Research on Multiagent-Driven Networked Sentiment Regulation Mechanism and a Credible Simulation Framework

Against the backdrop of the deep penetration of social media into social life, networked sentiment polarization has become an important challenge for public opinion research and governance analysis. To address the reactive and lagging nature of traditional public opinion response strategies, this study proposes a large language model–driven multiagent simulation framework for modeling networked sentiment dynamics and comparing regulation strategies. Based on multidimensional user profiles, including user identity, sentiment tendency, personality traits, and reading habits, the framework simulates content generation and interactive behaviors such as commenting, liking, and reposting, thereby forming a feedback loop between agent behaviors and the public opinion environment. This study introduces the “rate of change in sentiment entropy†as a core metric to quantify sentiment fluctuation intensity. On this basis, static and dynamic regulation strategies are designed and compared. Experiments based on two real-world Weibo public opinion events show that the proposed simulation environment can, to a certain extent, reproduce the dynamic characteristics of real public opinion data in terms of semantic content, stylistic features, sentiment evolution, and interaction behavior. The regulation experiments further reveal a “pulse-rebound†pattern and suggest that information-source credibility and adaptive triggering mechanisms play important roles in the recovery and stabilization of the simulated sentiment system. Overall, this study provides a controlled simulation environment and analytical reference for sentiment evolution modeling and regulation strategy comparison.