Public opinion dissemination simulation based on large language model multi-agent systems.

Area of Science:

• Computational Social Science
• Artificial Intelligence
• Social Network Analysis

Background:

• The internet significantly influences public opinion, but traditional simulations lack fidelity and are costly.
• Existing models struggle with evolutionary dynamics, behavioral homogenization, and high resource demands.

Purpose of the Study:

• To develop a realistic public opinion simulation system integrating macro-level diffusion and micro-level cognition.
• To address limitations of traditional models in generalizability and resource intensity.

Main Methods:

• Proposed a Large Language Model (LLM)-based multi-agent simulation framework for public opinion dissemination.
• Constructed behavior probability profiles calibrated with social media data (exponential/normal distributions).
• Integrated LLM as a cognitive core for dynamic content generation using PSOP and GISP.

Main Results:

• Framework demonstrated cross-scenario robustness, replicating opinion evolution patterns (incubation-eruption-decay) without parameter tuning.
• Achieved normalized agent behavior distribution entropy of 0.69 and Distinct-2 metric of 0.83, mitigating homogenization.
• Framework supports zero-shot cold start with negligible computational costs for domain adaptation.

Conclusions:

• The unified agent architecture enables efficient public opinion evolution simulation.
• Provides a low-cost, high-fidelity methodological paradigm for public opinion crisis management.
• Offers significant improvements over traditional deep learning methods in simulation efficiency and adaptability.