Google builds the data that teaches models to use tools backwards
ToolGrad runs the tools first to build a working chain, then writes the question that fits it. Generation success went from 63.8 to 99.8 percent. Apache 2.0.
A Google research team has published ToolGrad, a framework for producing the training data that teaches language models to use tools. The work was done with researchers from the University of Tokyo, RIKEN AIP and Tohoku University.
Tool use means the model calling an outside service instead of inventing an answer: looking up the weather, querying a database, running a calculation. Whether today's agents can do their job depends on it.
Why it is hard
To train a model on this you need pairs: a user's question, and the correct chain of tool calls that answers it. Producing those pairs at scale is both slow and expensive.
The usual method, ToolBench for example, starts from the question: it has a model invent a user query, then searches for a tool sequence that would answer it. The search often fails, and the compute spent on it is wasted.
ToolGrad inverts the order
ToolGrad first executes the tools for real and builds a chain that works. Once it holds a chain proven to run, it writes the user question that matches it.
The difference was measured: on the same API database the older approach succeeded 63.8 percent of the time, while ToolGrad reaches 99.8 percent. Almost every generated example is usable.
The team also released ToolGrad-500, a set of 500 samples produced with Gemini 2.5 Flash-Lite.
Why it matters
The code is on GitHub under Apache 2.0, the models are on Hugging Face, and there is a PyPI package. The method is not only described, it is handed over in usable form.
This is the bottleneck anyone building agents runs into: however good the model is, without data that teaches tool use properly the agent makes the wrong call in the field.
Source: MarkTechPost.
