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

This collection contains four recent papers on social media modeling, manipulation, and LLM behavior. MF-MDP: a mean-field + per-agent MDP framework that models latent per-agent opinion states to enable stable long-horizon social simulations (up to ~40k interactions) and sharply reduced long-horizon and reversal KL divergence vs. prior MF-LLM. ContentFuzz: a confidence-guided LLM-based rewriting system that preserves human-interpreted intent while flipping machine-inferred stance labels to route posts beyond existing affinity clusters, enabling evasion of stance-detection signals used in recs/