Google Research ToolGrad Cuts AI Tool-Use Data Costs With 99.8% Pass Rate
Google Research has introduced ToolGrad, a new framework designed to make it faster and cheaper to generate training data for AI models that use external tools. Instead of starting with a user request and asking an AI system to discover a solution, ToolGrad first builds a valid sequence of tool calls and then generates the user query that matches it.
The researchers say this change in direction produced a 99.8% pass rate in their data-generation experiments. The approach also generated more complex tool-use examples at lower cost than the query-first method used in earlier systems.
ToolGrad Reverses the Usual Training Process
Training an AI model to work with external tools requires large datasets showing how a user request should translate into a series of API calls and a final answer. Traditional approaches generally create a hypothetical user query first and then use an agent to search for a successful tool-use path.
That process can waste significant computing resources because the agent may try multiple unsuccessful paths before finding one that works. ToolGrad reverses the sequence by constructing a verified tool-use chain first and generating the corresponding user request afterward.
Four-Part System Builds Verified Workflows
ToolGrad uses four main components: an API proposer, API executors, an API selector and an LLM updater. The system first narrows a pool of available APIs, tests promising combinations and selects the strongest-performing call based on execution results.
The selected API is then added to the workflow, while the LLM updater adjusts the synthetic user query and AI response to match the expanded tool chain. The process can be repeated to create increasingly complex workflows.
99.8% Pass Rate in ToolBench Testing
Google Research tested ToolGrad using ToolBench, a database containing more than 16,000 real-world APIs. The researchers compared its answer-first approach with the conventional query-first method and found that ToolGrad generated tool-use data with a reported 99.8% pass rate while requiring fewer optimisation steps.
The researchers also created a smaller dataset called ToolGrad-500 and used it to fine-tune Gemma 3 models at 1-billion, 4-billion and 12-billion parameter sizes. The resulting models showed improved tool-use performance across the tested configurations.
Smaller AI Models Show Strong Results
The 12-billion-parameter model trained on ToolGrad-500 achieved a score of 83.1 on the Berkeley Function Calling Leaderboard. Google Research says that was close to the scores of several proprietary models tested at the time, including Gemini 2.5 Pro at 83.2 and Claude 4.5 Opus at 82.8.
The researchers also reported an unusual result: the ToolGrad-12B model outperformed the Gemini 2.5 Flash-Lite model that generated the training data. That finding suggests carefully generated tool-use examples can allow smaller models to learn effective behaviours from a more capable system.
ToolGrad Targets AI Agent Training
The research is aimed at a growing problem in AI development. Agents increasingly need to search the web, access files, call APIs, execute code and interact with software, but producing reliable training examples for those tasks can be expensive.
Google Research says ToolGrad could make that process more scalable by generating verified tool-use chains before creating the prompts that describe them. The team plans to explore the framework across larger and more dynamic API ecosystems, as well as applications involving continuous learning and personalised agent behaviour.
The work was presented at ACL 2026, where the researchers positioned ToolGrad as an alternative to the trial-and-error approach used in earlier tool-use dataset generation systems. Its reported results suggest that changing how training examples are constructed could become as important as increasing model size when building more capable AI agents.
