LLM Multi-Agent Systems Are Vulnerable to Herding and Runaway Cascades
LLM multi-agent systems are increasingly deployed in settings where one model's output becomes another's input, creating endogenous feedback loops that could destabilize collective judgment. This paper adapts the classic Music Lab experiment to test whether populations of LLM agents are vulnerable to herding and runaway cascades when making sequential song recommendations. Across 13 models and 168,480 agents, the authors find that LLM populations diverge drastically from independent judgment, produce unpredictable outcomes across replicate worlds, and largely fail to show the self-correcting dynamics observed in human benchmarks. The results matter because they shift evaluation from isolated model behavior to emergent population-level dynamics in feedback-driven systems such as recommender systems and automated evaluation.
The paper makes a timely and empirically rigorous contribution, but its boldest claims rest on comparisons that blend population type with task redesign. The core within-LLM findings are robust: every model exhibited significant herding (p<0.001), non-reasoning models were more volatile than reasoning models, and higher benchmark intelligence correlated with lower susceptibility. However, the human–LLM gap is difficult to interpret causally. The authors concede that "These differences limit attribution of human–LLM contrasts to population type" (Section 4), and without a matched human replication it is unclear how much of the divergence stems from LLM social learning versus the switch from audio to lyrics, downloads to recommendations, or grids to lists. The leap from song recommendations to implications for "unauthorized conspiracies" also outpaces the experimental evidence.
The parallel-worlds design and outcome measures are well-motivated and cleanly executed. KL divergence quantifies displacement from a population-specific independent baseline, while unpredictability captures path dependence across replicate worlds. The intelligence gradient is striking: "Both KL divergence and relative unpredictability decline as intelligence scores increase, regardless of influence condition" (Section 3.3), with Spearman $\rho = -0.851$ for divergence under strong influence. The authors also conduct sensible robustness checks for sampling temperature and consideration-set size that preserve the main patterns, and they are appropriately cautious that the intelligence association is "descriptive, not causal" (Appendix C.1).
The most significant limitation is that the human benchmark is not a true experimental control because the task was substantially redesigned for LLM agents. As the authors note, "A matched human replication of this study is needed to rule out these design differences as the explanation for observed behavioral differences" (Section 4). The experiment also studied only homogeneous populations of a single model acting without incentives, communication, or ground-truth accuracy, which limits external validity for real-world multi-agent systems. Additionally, the self-correction analysis comes with the caveat that "the results do not establish that correction could never occur over longer sequences" (Section 3.4), and the strong-influence condition is partially confounded by position bias, though the authors argue this would likely understate rather than overstate the human–LLM gap.
The evidence strongly supports the claim that endogenous feedback produces unpredictable cascades within the LLM populations. "In LLM populations, early movers predicted later popularity better than intrinsic appeal did in 12 of 13 models under weak influence and in all 13 under strong influence" (Section 3.2), which is a powerful demonstration of path dependence. Comparisons to related work are generally fair: the authors correctly distinguish their endogenous popularity-count design from exogenous conformity paradigms (Section 1). However, the abstract's framing that "unlike humans, LLM cascades were less able to self-correct" slightly overreaches, because the self-correction diagnostic depends on intrinsic appeal measured under a different task format (lyrics-only recommendations versus human audio downloads), making it impossible to separate cascade dynamics from task-specific sensitivity.
The methods are documented in unusual detail—prompts, personas, inference settings, and smoothing procedures are all provided in Appendix B—but the paper does not mention releasing code, raw outputs, or evaluation scripts, which would block straightforward independent reproduction of the 168,480-agent experiment. Hyperparameters are partially specified: non-reasoning models used a "sampling temperature of 1.0" (Appendix B.4), but reasoning models used "medium reasoning effort where an effort level was supported," a vague setting tied to proprietary APIs. The reliance on commercial endpoints with non-frozen weights also threatens exact reproducibility, since updated model versions could shift the collective dynamics even if the prompts were held constant.
Large language models (LLMs) increasingly operate in multi-agent systems where they collaborate by responding to others' choices, creating endogenous feedback loops with vulnerability to social influence, herding behavior and runaway cascades. Following the design of the classic Music Lab experiment, we tested this vulnerability in large populations of LLM agents that recommended unfamiliar songs, with or without exposure to cumulative recommendation counts. Exposure led to runaway cascades that generated extreme and unpredictable collective recommendations increasingly divergent from a song's intrinsic appeal. Compared to humans, LLM populations were generally far more susceptible to herding behavior, and unlike humans, LLM cascades were less able to self-correct. This vulnerability has implications for multi-agent collaboration in recommender systems, automated evaluation, complex problem-solving, and unauthorized conspiracies.
Pick a starting point or write your own. Challenges run in the background, so you can keep reading while the AI investigates.
No challenges yet. Disagree with the review? Ask the AI to revisit a specific claim.