Researchers improved AI tool usage by generating answers before queries
The new ToolGrad framework increases the success rate of AI agents by creating task solutions first and then deriving user prompts.
Researchers introduced ToolGrad, a system that reverses how AI agents learn to use digital tools like search engines and file managers. Instead of starting with a user request and searching for a solution, this framework generates a successful chain of tool actions first. It then uses that chain to create the corresponding user instruction. This approach achieved nearly a 100 percent success rate in generating training data.
The team applied this method to fine-tune Gemma-3 models. In tests on the Berkeley Function Calling Leaderboard, a 12 billion parameter version of the model reached a score of 83.1. This performance places it near advanced proprietary models like gemini-2.5-pro and claude-4.5 Opus.
ToolGrad demonstrates that high-quality tool-use datasets can be generated more efficiently and reliably through an answer-first paradigm.
ToolGrad relies on a four-stage process that proposes APIs, executes them, selects the best options through textual feedback, and updates the final response. The researchers noted that this method can help smaller, open-source models outperform the larger systems used to train them.
The study currently focuses on existing tool libraries. The researchers have not yet scaled the framework to dynamic, real-world API ecosystems that change continuously, nor have they tested its ability to perform long-term personalized learning for individual users.
Source: Google Research Blog