Multi-LLM Agent Simulation Reproduces Hate Speech Cascades with 97% Hostile Stance Accuracy
Researchers find hate speech spreads star-like on Bluesky, multi-LLM simulators can now model it faithfully.
A new arXiv paper by Fan Huang tackles the challenge of modeling hateful content propagation on online platforms. The author compares three real hateful Bluesky cascades against a benign control, finding that 97.4–99.7% of reposters adopt a hostile stance. Toxicity-engagement homophily is higher on the diffusion tree than on the follower graph, and hateful cascades exhibit a star-like topology (most reposts come directly from the root) versus a tree-like structure for benign content.
To simulate these dynamics, Huang deploys a multi-agent LLM system where each agent’s reshare decision depends on user profile, community context, and post content. The simulator faithfully reproduces the stance monoculture and the direction of toxicity deltas. A structured ablation identifies agent heterogeneity as the leading fidelity factor. An intervention strategy targeting amplifiers on dense networks reduces hateful reshares by 7.5–12.9% while affecting only 5.7% of benign posts. This work opens the door to deploying LLM-driven simulations for testing moderation strategies at scale.
- In real Bluesky hate cascades, 97.4–99.7% of reposters show hostile stances, and topology is star-like (single root) vs. tree-like for benign content.
- A multi-LLM agent simulator using profile, community, and content features reproduces stance monoculture and toxicity-delta direction better than classical cascade models.
- Targeting amplifiers on dense networks reduces hateful reshares by 7.5–12.9% with only 5.7% benign collateral, and agent heterogeneity is the top fidelity factor.
Why It Matters
Platforms can now use LLM-driven agent simulations to design and test hate speech interventions before real-world deployment.