Google DeepMind researchers have released ToolGrad, a new framework that dramatically improves how large language models (LLMs) learn to call external tools and APIs. The breakthrough inverts the traditional training pipeline: instead of asking an AI to invent a user query and then search for the right tool chain, ToolGrad builds a verified tool chain first, then writes a matching query. The result is a 99.8% success rate in generating training data, compared to just 63.8% with previous approaches.

Why Does This Matter for AI Development?

Training an LLM to reliably call tools requires massive datasets that pair user questions with correct sequences of API calls. Creating that data at scale has been slow, expensive, and error-prone. The old method, called query-first, would ask an AI to imagine a user instruction, then dispatch a search agent to find a tool chain that satisfies it. When the search failed, all the computing power spent on exploration was wasted. ToolGrad solves this by working backward: it executes real APIs to build a ground-truth tool chain, then annotates it with a matching user query. This approach is far more efficient because the chain already works.

How Does ToolGrad Actually Work?

• API Proposer: Narrows a sampled set of APIs down to a few candidates that could extend the current workflow.
• API Executors: Run those candidates in parallel and produce detailed execution reports showing what each API actually returns.
• API Selector: Reviews the reports, picks the single best-performing call, and appends it to the workflow, providing directional feedback.
• LLM Updater: Rewrites the synthetic user query and AI response so they match the new API set, keeping everything aligned.

The system repeats this loop to build longer, more complex tool-use chains. In the default configuration, it runs 10 iterations over 50 sampled APIs per workflow, generating one complete training sample each time.

What Do the Numbers Show?

The research team tested ToolGrad against the older DFS-based query-first approach using the ToolBench API database, which contains over 16,000 real-world APIs. The improvements were dramatic. Pass rates jumped from 63.8% to 99.8%, meaning ToolGrad successfully generated valid training samples in nearly every attempt. The framework also produced longer tool-use chains, increasing from 2.1 steps per sample to 3.4 steps, while actually reducing the total number of tool-use steps needed from 34.3 to 20.0.

The researchers then generated a 500-sample dataset called ToolGrad-500 using Gemini 2.5 Flash-Lite and used it to fine-tune Gemma-3 models at three different sizes: 1 billion, 4 billion, and 12 billion parameters. When tested on the Berkeley Function Calling Leaderboard, which uses a completely different set of tools to ensure a fair out-of-distribution test, the results were striking. Gemma-3-12B scored 83.1, matching Gemini 2.5 Pro at 83.2 and outperforming Claude 4.5 Opus at 82.8 and GPT-5 at 74.4.

What Makes This Result Surprising?

The most remarkable finding is that a smaller, open-source model trained on just 500 samples can match the performance of Google's own frontier model. The 12-billion-parameter Gemma-3 student model even outperformed Gemini 2.5 Flash-Lite, the teacher model that generated its training data. This suggests that the quality of training data matters far more than raw model size or the resources of the organization building it. The framework also led other open-source tool-use specialists, including ToolACE and Hammer-2.1-7B.

Is This Actually Usable?

Yes. Google has released ToolGrad under an Apache 2.0 open-source license, making it freely available for research and commercial use. The ToolGrad-500 dataset and the fine-tuned Gemma-3 models at 1B, 4B, and 12B parameters are available on Hugging Face, a popular platform for sharing AI models. There is also a PyPI package for easy installation. The reproduction scripts were verified on a single NVIDIA A100 40GB GPU, making it accessible to researchers and developers without massive computing budgets.

This release represents a significant shift in how the AI community approaches tool-use training. By open-sourcing both the framework and the trained models, Google is enabling smaller organizations and independent researchers to build AI systems that can reliably interact with external tools and APIs. The dramatic improvement in data generation efficiency, combined with the strong benchmark performance, suggests that ToolGrad could become a standard approach for training tool-use capabilities in future LLMs.