Persona-Driven AI Agent Laboratories for Opinion Dynamics: Learning Effective Models From LLM Debates
This project builds AI agent laboratories where large language models with assigned personas engage in structured debates to simulate opinion dynamics in social systems. By learning effective low-dimensional models from these simulated debates—capturing polarization, consensus formation, and information cascades—the project bridges multi-agent AI simulation with rigorous mathematical modeling of collective behavior.
🧑🎓 IOL Project Members
🤝 External Project Members
Principal Investigator
🪙 Funding
This project is being funded by the Berlin Mathematics Research Center MATH+ (project ID EF-MA-Sys-3), itself funded by the German Research Foundation (DFG) under Germany's Excellence Strategy (EXC-2046/1, project ID 390685689) from October 2026 to September 2029.
🔬 Project Description
Opinion dynamics models the evolution of beliefs, attitudes, and opinions in social systems through interaction networks. Traditional models (e.g., bounded-confidence, voter, and Ising-type models) provide mathematical tractability but often fail to capture the nuanced, context-dependent reasoning that drives real human opinion change.
This project introduces persona-driven AI agent laboratories: structured environments where LLM-based agents, each assigned a detailed persona with demographic, ideological, and epistemic characteristics, engage in multi-turn debates on contested topics. From these simulated interactions, we learn effective, reduced-order opinion dynamics models that:
- Capture the statistical signatures of polarization, echo chambers, and consensus formation observed in real social data;
- Are calibrated against empirical survey and social-media data through Bayesian inference;
- Enable counterfactual analysis and scenario forecasting (e.g., how does a misinformation intervention change the trajectory of a debate?).
The project combines multi-agent LLM orchestration, Bayesian system identification, and the mathematics of interacting particle systems to create a new paradigm for data-driven social science.