LLM multi-agent systems are susceptible to herd effects and runaway cascades

Drawing on the classic Music Lab experiment, the researchers tested the behavior of large-scale LLM agent groups recommending unfamiliar songs, and found that exposure to cumulative recommendation counts triggers runaway cascades, making collective recommendations extreme, unpredictable, and increasingly divergent from the songs' intrinsic appeal. Compared with humans, LLM groups are generally more susceptible to herd effects, and the cascades are not self-correcting as they are in humans. The authors note that this vulnerability has implications for multi-agent collaboration scenarios such as recommender systems, automated evaluation, complex problem solving, and unauthorized collusion.