Towards efficient optimization of multi-agent social simulation via large language models

Public opinion on social media during major societal events is highly vulnerable to manipulation by misinformation and extreme emotions, posing severe threats to mainstream values and public trust. While most existing studies focus on passive detection of negative content, systematic approaches for proactively optimizing opinion guidance remain scarce. To address this issue, we propose a Social Influence Text Generation Network, namely SITGNet, a novel multi-agent social simulation framework powered by large language models (LLMs). SITGNet populates a digital social sandbox with cognitively sophisticated agents, each possessing distinct profiles, memories, and behavioral patterns grounded in empirical data. This enables the high-fidelity reproduction of opinion dynamics, capturing emergent phenomena from complex user interactions and psychological transitions. Within this platform, we introduce Guiding Agents, a specialized class of agents designed to proactively steer discourse. These agents leverage a multi-stage retrieval-augmented generation (RAG) pipeline to synthesize information from external knowledge bases, grounding their outputs in factual evidence to generate nuanced and influential content. Through extensive, large-scale simulations, we systematically evaluate various intervention strategies. Our results demonstrate that a hybrid deployment of Guiding Agents can effectively suppress negative sentiment propagation, alleviate group polarization, and guide opinion trajectories toward rational outcomes. In addition to building the simulation framework, this study treats Social Influence Text Generation as a key modeling objective, enabling systematic evaluation of text-based opinion guidance strategies.