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 bymisinformation and extreme emotions, posing severe threats to mainstream values and public trust.While most existing studies focus on passive detection of negative content, systematic approachesfor proactively optimizing opinion guidance remain scarce.To address this issue, we propose a SocialInfluence Text Generation Network, namely SITGNet, a novel multi-agent social simulation frame-work powered by large language models (LLMs). SITGNet populates a digital social sandbox withcognitively sophisticated agents, each possessing distinct profiles, memories, and behavioral patternsgrounded in empirical data. This enables the high-fidelity reproduction of opinion dynamics, captur-ing emergent phenomena from complex user interactions and psychological transitions. Within thisplatform, we introduce Guiding Agents, a specialized class of agents designed to proactively steerdiscourse. These agents leverage a multi-stage retrieval-augmented generation (RAG) pipeline to syn-thesize information from external knowledge bases, grounding their outputs in factual evidence togenerate nuanced and influential content. Through extensive, large-scale simulations, we systemati-cally evaluate various intervention strategies. Our results demonstrate that a hybrid deployment ofGuiding Agents can effectively suppress negative sentiment propagation, alleviate group polariza-tion, and guide opinion trajectories toward rational outcomes. In addition to building the simulationframework, this study treats Social Influence Text Generation as a key modeling objective, enablingsystematic evaluation of text-based opinion guidance strategies.