Qwen3.5-4B agent swarms herding persists after RL

ashtonchew.com24 min read

The author investigates the phenomenon of herding in multi-agent AI swarms, finding that agents frequently drop correct private clues or facts to conform with their peers or misleading context. Through experiments with small Qwen models on search and puzzle tasks, the author observes that neither scaling agent count nor applying reinforcement learning with team or individual credit schemes successfully mitigates this conformity behavior. Open article →