source
What the sources say
- MarkTechPost — Reports the framework’s method, pass rates, and the Gemma fine-tuning benchmark.
Sources
The original reporting. Follow these — they did the work.