Simulation of Persuasive Communication Using Large Language Model Agents

Abstract:
Large language models (LLMs) hold significant potential for simulating social phenomena. Drawing on Social Judgment Theory (SJT) and opinion dynamics, this study employs an LLM-driven agent-based simulation approach to investigate how persuasion and attitude evolution unfold within social networks, as well as the key factors influencing these processes. Using the AgentSociety simulation platform, we construct an SJT-based agent model that incorporates a continuous opinion interaction mechanism to simulate how individuals’ attitudes change under topics with varying levels of controversy. We further analyze the roles of network structure, macro-level social opinion, and micro-level psychological factors.

Experimental results reveal that: (1) Group attitudes exhibit a clear convergence trend during interactions, ultimately forming either consistent supportive or neutral opinions; (2) At the micro level, negative emotions serve as the key driver of attitude change, whereas individual states and positive emotions exert relatively weaker influence; (3) Network topology significantly affects the speed of attitude evolution — BA scale-free networks accelerate opinion convergence due to the presence of hub nodes, while WS small-world networks slow down the diffusion process because of their local clustering; (4) Macro-level social opinions shape the overall direction of group attitude evolution.

By integrating communication theories with LLM-based simulation, this study proposes a multi-level persuasion communication model, revealing how network structures and emotional mechanisms jointly shape opinion dynamics. The findings provide theoretical implications for public opinion governance and policy-oriented persuasion strategies. Full Paper Link: https://zhou-zhen-feng.github.io/assets/pdf/agent_simulation_paper.pdf