As large language model (LLM) assistants become commonplace in everyday life, continuously adapting to individual user preferences is essential for long‑term interaction quality. In practice, preferences are rarely stated explicitly; they emerge from heterogeneous, latent, and noisy signals. Existing approaches either rely on raw interaction logs or require costly reward‑based optimization, which limits reliability and scalability.
We introduce HypReflect, a framework that infers explicit, uncertainty‑aware preference hypotheses from diverse user signals. The system reflexively refines these hypotheses as new evidence arrives and incorporates the resulting user model into the language model via hypotheses‑guided self‑distillation. The pipeline consists of: (1) extracting latent preference features from dialogues, clicks, and behavioral data; (2) constructing a Bayesian, uncertainty‑aware set of preference hypotheses; (3) performing self‑distillation guided by the posterior of these hypotheses, allowing the model to retain general capabilities while progressively enhancing personalization.
Experiments across three representative settings—online real‑time personalization, multi‑session adaptation, and implicit behavioral signal modeling—show that HypReflect outperforms baselines that use raw histories or incremental updates in terms of accuracy, user satisfaction, and robustness. Additional cross‑user and cross‑domain evaluations demonstrate hypothesis reusability and strong generalization to unseen users, while maintaining stable performance under varying context budgets, enabling more focused and interpretable personalization.
These findings suggest that explicit, revisable preference hypotheses combined with self‑distillation provide a reliable and scalable path toward continual personalization.
Review