Hypotheses-Guided Self Distillation for Continual Personalization

As people increasingly rely on large language model assistants, adapting to individual user preferences over time has become essential. However, users rarely articulate their preferences fully, leaving models to infer them from heterogeneous, noisy, and indirect behavioral signals. Current approaches often depend on processing raw interaction histories or using costly reward-based optimization, which can be difficult to scale reliably.

To address these challenges, the researchers introduce HypReflect, a framework designed for continual personalization. The system infers explicit, uncertainty-aware preference hypotheses from diverse user interactions, reflectively updates these hypotheses as new evidence arrives, and integrates this refined user model using hypotheses-guided self-distillation.

Across evaluations covering online personalization, multi-session interactions, and implicit signals, HypReflect outperformed traditional raw-history and incremental-update baselines. It also demonstrated strong generalization to new users and cross-domain tasks while maintaining stability across different context budgets, pointing toward more scalable and dependable personalized assistants.