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.
Figures & tables
Figure 1: Experimental design. Each of 13 model populations evaluated the same 48 unfamiliar songs, with agents assigned personas combining demographic attributes and preferred music genres. For each model, one independent world contained 1,440 agents, and eight worlds in each influence condition contained 720 agents each. Titles were randomly ordered without recommendation counts in the independent condition, randomly ordered with counts under weak influence, and ordered by decreasing counts under strong influence. Walls indicate the absence of social feedback; dashed and solid arrows indicate feedback under weak and strong influence, respectively. The bottom sequence illustrates one agent’s turn in the strong-influence condition. Agents first selected four titles, then read the lyrics and decided whether to recommend each song. In the influence conditions, cumulative counts were visible in both the selection and recommendation stages and never shown in the independent condition. Each recommendation increased the song’s count by one, updating the information available to subsequent agents in the same world. All worlds began with zero counts and evolved separately.
Figure 2: Compared with humans, LLM populations showed greater divergence from independent judgment. Kullback–Leibler (KL) divergence measures departure of influence-world market shares from the corresponding independent baseline, averaged over eight worlds per condition. Higher values on the logarithmic vertical axis indicate greater divergence. Divergence was computed with Jeffreys smoothing (pseudocount 0.5 per song). Gold and purple denote strong and weak influence, respectively. Hatched bars show human values from the original Music Lab data ( Salganik et al., 2006 ) . (A) Mean divergence was greater among non-reasoning models than among reasoning models, and greater in both classes than among humans. (B and C) . GPT-5 nearly matched the human benchmark under strong influence; Gemma 3 4B had the greatest divergence under strong influence and Gemma 3 27B under weak influence. Differences between independent and social influence conditions were significant for all 13 LLMs under both influence conditions ( p<0.001 ).
Figure 3: Social influence increased unpredictability, especially among non-reasoning models. As in the original Music Lab experiment ( Salganik et al., 2006 ) , unpredictability (U) measures the mean absolute difference in a song’s market share across all 28 pairs of eight influence worlds, averaged over the 48 songs. Unpredictability in the independent world was estimated by randomly splitting the population into two equal groups of 720 agents and averaging over 1,000 random splits. Gold, purple, and blue denote strong influence, weak influence, and independent evaluation, respectively. The hatched bar is the human benchmark for comparison with unweighted class means among the five reasoning and eight non-reasoning models. (A) Unpredictability was greater among non-reasoning models compared to reasoning models and humans, regardless of influence level. (B and C) Model-specific results show that unpredictability in the independent condition was near zero for each model and only slightly higher among humans, indicating stable aggregate shares under independnet evaluation. Social influence increased unpredictability in every model, although some reasoning models fell slightly below the human benchmark. In every population, unpredictability was greater in both influence conditions than in the independent condition ( p<0.001 ), except for humans ( p<0.01 ) and DeepSeek-R1 under weak influence ( p<0.05 ).
Figure 4: Higher model intelligence is associated with lower divergence and relative unpredictability. Each point represents one of 13 LLMs plotted against its Artificial Analysis Intelligence Index score, version 4.1.1 ( Artificial Analysis, 2026 ) . The index aggregates performance across benchmark agentic tasks, coding, general problem-solving, and scientific reasoning; higher scores indicate higher intelligence. The associations are summarized as a least-squares linear fit and Spearman correlation. Stars denote reasoning models and dots denote non-reasoning. The dashed line is the human benchmark. Within each influence condition, KL divergence is the mean over eight influence worlds and unpredictability is the mean over 28 pairs of influence worlds, normalized by unpredictability in the independent condition ( Uinf/Uind ). The association with divergence depends on the strength of social influence, going from ρ=−0.851 ( p<0.001 ) when influence is strong (A) to ρ=−0.551 ( p=0.051 ) when influence is weak (B) . The association with unpredictability (C and D) remains strong regardless of influence condition ( ρ=−0.733 , p<0.01 and ρ=−0.719 , p<0.01 for strong and weak influence, respectively). The models are clustered by reasoning class, indicating the expected association with intelligence score.
Figure 5: Human cascades consistently self-corrected; LLM cascades generally did not. Following van de Rijt ( Van de Rijt, 2022 ) , self-correction was measured as the change in Spearman correlation between popularity and intrinsic appeal between the first and second halves of the sequential recommendations. The change in correlation was calculated for each of the eight influence worlds in each influence condition (indicated by a dot) and then averaged over the eight worlds (indicated by a star). A higher second-half correlation indicates closer alignment with the independent ranking; negative changes indicate movement of popularity away from intrinsic appeal. Shaded boxes and lines show the interquartile range and total range respectively. Human correlations increased in seven of eight strong-influence worlds and all eight weak-influence worlds, with mean changes of +0.085 ( p<0.05 ) and +0.140 ( p<0.01 ), respectively. In contrast, LLM changes varied in direction across models and worlds and averaged near zero under strong influence. As expected, self-correction requires intrinsic appeal that is strong relative to social influence.
Appendix figures & tables12 assets
Supplementary material from the paper’s appendix.
Appendix
Title
Words
Title
Words
Map Beneath the Stars
172
Briefly Burning
156
Voice Like Silk
158
Electric Soulmates
155
Neon Garden
137
Concrete Undertow
145
Uncaught Thoughts
165
Plastic Moons
150
Where the Day Gives Up
158
Clouds in My Tea
137
Drift Until It Ends
144
Stealing Stars
162
Appendix
Table S1: Song stimuli and lyric length. Word counts were computed using whitespace-separated tokens, including section markers.
Figure S1: Prompts used in the strong-influence condition. The upper blue box contains the prompt for the selection stage and the lower blue box contains the prompt for choosing songs to recommend. The yellow boxes show illustrative responses. Braced fields identify persona, song, and recommendation-count inputs. Counts reflect the cumulative number of recommendations before the agent’s decisions and were displayed in both stages. The weak-influence condition used the same templates except that the clause describing popularity-based ordering was omitted and title order was randomized independently for each agent.
Figure S2: Prompts used in the independent condition. The upper and lower blue boxes contain selection- and evaluation-stage prompt templates, with illustrative responses in the yellow boxes. Persona framing, task structure, and response format matched the influence conditions, but recommendation counts and references to earlier choices were omitted. Title order was randomized independently for each agent.
Influence vs. independent
First 30 vs. later agents
Population
Weak
Strong
Weak
Strong
Human data
0.740
0.557
0.316
0.723
GPT-5 nano
0.533
0.389
0.844
0.963
GPT-5
0.885
0.815
0.823
0.889
Claude Sonnet 4.6
0.741
0.779
0.947
0.858
Gemini 3 Flash
0.860
0.831
0.889
0.857
Appendix
Table S2: Correspondence of popularity with intrinsic appeal and with early movers in the social influence conditions. The first two columns report Pearson correlations between song market shares under social influence and in the independent condition, pooling the 48 songs across the eight influence worlds (384 paired song–world observations per population and condition). Higher values indicate greater alignment with intrinsic appeal. The last two columns report Pearson correlations between market shares among the first 30 agents (participants for humans) in each influence world and market shares among all later agents in the same world, pooled in the same way; higher values indicate that later agents reproduced the ranking established by early movers. Class means are arithmetic averages of model-level coefficients.
Model
Developer
Class
Intelligence Index
GPT-5
OpenAI
Reasoning
35
GPT-5 nano
OpenAI
Reasoning
19
Gemini 3 Flash
Google DeepMind
Reasoning
39
DeepSeek-R1
DeepSeek
Reasoning
19
Claude Sonnet 4.6
Anthropic
Reasoning
48
GPT-4o
OpenAI
Non-reasoning
11
Appendix
Table S3: AAII scores by reasoning class. Models were classified by developer designation and the inference mode used in the experiments. Artificial Analysis Intelligence Index v. 4.1.1 scores were used in the analyses [ Artificial Analysis, 2026 ] . Higher scores indicate greater benchmarked intelligence.
Figure S3: Social influence generally increased market-share inequality. Inequality was measured by the Gini coefficient of market shares across the 48 songs; larger values indicate greater concentration of popularity among fewer songs. Gold, purple, and blue denote strong influence, weak influence, and independent evaluation, respectively. The gold and purple bars report the Gini coefficient averaged over the eight worlds in each influence condition. Blue bars show the Gini coefficient averaged over 1,000 bootstrap samples drawn from the independent world. Each sample was the same size as an influence world (720 agents for LLMs). Hatched bars show human values from the original Music Lab data [ Salganik et al., 2006 ] for comparison with unweighted class means among the five reasoning and eight non-reasoning models. A. On average, both model classes were more unequal than humans in the independent condition. Relative to their independent baselines, reasoning models showed smaller mean influence-driven increases than humans, whereas non-reasoning models showed larger increases under both influence conditions. B and C. Inequality exceeded the independent baseline in all 13 models in the strong-influence condition and in 12 models under weak influence, with GPT-5 nano the lone exception.
Figure S4: Position bias among LLMs and humans . Gold, purple, and blue denote the strong influence, weak influence, and independent conditions, respectively. The human data show the corresponding distributions under strong influence (orange) and independent evaluation (green). Bars show the fraction of the first-stage selections made at each of the 48 display positions and solid lines show the smoothed distribution. The dashed line indicates uniform selection (equal fractions at each position). Position bias was measured in the independent condition using the fraction of first-stage selections made at each of the 48 display positions (not songs). For most models, the distribution approaches uniformity in the randomly ordered independent and weak-influence conditions, indicating a low level of position bias. Only Gemma 3 4B and GPT-4o mini showed a modest preference for early positions, and even here, the distributions are closer to uniformity than to the strong-influence distributions. The figure also reports stronger position bias in the independent condition among humans than among AI agents.
Figure S5: Dynamics of KL divergence as recommendations accumulated. The plots display KL divergence from the corresponding independent distribution over the course of the experiment for the 13 LLM populations and the human benchmark, under weak and strong influence. Each line shows one of the eight worlds in an influence condition. The horizontal axis is the number of participants (humans or agents) who had completed the task; the vertical axis is logarithmic. Market shares at each stage were calculated from accumulated recommendations for LLMs and downloads for humans. Gold and purple denote strong and weak influence for LLMs; orange and green denote strong and weak influence for humans. Final divergence values correspond to those reported in fig. 2 , where they are averaged over the eight worlds in each influence condition. Divergence persisted over the observed trajectories, although its development varied across models and worlds.
Figure S6: Dynamics of unpredictability as recommendations accumulated. The plots report unpredictability over the course of the experiment for the 13 LLM populations and the human benchmark, under weak and strong influence. Unpredictability measures the mean absolute difference in a song’s market share across all 28 pairs of the eight worlds in an influence condition, averaged over the 48 songs, with market shares calculated from the recommendations (downloads for humans) accumulated at that stage. The horizontal axis is the number of participants (agents or humans) who had completed the task. Gold and purple denote strong and weak influence for LLMs; orange and green denote strong and weak influence for humans. Final unpredictability values correspond to those reported in fig. 3 . Unpredictability generally stabilized or declined over time, with strong influence producing greater final unpredictability than weak influence in most models.
Figure S7: Effects of social influence were robust to a lower sampling temperature. The figure reports KL divergence (left) and unpredictability (right) for the eight non-reasoning models under strong influence at temperatures 1.0 (solid bars) and 0.7 (hatched bars). KL divergence compares each influence world with the independent baseline generated at the same temperature, averaged over the eight influence worlds; the vertical axis is logarithmic. Unpredictability is shown for both influence (gold) and independent (blue) conditions; influence values are the mean absolute difference in a song’s market share across all 28 pairs of the eight influence worlds, averaged over the 48 songs. Outcomes in the independent condition were estimated from 1,000 random splits into two equal subpopulations. Lowering temperature did not attenuate the effects of social influence on average, and unpredictability in the independent condition hardly changed. Averaged over the eight models, KL divergence was 1.62 at temperature 1.0 and 1.75 at temperature 0.7, with greater divergence at the lower temperature in seven of the eight models. Unpredictability in the influence condition averaged 0.0292 and 0.0306, respectively, compared with 0.00288 and 0.00284 in the independent condition.
Figure S8: KL divergence was robust to an increase in the number of songs that could be selected for assessment. The figure reports KL divergence between strong-influence worlds and their matched independent baselines when agents selected four (gold) or 12 (brown) of the 48 songs for evaluation. Each model’s divergence is averaged over eight influence worlds. In panel A, bars indicate the unweighted means across the five reasoning and eight non-reasoning models. The vertical axis is logarithmic. Although the larger selection budget reduced divergence in all 13 models, the ratio of mean divergence in non-reasoning to reasoning models was nearly unchanged, as shown by the similar vertical distances between the two gold bars and between the two brown bars in panel A.
Figure S9: Unpredictability was robust to an increase in the number of songs that could be selected for assessment. The figure reports unpredictability under strong influence and independent evaluation when the number of songs agents could evaluate was increased from four to 12 out of 48. Unpredictability in the influence condition is the average absolute difference in market shares for the 48 songs across all 28 pairs of eight worlds. Unpredictability in the independent condition is the average of 1,000 random splits into two groups of 720 agents. A. Bars show unweighted means across the five reasoning and eight non-reasoning models. Gold and brown denote strong influence with four and 12 selections, respectively; blue and green denote the corresponding independent conditions. Increasing the selection budget reduced unpredictability on average across the 13 models (panel A), but every influence population remained more unpredictable than every independent population at both selection sizes.
Large language model (LLM) agents are increasingly deployed in interacting populations, raising the question of what such populations come to believe collectively. Whether a population aggregates genuine knowledge or collapses into a false consensus directly affects how much such systems can be trusted. Classical social-network models assume that the network itself determines how beliefs combine. This assumption breaks down for LLM agents, whose limited attention takes in only part of what they are exposed to, so these models overstate how much information a population actually pools and cannot tell genuine consensus from herding. We introduce SNLA, a framework that models how much each agent actually influences others, rather than merely how the network connects them. This influence depends on each agent's position in the network and on how sharply attention focuses. Theoretically, we show on a tractable proxy that narrow attention causes herding, where the effective sample size stays bounded regardless of population size, while wide attention recovers wisdom-of-crowds behavior only when the exposure graph is undirected and degree-regular. Empirically, a controlled testbed validates these predictions directly, and the herding-wisdom transition reproduces on operator-controlled variants of three multi-agent LLM benchmarks.
Kaixuan Liu, Guojun Xiong, Weinan Zhang +1
Department of Computer Science Emory University Atlanta, GA 30322 · School of Computer Science Shanghai Jiao Tong University Shanghai, China
Large Language Model-based Multi-Agent Systems (LLM-MAS) are increasingly applied to complex collaborative scenarios. However, their collaborative mechanisms may cause minor inaccuracies to gradually solidify into system-level false consensus through iteration. Such risks are difficult to trace since errors can propagate and amplify through message dependencies. Existing protections often rely on single-agent validation or require modifications to the collaboration architecture, which can weaken effective information flow and may not align with natural collaboration processes in real tasks. To address this, we propose a propagation dynamics model tailored for LLM-MAS that abstracts collaboration as a directed dependency graph and provides an early-stage risk criterion to characterize amplification risk. Through experiments on six mainstream frameworks, we identify three vulnerability classes: cascade amplification, topological sensitivity, and consensus inertia. We further instantiate an attack where injecting just a single atomic error seed leads to widespread failure. In response, we introduce a genealogy-graph-based governance layer, implemented as a message-layer plugin, that suppresses both endogenous and exogenous error amplification without altering the collaboration architecture. Experiments show that this approach prevents final infection in at least 89% of runs across operating modes and significantly mitigates the cascading spread of minor errors.
Yizhe Xie, Congcong Zhu, Xinyue Zhang +5
Faculty of Data Science, City University of Macau · Minzu University of China
LLM-powered AI assistants acting on behalf of users can produce poor collective outcomes at scale. We introduce a framework for evaluating their emergent behaviour in social dilemmas, applied to three iterated games (Public Goods, Collective Risk, Common Pool Resource). We prompt each model to produce a natural-language strategy, then have the same model translate it into code. This aims to isolate strategic reasoning from input-parsing, enables pre-deployment inspection, and scales to populations of hundreds of agents. We propose three analyses: behavioural fingerprinting via exhaustive evaluation over opponent histories; self-play robustness across mixtures of a model's strategies with either a Selfish or Collective disposition; and cultural evolution under payoff-biased imitation. Applied to three state-of-the-art LLMs, we find substantial cross-model differences in self-play welfare, and that cultural evolution converges to low-welfare, Selfish-dominant equilibria in larger groups.
Richard Willis, Jianing Zhao, Yali Du +1
King’s College London London, United Kingdom · King’s College London · Turing Institute +1