A new paper introduces PocketTSLM, an on-device soft-prompt time-series language model designed for privacy-preserving smartphone reasoning. The authors argue that most mobile systems currently reduce physiological streams to summaries or text, which discards the temporal structure that matters for personal health analysis.
PocketTSLM instead processes raw time-series data directly, using soft prompts to adapt the model for mobile hardware. By keeping inference on the device, the approach avoids sending sensitive physiological data to remote servers, addressing a key privacy concern for health-related smartphone applications.
The paper is available on arXiv and describes the model's design and deployment strategy, though the abstract does not include benchmark results or comparisons with existing methods.