EvoUser: From Interaction History to Evolving User State for Long-Term LLM Personalization

Personalized large language models aim to generate responses tailored to individual users by leveraging their historical interactions or user profiles. However, existing approaches either represent users through textual memories, which are effective for explicit facts but struggle to capture weak and distributed signals emerging across interactions, or rely on user-specific model adaptation, which incurs substantial computational cost.We propose EvoUser, a long-term personalization framework that combines a fixed-length, continuously evolving user state with query-driven textual retrieval. The framework uses the evolving user state to capture implicit information accumulated across interactions, while leveraging query-relevant textual retrieval to preserve explicit and traceable contextual evidence. We jointly train the state interface and response model through RAG-conditioned supervised learning and GRPO optimization. The complete system achieves accuracies of 82.68% and 67.53% on PersonaMem-v2 and PrefEval, respectively. Further controlled analyses demonstrate that even when the user states are frozen, they retain decodable user-specific attributes and preferences, validating that the continuous user state learns and stores effective personalized representations.