SensorPersona: An LLM-Empowered System for Continual Persona Extraction from Longitudinal Mobile Sensor Streams

A team of researchers has introduced SensorPersona, a novel system that leverages Large Language Models (LLMs) to continuously build comprehensive user profiles directly from the sensor streams of mobile devices. Unlike current methods that rely on self-disclosed chat histories, SensorPersona performs 'person-oriented context encoding' on longitudinal, multimodal data—such as location, movement, and app usage—collected unobtrusively over time. Its core innovation is a 'hierarchical persona reasoning' engine that integrates analysis within single episodes (intra-episode) and across longer time spans (inter-episode) to infer stable personas encompassing a user's physical habits, psychological traits, and life context.

The system was rigorously evaluated on a substantial, real-world dataset containing 1,580 hours of sensor data from 20 participants, collected over up to three months across 17 cities on three continents. The results demonstrate a significant leap in capability: SensorPersona achieved up to 31.4% higher recall in persona extraction compared to state-of-the-art baselines. More importantly, when these extracted personas were used to guide an AI agent's responses, they resulted in an 85.7% win rate in user preference tests and notable improvements in user satisfaction. The system also includes mechanisms for 'clustering-aware incremental verification' and 'temporal evidence-aware updating' to adapt personas as user behaviors evolve, ensuring the profiles remain accurate over time.