Simulating Hate Speech Cascades with Multi-LLM Agents: Empirical Grounding, Modeling Fidelity, and Intervention Strategies

The study presents a multi-agent large language model (LLM) simulator designed to model the propagation of hate speech on online platforms more accurately than classical cascade models. The research analyzes three hate speech cascades from Bluesky, revealing that 97.4–99.7% of reposters adopt a hostile stance and that the network topology for hateful content is star-like, contrasting with the tree-like structure of benign content. The findings highlight the importance of agent heterogeneity in improving modeling fidelity and suggest that targeted interventions can reduce harmful spread by 7.5–12.9% while maintaining low collateral impact.