USHB: A Unified Framework for Simulating Human Behaviors in Agent Society through User-and-Item Modeling
Published in WWW 2025, 2025
Renhuo Zhao*, Hailong Yang*, Mingxian GU, Jianqi Wang, Wu Long, Zhaohong Deng. WWW 2025 (CCF A). https://doi.org/10.1145/3701716.3719227
USHB
Emulating real-world human behaviors within Artificial Intelligent (AI) agent systems continues to be a formidable challenge, as existing methodologies frequently have difficulty integrating the intricacies of real-world scenarios and personal preferences.To address this issue, we propose USHB, a mulit-agent framework that emphasizes advanced user and item modeling along with communication style simulation. USHB consists of 3 modules: a Knowledge-Mining Module (KMM), a User-and-Item Modeling Module (UIMM), and a Reasoning Module (RM). USHB utilizes Large Language Models (LLMs) to predict responses, reviews, and ratings tailored to individuals, guaranteeing results that are both coherent and contextually appropriate. USHB is capable of delivering precise, detailed simulations that closely mimic human behavior. We evaluated USHB using datasets from Yelp, Amazon, and Goodreads, where it consistently outperformed baseline methods. Moreover, USHB demonstrated strong generalization and maintained stable performance across a variety of model configurations. These advancements make USHB a valuable contribution to dynamic, context-aware behavior simulation, achieving a top-three ranking in the 2025 AgentSociety Challenge.