Multi-Agent Systems with LLMs for Cyberbullying Intervention Training and Safe Social Simulation
Ongoing research project
Demonstration
The run below is Scenario G (a tackle at break broke a popular boy’s wrist before the tournament) in human-victim mode: three DeepSeek V4 Flash agents play the bully, his supporter and the victim’s defender, while a person sits in the victim seat and types under the same time pressure a real group chat imposes. The left side is the chat as the human saw it. The right side is what the human never sees: each round’s bids (every agent names an urgency or passes, and the speaker is drawn in proportion to urgency), the floor policy waiting on the human, and each agent’s turns and harm labels accumulating.
Six and a half minutes of wall clock compressed to about a minute and a half. The human's messages are verbatim, typos included. The run was stopped by hand at message 37; that is not a social outcome.
Full annotated transcript: every bid, pass and floor-policy window, with the labeler's debrief
Each agent message unfolds to show who else wanted the floor that round and why. Open it in its own tab →
Keywords: multi-agent simulation, urgency-proportional turn-taking, emergent termination, LLM-as-judge annotation, synthetic dataset generation, cyberbullying detection.