New Framework Personalizes LLM Assistants with User Preference Hypotheses

Large Language Model (LLM) assistants are becoming integral to daily life, making continuous adaptation to individual user preferences crucial for effective long-term interactions. Current personalization methods often rely on raw interaction histories or expensive reward-based optimization, struggling with the latent, noisy, and unstated nature of user preferences. Researchers have introduced HypReflect, a novel framework designed for reliable and scalable continual personalization. This system infers explicit, uncertainty-aware preference hypotheses from various user signals, then reflectively refines these hypotheses as new evidence emerges. The resulting user model is integrated through a process called hypotheses-guided self-distillation. Experiments demonstrate that HypReflect surpasses several baseline methods, including those using raw history or incremental updates, across online personalization, multi-session interactions, and implicit behavioral signals. The framework also shows strong generalization capabilities to new users and domains, maintaining stability across different context budgets and enabling reusable, focused personalization.