CORE Improves Personalized LLM Alignment by Limiting Persona Drift

Summary

Persona drift is a persistent challenge for language models that maintain user profiles across long conversations: the model must recognize genuine preference changes without treating temporary, ambiguous, or unresolved observations as durable updates. The paper proposes CORE, a method that separates evidence local to the current turn from revision of a persistent persona state and selectively updates grounded preferences through uncertainty-aware belief revision. It also introduces PERSIST, a held-out post-anchor benchmark designed to stress persona-state robustness during sequential interaction, including ambiguity, conflict, and controlled social influence. Across ALOE, PersonaChat, and PERSIST, the authors report gains in personalized alignment and robustness, along with complementary improvements in normalized closed-slot state fidelity. Human evaluation and mechanistic control experiments further support the value of explicit update control, beyond improvements attributable to stronger generation or persistent memory alone.