Autofinetune automates model tuning experiments on TPUs

Google describes an experimentation loop in which an agent adjusts the settings, runs the training and retains the measured improvements.

On September 11, Google presented autofinetune, a project combining Tunix, Gemma and TPUs with an agent in charge of piloting the experiments. A human-written framework sets the modifiable parameters, constraints and evaluation metric.

The demonstration includes twenty supervised adjustment experiments on FunctionGemma and forty reinforcement experiments on Gemma 3 1B for mathematical exercises. The system retains the changes deemed favorable and cancels the regressions. These tests illustrate the automation of a search for settings within a defined perimeter, without establishing a general capacity for autonomous scientific research.