Personalizing Human-LLM Interactions through Mixed Profiling

While Large Language Modelss (LLMss) are transitioning into the role of companions, human-LLMs interaction remains largely indifferent to the nuanced individualities of the human user. Most current personalization of LLMss relies on behavioral data, while established psychometric methods, such as questionnaires that assess individual traits and states, are rarely leveraged. We propose Mixed Profiling, a novel approach that combines standardized short questionnaires with adaptive LLMs dialogs to create precise user profiles. We implemented and evaluated the Mixed Profiling approach for the mental health context as an example in an online study (N=40). Participants completed questionnaires and participated in an LLMs-based chat. They rated the Mixed Profiling approach’s results as significantly more trustworthy than those of a basic questionnaire-based profile. By bridging explicit psychometrics and data-driven personalization, our work lays the foundation for a more psychologically grounded and transparent personalization in human-LLMs interaction.