LLM Multi-Agent Systems Are Vulnerable to Herding and Runaway Cascades

[arXiv] Farnoosh Hashemi, Shiyu Ji, Joice Chen, and Michael W. Macy, modeled on the classic Music Lab experiment, tested social influence when unfamiliar songs are recommended in large-scale populations of LLM agents. The authors say that when agents can see cumulative recommendation counts, runaway cascades occur, collective recommendations become extreme and unpredictable, and they increasingly deviate from the songs' intrinsic appeal. The paper says that compared with humans, LLM populations are generally far more prone to herd behavior, and cascades are harder to self-correct; this vulnerability matters for multi-agent scenarios such as recommendation, automated evaluation, collaborative problem-solving, and unauthorized collusion.