Researchers Build On-Device AI That Keeps Learning From User Behavior on Android

Developers have created Online-SDFT, an open-source prototype that enables a 230-million-parameter language model to continuously learn from user interactions directly on an Android device. The system uses a rank-4 LoRA adapter and ONNX Runtime Training to fine-tune the model locally, without sending data to external servers. Rather than relying on clean labels or explicit rewards, it uses a 'hindsight' approach where the same base model — with its adapter disabled — reviews completed interactions to generate soft training targets for future decisions. The prototype was tested on a notification-routing task, where the model learns to better decide whether to show, save, or archive alerts based on how users respond over time. By keeping inference, interaction storage, and model updates entirely on-device, the system addresses privacy concerns while enabling personalized AI that improves after deployment.