ToolGrad: Efficient tool-use dataset generation with textual "gradients"

Google Research is trying to fix a very boring but very important problem: how to make AI agents good at using tools without paying humans to annotate endless examples. The new method, ToolGrad, flips the usual process. Instead of starting with a user question and hunting for a tool path, it starts with a verified tool-use chain and then writes the matching prompt around it.

That matters because the old approach is expensive and clumsy. Prior methods such as ToolBench and ToolACE used an agent to search for a useful path through trial and error, often with depth-first search. ToolGrad takes the cleaner route. One LLM step can turn a known workflow into a usable training example, which makes the whole thing easier to scale and better suited to long, multi-step tasks.

The system itself is built around four parts: an API Proposer, API Executors, an API Selector, and an LLM Updater. The selector is where the “textual gradients” idea comes in. It reads the execution reports, picks the best call, and feeds that directional feedback back into the workflow. Google says that repeating this loop yields a user query, a verified API workflow, and a final AI response.

For the data source, the team used ToolBench, which contains more than 16,000 real-world APIs. It compared ToolGrad with the original query-first, DFS-based setup and found the answer-first method could generate more complex tool-use data, with a higher pass rate and lower cost. On top of that, the generated samples weren’t just cheaper; they were useful enough to train models that held up well outside the original tool set.

Google fine-tuned Gemma-3 models at 1B, 4B, and 12B parameters on a small dataset called ToolGrad-500, then tested them on the Berkeley Function Calling Leaderboard. The 12B model scored 83.1, just behind gemini-2.5-pro at 83.2 and ahead of claude-4.5 Opus at 82.8 and gpt-5 at 74.4. Google also says a model trained on data generated by gemini-2.5-flash-lite beat that same teacher model, which is the kind of result that makes the whole “teacher-student” story a little less neat.