LLM Opinion Dynamics Simulation: Exploring Group Opinion Emergence and Social Evolution in Large Language Models

This project uses multi-agent simulation to conduct in-depth research on how large language models (LLMs) simulate the formation, spread, and polarization of opinions in human society. It aims to reveal the laws of group opinion emergence in AI systems and provide new perspectives for computational sociology and AI safety research.

Research Background: When AI Meets Social Sciences

Human society is a complex adaptive system where interactions between individual opinions evolve into macroscopic phenomena (such as polarization and consensus). Traditional rule-based computational models (DeGroot, Hegselmann-Krause) simplify cognition and language, making it difficult to capture real complexity. The rise of LLMs provides agent tools that understand natural language and have reasoning capabilities, and this project attempts to use them to simulate human opinion formation and group dynamics.

Technical Implementation Architecture

Multi-Agent Simulation Framework

The system adopts an LLM-driven multi-agent architecture, where agents interact through social networks. Network structures include fully connected, small-world, scale-free, and community-structured networks.

Opinion Representation and Evolution Mechanism

Opinions are represented as multi-dimensional vectors or natural language statements. In each iteration, agents receive neighbors' opinions and update their own opinions by combining their cognitive characteristics (openness, stubbornness). This supports complex semantic positions rather than binary choices.

Emergent Phenomenon Detection

Group emergent patterns are quantified using indicators such as consensus degree (variance/entropy), polarization index (bimodal shape), number of clusters, and convergence speed.

Key Research Findings and Insights

Opinion Resilience and Persuasion Threshold

LLM agents have a "persuasion threshold": when the opinion difference is too large, they refuse to change and may even show a rebound effect, echoing the cognitive dissonance theory.

Impact of Network Structure

• Fully connected network: Reaches consensus quickly but easily forms echo chambers
• Community-structured network: Stabilizes opinion clusters and simulates political polarization
• Small-world network: Balances consensus and diversity, with efficient spread and low polarization

Collective Intelligence and Bias Amplification

LLM groups can exhibit collective intelligence (collective decisions are better than the average of individuals), but may also amplify biases (majority stereotypes persuade the minority).

Methodological Innovation and Academic Value

Paradigm Shift in Computational Sociology

Shifting from simplified mathematical models to "LLM-based agent simulation", the advantages include high fidelity (processing natural language), scalability (generating diverse agents), and interpretability (insight into cognitive mechanisms).

Implications for AI Safety and Alignment Research

Understanding multi-agent interaction dynamics provides a theoretical basis for designing fair and inclusive AI systems, helping to address challenges such as opinion polarization and fake news spread.

Limitations and Future Directions

Current Limitations

• Lack of real human emotions and embodied experience
• Biases in training data
• Lack of repeatability of simulation results due to LLM randomness

Future Research Directions

• Heterogeneous agents (mixing different LLM architectures)
• Dynamic networks (adjusting structure as opinions evolve)
• External information injection (impact of news, authoritative opinions)
• Human-machine hybrid experiments (comparison between LLMs and real humans)

Conclusion

This project is at the cross-frontier of AI and social sciences, providing a new path for understanding LLM behavioral characteristics and innovating computational sociology methodologies. In the future, it is expected to build simulation models closer to real society to address challenges such as opinion polarization in the information age.