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

Researchers introduce SensorPersona, an LLM-based system that continuously extracts user personas from mobile sensor data rather than chat histories, achieving 31.4% higher recall in persona extraction and 85.7% win rate in personalized agent responses. The system processes multimodal sensor streams to infer physical patterns, psychosocial traits, and life experiences across longitudinal data collected from 20 participants over three months.

SensorPersona addresses a fundamental limitation in current LLM personalization: reliance on explicit chat data that reveals only curated self-presentation rather than authentic behavioral patterns. By leveraging continuous, unobtrusively collected mobile sensor streams—including location, activity, and contextual signals—the system captures what users actually do versus what they say about themselves. This distinction matters because behavioral patterns often diverge from self-reported preferences, enabling more accurate personalization.

The technical approach employs hierarchical reasoning that processes sensor contexts at multiple levels, distinguishing between individual episodes and longer-term patterns to infer stable personas. The clustering-aware incremental verification and temporal evidence-aware updating mechanisms allow the system to adapt as user behaviors evolve, addressing the reality that personas are not static. Testing on 1,580 hours of real-world sensor data across 20 users demonstrates practical viability at meaningful scale.

For the AI industry, this research signals movement toward ambient intelligence systems that understand users through behavioral observation rather than explicit input. The 31.4% improvement in persona extraction and 85.7% win rate in personalized responses suggest substantial practical gains. Privacy considerations become critical here—continuous sensor monitoring raises data protection questions that implementations must address through on-device processing and user consent frameworks.

Looking forward, SensorPersona-type systems could reshape how AI agents personalize at scale, particularly for mobile applications. Adoption depends on resolving privacy-regulatory tensions and demonstrating user value clearly exceeds surveillance concerns. Integration with commercial LLM platforms remains uncertain but represents logical next steps for companies pursuing differentiated personalization capabilities.

• →SensorPersona infers user personas from continuous mobile sensor data, achieving 31.4% higher recall than chat-history-based approaches.
• →The system uses hierarchical reasoning to extract physical patterns, psychosocial traits, and life experiences from behavioral observations.
• →Real-world evaluation on 1,580 hours of sensor data from 20 users showed 85.7% win rate in persona-aware agent responses.
• →Adaptive mechanisms allow the system to track evolving personas through clustering-aware verification and temporal evidence updating.
• →Behavioral data from sensors captures authentic patterns that diverge from self-reported preferences, enabling more accurate LLM personalization.