Google Research builds valid tool sequences before writing related queries, to produce training data more efficiently.
Presented on September 10, ToolGrad proposes to start with the solution: the system assembles and checks tool calls, then generates the corresponding user request. Text returns guide the addition of new steps and the selection of calls that work.
Google reports that this method produces more complex routes at lower cost than the reference approach studied. Fitted Gemma 3 models with 500 examples improve their results on a benchmark using other tools. These measures focus on accurate assessments and do not demonstrate reliable autonomy in all environments.
Read the full article on the original source
Go to the source →