Multi-LLM Agents Simulate Hate Speech Propagation for Moderation Research

A new study explores the use of multi-agent large language model (LLM) systems to accurately model the spread of hateful content on online platforms. Traditional cascade models often fall short because they don't explicitly account for user profiles, community dynamics, or specific content characteristics that drive hate speech propagation. This research investigates whether LLM-based agents, which can make reshare decisions based on these nuanced factors, can offer more faithful simulations. The team analyzed three real-world hate speech cascades and one benign control cascade from Bluesky. Their empirical observations revealed distinct patterns for hateful content: a high percentage of hostile reposters, stronger toxicity-engagement homophily, and a star-like diffusion topology where most reposts originate directly from the initial post. In contrast, benign content showed a more tree-like, multi-hop propagation. The multi-LLM agent simulator successfully replicated the "stance monoculture" and toxicity-engagement patterns observed in the real data. A key finding was that agent heterogeneity—meaning diverse user profiles and behaviors among the agents—was the most critical factor for achieving high simulation fidelity. The study also demonstrated that targeted interventions, specifically focusing on "amplifiers" within dense networks, could reduce hate speech propagation by 7.5-12.9% while minimizing impact on benign content.