Researchers Boost On-Device Mobile AI with Extended Function-Calling Model

Mobile assistants that operate entirely on-device face a significant challenge: converting spoken or typed instructions into actual system actions without relying on cloud APIs. Researchers have now demonstrated a more capable solution by extending a compact language model with a new dataset tailored to real-world Android workflows.

What the researchers built#

The team extended FunctionGemma, a 270-million-parameter model, by creating MOBILEACTIONSEXTENDED—a dataset of roughly 9,500 synthetic conversations covering fifteen distinct device-control categories. These span messaging, phone calls, camera and screenshot capture, brightness adjustment, device status checks, flashlight toggling, and app management. They then fine-tuned the model using supervised learning with completion-only loss via TRL.

Performance gains across device categories#

The specialized model achieved a sharp improvement: end-to-end accuracy jumped to 76.5% on the new dataset, compared to just 29.3% for the base FunctionGemma and 17.2% for Google's existing Mobile-Actions variant. The researchers also trained a combined model jointly on both their dataset and Google's MOBILEACTIONSGOOGLE benchmark, which retained 76.5% accuracy on their tasks while reaching 82.3% on Google's benchmark—an 8-percentage-point trade-off for achieving nearly double the category coverage.

Why this matters for mobile AI#

The approach prioritizes two key benefits over cloud-based alternatives: low latency and privacy preservation. By keeping function calling local to the device, users avoid the delays and data-sharing concerns that come with server-based processing. The researchers released their dataset, fine-tuned models, training pipeline, and an Android demonstration, positioning on-device function calling as a practical route toward practical personal assistants.