What is ToolGrad?

ToolGrad is a method from Google Research for making the practice examples that teach an AI model to call tools. Tool calling is the model writing a request to an outside tool, such as a search or a map, which the app (the harness, the software around the model) then runs. ToolGrad builds a working chain of tool calls first and writes the user's question afterwards, to fit the chain. Take a chain of two calls: find_pharmacy with open now as its one field, then get_directions. Only then is the question written: which pharmacy near me is open now, and how do I walk there? The older method wrote the question first and then searched for calls to answer it. Google counted 63.8% of its examples as usable, against 99.8% of ToolGrad's. Every question ToolGrad writes already has a chain of calls that answers it, because the chain was built first.

ToolGrad is a method from Google Research for making the practice examples that teach an AI model to call tools, and it builds each example backwards: the chain of tool calls first, the user's question second. Tool calling is the model writing a request to an outside tool, such as a search or a calendar, which the app (the harness, the software around the model) runs before handing the result back. A model learns to do that from practice examples during training. ToolGrad runs before training, to make those examples. When you use the app (inference), only the trained model runs. Each example is a question, the right tool calls in order, and the final answer. Writing thousands of those by hand is too slow, so labs have an AI generate them, which is what synthetic data means. The older way started with the question. An AI agent then searched by trial and error for tool calls that answered it, and when no chain worked the example was thrown away. Google's count: 63.8% came out usable, so 36.2 in 100 were wasted. ToolGrad starts with the chain. It proposes a few tools that might help, tries them out, and has an AI judge write a short note on which call worked best. That note is the textual gradient. A gradient is the signal in training that says which way to adjust, and this one does the same job in words. The best call joins the chain and the loop runs once more. When the chain is complete, the AI writes the question and the answer to match it. Of those examples, 99.8% came out usable, and the chains were longer, 3.4 calls against 2.1. Google made 500 examples this way and trained its open model Gemma 3 12B (one anyone can download) on them. On BFCL, a test of tool calling, the model scored 83.1, which is 6.3 points above its score before that training and level with Gemini 2.5 Pro's 83.2 at the time. Two limits. The comparison with those models rests on one test and 500 examples. And a question written to fit an answer may not read as something a person would ask.

ToolGrad is a method for making the practice examples that teach an AI model to call tools, and it builds the chain of tool calls first and writes the question afterwards. Setting a crossword works the same way. The setter fills the grid with words first and writes the clues last. Each clue has an answer, because the answer was in the grid before the clue existed. A setter who wrote 30 clues first and then tried to fit answers into the grid would be left with clues that no word of the right length can satisfy. The filled grid is the chain of tool calls built first, and the clues are the questions written afterwards. The setter with unfillable clues is the older method, which wrote the question first and threw away the examples it could not answer. The match ends here: a setter fills the grid in one sitting, while ToolGrad builds its chain one call at a time, trying tools and holding on to the one that worked.