Language models improve tracking of changing user preferences
Toward Robust Personalized Alignment for LLMs: Mitigating Persona Drift in Multi-Turn Dialogue
Artificial Intelligence
Summary
When people chat a lot with AI language models, their preferences and personality details can change over time. The authors found that typical AI models struggle to keep track of these changes accurately, sometimes getting confused by short-term or unclear information. They created a method called CORE that helps the AI decide when to update what it knows about a person’s preferences and when to hold back. They also introduced a new test called PERSIST to check how well AI keeps a consistent sense of a user's persona across many conversations. Their method improved how well AI stays true to users' personalities over time.
What this means in practice
- •For chatbot developers: Improve chatbot ability to track and update user preferences accurately in long conversations.
- •For customer support teams: Enhance virtual agents to remember changing customer details and avoid mistakes from temporary or ambiguous inputs.
Authors
Youyuan Zhang, Siyuan Li, Fangming Liu, Jing Li
Abstract
Persona drift remains a central challenge for personalized language models, as user profiles evolve over long interactions rather than remain permanently fixed. Models must therefore revise persistent persona states when preferences genuinely change, while avoiding updates driven by transient, ambiguous, or unresolved observations. We propose CORE, which separates turn-local evidence from persistent persona-state revision and selectively updates grounded user preferences through uncertainty-aware belief revision. We also introduce PERSIST, a held-out post-anchor benchmark for persona-state robustness under sequential interaction stress, covering ambiguity, conflict, and controlled social influence. Across ALOE, PersonaChat, and PERSIST, CORE improves personalized alignment and robustness, with complementary gains in normalized closed-slot state fidelity. Human evaluation and mechanistic controls further support explicit update control beyond stronger generation or persistent memory alone.