IRIS Learns Dynamic User Personas from Implicit Interactions.

Personalizing large language models (LLMs) is crucial for enhancing user experience, but current methods often depend on explicit feedback like pairwise comparisons or demographic data, which limits their applicability in natural interaction settings. Researchers have developed IRIS (Iterative Refinement of Implicit Streams), a framework designed to learn dynamic user personas from implicit interaction streams. IRIS extracts behavioral signals from everyday conversations and continuously refines persona representations through a prediction-driven closed loop, eliminating the need for explicit user feedback. This approach offers a scalable alternative to traditional personalization methods. A proof-of-concept study using synthetic data demonstrated IRIS's ability to produce stable personas and differentiate individual users. Further validation on real-world Reddit data showed IRIS achieving the highest decision prediction accuracy (61.0%) among evaluated methods, outperforming static personas, memory-only retrieval, and non-personalized baselines. These results highlight the potential of implicit behavioral modeling for adaptive conversational systems and embodied agents.