Coupling Macro Dynamics and Micro States for Long-Horizon Social Simulation

A research team led by Yunyao Zhang has introduced MF-MDP, a novel framework for simulating social network dynamics that addresses a critical flaw in current LLM-based simulators. While existing models like MF-LLM focus primarily on aggregate group behavior, they largely ignore the internal cognitive states of individuals, making them unreliable for long-term predictions and incapable of capturing gradual opinion shifts that lead to reversals. MF-MDP solves this by tightly coupling macro-level collective dynamics with micro-level individual states, explicitly modeling each agent's latent opinion through a state transition mechanism.

Technically, MF-MDP combines individual Markov Decision Processes at the micro level with a mean-field collective framework at the macro level. This architecture allows agents to change their internal states gradually rather than reacting instantly to stimuli, enabling the system to distinguish between agents who are close to switching opinions and those who are firmly entrenched. The results are dramatic: MF-MDP supports stable simulation of social processes with up to 40,000 interactions, compared to just 300 in the baseline MF-LLM model, while reducing long-horizon prediction error (measured by KL divergence) by 75.3% and reversal prediction error by 66.9%.

The framework's ability to maintain accuracy over extended horizons represents a significant breakthrough for applications requiring long-term social forecasting. By mitigating the 'drift' phenomenon where simulations become increasingly inaccurate over time, MF-MDP enables more reliable modeling of complex social phenomena like political polarization, market adoption curves, and public health behavior changes. The code has been made publicly available, allowing researchers and developers to build upon this architecture for more accurate social simulations.