# Google Research — ToolGrad: Efficient tool-use dataset generation with textual "gradients"

- Company: Google Research (research.google)
- Announced: 2026-09-10T22:50:22+00:00
- Category: research-paper
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://research.google/blog/toolgrad-efficient-tool-use-dataset-generation-with-textual-gradients/
- Record: https://forck.live/items/9974-toolgrad-efficient-tool-use-dataset-generation-with-textual-gradients
- Subject: Research
- Models affected: Gemma-3, ToolGrad-1B, ToolGrad-4B, ToolGrad-12B

Google Research introduces ToolGrad, a data generation framework that reverses the traditional paradigm by first generating tool-use answers before user queries, enabling LLMs to achieve better tool-use performance. The framework was presented at ACL 2026. Experiments show that fine-tuning Gemma-3 models on ToolGrad-500 data improves tool-use performance across all tested parameter sizes, with ToolGrad-12B achieving a score of 83.1 on the Berkeley Function Calling Leaderboard.

## Evidence

Verbatim from https://research.google/blog/toolgrad-efficient-tool-use-dataset-generation-with-textual-gradients/:

> ToolGrad is a data generation framework that reverses the traditional paradigm by first generating tool-use answers before user queries. We show this design enables LLMs to achieve better tool-use performance.

---

Record: https://forck.live/items/9974-toolgrad-efficient-tool-use-dataset-generation-with-textual-gradients
Catalogue: https://forck.live/llms.txt
Current issue: https://forck.live/feed.md