A new paper on arXiv examines how always-on virtual assistants personalize on-device to new user-specific commands via few-shot fine-tuning, such as LoRA. The authors note that this adaptation is evaluated almost exclusively on whether the model learns the new commands, and not on other consequences of that adaptation.
The paper's title asks whether personalization breaks trust, and its proposed focus is open-set rejection calibration under few-shot on-device intent adaptation. In other words, the work centers on calibrating the assistant to reject inputs it should not act on, rather than only recognizing newly added intents.
Because the abstract is truncated in the source, the full method and results are not available here; this summary covers only the stated motivation and proposed direction.