HypReflect Uses Preference Hypotheses for Continual LLM Personalization

Summary

Researchers introduce HypReflect, a framework for continually personalizing LLM assistants. It infers explicit, uncertainty-aware hypotheses about user preferences from heterogeneous signals, revises them as evidence accumulates, and uses them to guide self-distillation. Across online, multi-session, and implicit-signal settings, HypReflect outperforms raw-history and incremental-update baselines, while generalizing to unseen users and cross-domain scenarios.