Unveiling the Potential of LLMs in Simulated Society: A Knowledge-Driven LLM Agent Framework for User Modeling

User modeling serves as the cornerstone of modern recommendation systems, focusing on the precise identification of user preferences and behavioral signatures to enable personalized service delivery. Existing recommendation methods face issues like sparse data and lack of transparency. The advent of large language models (LLMs) has brought new possibilities for inferring user preferences, which provides natural-language-based recommendation rationales and affective expressions. Motivated by AgentSociety Challenge @ WWW 2025 focused on simulating user-item interactions, we propose a knowledge-driven LLM agent framework for simulating user-item review interactions, which includes three modules: preference refinement, dual-signal injection, and category distinguisher. Specifically, preference refinement aims to model user and item profiles, enabling LLMs to perceive users and items comprehensively. Dual-signal Injection aims to incorporate external knowledge, such as collaborative filtering and distribution knowledge. Category distinguisher is designed to analyze the differences between true data and simulated data. Both online and offline experimental results demonstrate the effectiveness of this framework and prove the great potential of LLMs in user modeling.