PaperlayerView exact revision on arXiv

Source-linked AI summary

Emergence of polarization in networks of large language model agents

Jinghua Piao, Zhihong Lu, Chen Gao, Fengli Xu, Qinghua Hu, Fernando P. Santos, Yong Li, James Evans

arXiv:2501.05171v3cs.SIcs.CY

TL;DR

How autonomous LLM agents form collective opinions through interaction remains incompletely understood. The paper simulates thousands of networked agents whose LLM-driven interactions produce polarization and self-organized social networks. Their patterns align with human social phenomena and opinion-dynamics models, while quantitative correspondence to real-world behaviour remains to be validated.

  • Problem

    Evidence remains limited on how autonomous LLM agents develop collective opinions and behaviour through evolving social interactions.

    Source: Introduction · p. 1

  • Method

    The authors simulate thousands of LLM-driven agents that form relationships, communicate, and update political opinions through social interactions.

    Source: Introduction · p. 2

  • Results

    LLM agents spontaneously formed polarized opinions and homophilic networks, with behaviours and collective patterns consistent with human observations and established opinion-dynamics models.

    Source: Introduction · p. 2, Discussion · p. 6

  • Takeaways & Limitations

    The system offers a synthetic testbed for exploring hypothetical interventions in networked LLM-agent systems.

    Source: Introduction · p. 2

    The quantitative correspondence between simulated patterns and real-world human behaviour requires further validation, and prompt-based design reduces interpretability.

    Source: Discussion · p. 7

Abstract

from arXiv · show

Rapid advances in large language models (LLMs) have not only empowered autonomous agents to generate social networks, communicate, and form shared and diverging opinions on political issues, but have also begun to play a growing role in shaping human political deliberation. Our understanding of their collective behaviours and underlying mechanisms remains incomplete, however, posing unexpected risks to human society. In this paper, we simulate networked systems involving thousands of LLM agents across different backbone models (GPT-3.5, GPT-4o, ChatGLM, Llama-3, and DeepSeek-V3), in which agents interact through LLM-guided conversations and update their opinions over time, resulting in the emergence of opinion polarization. We discover that these agents spontaneously develop their own social network with properties characteristic of human social networks, including homophilic clustering. The collective opinions of these LLM agents evolve in ways that exhibit behavioural patterns consistent with social phenomena and mechanisms widely discussed in studies of human behaviour. Overall, these behaviours, patterns, and emergent phenomena produced by LLM agents are consistent with real-world observations and established opinion-dynamics models. This consistency suggests that LLM agents can serve as a valuable synthetic testbed for exploring hypothetical intervention strategies in networked LLM-agent systems.

Introduction

Prior work leaves open how autonomous LLM agents develop collective opinions through large-scale interaction. Here, thousands of LLM agents spontaneously form polarized opinions and homophilic networks, with patterns consistent with human social phenomena and opinion-dynamics models.

Introduction

  • Prior studies largely treated LLMs as isolated counterparts to humans or constrained agents, limiting evidence about autonomous, large-scale collective opinion dynamics.The introduction also identifies the challenge of understanding black-box opinion formation and predicting bottom-up collective behaviour.Source: Introduction · pp. 1–2

Results

  • Thousands of freely interacting LLM agents developed polarized opinions and homophilic networks, with collective patterns consistent with human social phenomena and established opinion-dynamics models.The authors present the system as a synthetic testbed for hypothetical interventions in networked LLM-agent systems.Source: Introduction · p. 2
  • Agents used LLM-driven stages for self-expression, communication, and opinion updating, with minimal initial opinions and network structure rather than predefined behavioural rules.Prompts acted as soft parameters shaping agent behaviour.Source: Opinion dynamics in a synthetic system… · pp. 2–3
  • 40% initial neutrality fell to 22.5%, 0.4%, and 5.1% for partisan alignment, gun control, and abortion ban, respectively, with left-skewed polarization.Self-regulation later produced balanced left- and right-leaning camps while polarization persisted.Source: Opinion dynamics in a synthetic system… · p. 3, Polarization patterns emerge from… · p. 4
  • Right-leaning agents showed more self-inconsistency and opposing transitions; iterative self-checking reduced inconsistency by 9.4%-52.2% and yielded balanced polarization.Opinion updating was particularly prone to self-inconsistency.Source: Polarization patterns emerge from… · pp. 3–4
  • Same-camp interactions rose by 156.8%-382.7%, reaching 48.5%-88.3%; agents formed homophilic clusters and, in one system, two opinion-based communities.The authors report that clustering and avoidance of opposing agents emerged without mechanisms prompting either behaviour.Source: Collective behavioural patterns behind… · p. 4
  • Homophilic exposure to more radical neighbours increased polarization, while opposing exposure could also increase it; these echo-chamber and backfire patterns varied across individuals.Agents’ opinion formation also exhibited patterns consistent with selective exposure and confirmation bias, and increasing either behaviour increased system polarization.Source: Collective behavioural patterns behind… · pp. 4–5
  • In polarized systems, individual-level open-mindedness strategies reduced polarization more than network edits: no confirmation bias reduced it by 11.8% and neutral elite signaling by 8.8%.No selective exposure had limited effects as the existing network crystallized; all tested interventions promoted more inclusive conversations.Source: Synthetic testbed for exploring… · p. 6

Discussion

The discussion frames the findings as evidence that networked LLM-agent societies can reproduce patterns consistent with human behaviour and opinion-dynamics models. It also emphasizes limits to real-world validation and interpretability.

Discussion

  • Across individual and collective levels, LLM agents exhibited behaviours and network patterns consistent with human social mechanisms, including selective exposure, confirmation bias, homophilic clustering, and polarization.The authors report supporting experiments across additional models, topics, and initial conditions.Source: Discussion · p. 6
  • The observed polarization aligns with argument-accumulation models, but such models abstract from affective responses and identity-based dynamics.Agents who became more polarized tended to accumulate more supporting arguments over time.Source: Discussion · p. 6
  • Self-inconsistency can make agents drift from assigned political personas and contribute to left-skewed polarization; self-regulation yielded balanced left-right polarization.The discussion reports only slight differences in polarization speed between original and self-regulated agents.Source: Discussion · p. 7
  • The quantitative correspondence between simulated patterns and real-world human behaviour remains to be validated; prompt-based behaviour and simplified opinions also reduce interpretability.The authors characterize intervention results as hypothesis-generating and call for human experiments and systematic calibration against fine-grained real-world data.Source: Discussion · p. 7

Methods

The methods build LLM agents that express reasons, communicate, and update opinions, then test pairwise opinion bias and define measures of polarization and self-inconsistency. A self-regulation procedure and controlled intervention design support these experiments.

Methods

  • Agents iteratively generate reasons for their opinions, choose whether to maintain social contact or encounter a random partner, exchange persuasive messages, and update opinions from received messages.The study uses near-Gaussian initial opinions and a Watts–Strogatz network; agents self-report opinions on a five-category scale.Source: M1 Design of the networked system of… · pp. 7–8

M2 Pairwise interaction-based bias evaluation experiment

  • In the pairwise bias experiment, agents with identical initial opinions are randomly paired for one persuasive exchange, isolating opinion changes from network structure.The recipient’s opinion change is measured after that single communication round.Source: M2 Pairwise interaction-based bias… · p. 8
  • Agents self-check whether their messages, communication, and opinion updates align with their current opinions and received messages, regenerating outputs until consistent or a retry limit is reached.Agents become inactive in a stage if they reach the maximum number of retries.Source: M3 Self-regulation strategy for the… · p. 8

Methods

  • The intervention experiment modifies a polarized partisan-alignment system from t = 35 to t = 40 and compares treated systems with the original control.Effectiveness is assessed through polarization and changes in homophilic-interaction proportions.Source: M4 Evaluation of hypothetical… · p. 9
  • Polarization is measured as the average distance of opinions from neutral, while self-inconsistency compares observed opinion transitions with an ideal identity matrix.The polarization score ranges from 0 to 2; the inconsistency measure assumes same-opinion communication should not change opinions.Source: M5 Metrics · p. 9

Funding Statement

This block combines funding and competing-interest disclosures with supplementary information and methods for agent interactions, self-regulation, and hypothetical interventions.

Funding Statement

  • The study received support from China’s National Key Research and Development Program and National Natural Science Foundation; funders had no role in the research or publication process.The listed grants are 2024YFC3307603, U23B2030, and 72442026.Source: Funding Statement · p. 15

Competing interests

  • J.E. reports a commercial relationship with Google through a visiting-researcher role, while Google had no role in the study; other authors report no competing interests.The stated Google non-involvement covers study conceptualization, data collection or analysis, manuscript preparation or editing, and publication decision.Source: Competing interests · p. 15

Additional information

  • The manuscript points readers to supplementary information, and Figure 1 depicts the three-stage agent loop and opinion distributions across three political issues.The figure describes experiments with 1,000 agents per issue.Source: Additional information · p. 16, Figure 1 · p. 17

1.1 System Description

  • Agents self-express reasons, choose whom and what to communicate, and update opinions based on messages from socially connected peers.Opinion updating and the next self-expression are combined, so the updated opinion and its supporting reason carry into the next timestep.Source: §1.1 · pp. 23–24

1.2 Pairwise Interaction-based Bias Evaluation Experiment

  • The pairwise experiment limits each agent to one partner with the same opinion to separate inherent agent issues from social influence.This design avoids network structures confounding the examination of opinion changes.Source: §1.2 · p. 24

1.3 Self-regulation Strategy

  • Self-regulation adds a double-check procedure that asks agents to verify whether their generated behaviors align with their current status.Agents regenerate messages, communication, or updated opinions until consistent, subject to a maximum retry number after which they remain inactive in that stage.Source: §1.3 · pp. 24–25

1.4 Intervention Experiments

  • The intervention study tests five hypothetical strategies, with two altering social networks and three adjusting agent traits or behaviors.The strategies are designed to explore polarization reduction in a controlled synthetic testbed.Source: §1.4 · p. 25
  • Network interventions randomize partners or restrict received messages to friends with moderate opposing opinions, while individual interventions target selective exposure, confirmation bias, or elite signaling.Neutral elite signaling uses an external influencer who broadcasts neutral messages to agents each timestep.Source: §1.4 · p. 25

2 Supplementary Note 1: Political Polarization in Human Society

The supplement compares empirical political opinion distributions with simulated agent distributions, while noting that empirical patterns and category boundaries vary across datasets. On a coarse three-camp division, the simulated patterns are similar to human distributions.

Supplementary Note 1: Political Polarization in Human Society

  • The empirical distributions are approximations because public opinion-distribution data are limited and some values were extracted from figures.The supplement describes processing open-source datasets and estimating distributions from published plots.Source: §2 · p. 26
  • Table S1 reports proportions for left, moderate left, neutral, moderate right, and right opinions, with camp difference defined as left-leaning minus right-leaning share.Left-leaning combines left and moderate-left opinions; right-leaning combines right and moderate-right opinions.Source: Table S1 · p. 26
  • Across the empirical datasets, polarizing camps exceed 50% of the population, but their balance against each other and neutral camps varies by source, issue, and collection time.Except for Twitter-Politics, left- and right-leaning camps are relatively balanced; moderate-opinion proportions also depend on category criteria.Source: §2 · p. 26
  • The regulated-agent simulations differ from empirical distributions by an average of 0.21, versus an average difference of 0.45 across empirical datasets (0.49 excluding Twitter-Gun Control and Twitter-Abortion).The comparison uses recent empirical datasets and coarse left-leaning, neutral, and right-leaning camps.Source: §2 · pp. 26–27

3 Supplementary Note 2: Large-scale Simulation of LLM Agents

A 2,000-agent simulation preserves the reported collective behaviors and emergent patterns seen at smaller scale, including polarization and homophilic clustering.

Supplementary Note 2: Large-scale Simulation of LLM Agents

  • At 2,000 agents, system scale does not change the reported collective or emergent behaviors.The supplement reports this comparison in Figure S2.Source: §3 · p. 27
  • Free-form social interactions still produce opinion polarization in the 2,000-agent system.This pattern is reported in Figure S2a-c.Source: §3 · p. 27
  • Agents still form a social network in which agents with similar opinions cluster and agents with opposing opinions avoid mutual interaction.The reported homophilic clustering occurs in the large-scale simulation.Source: §3 · p. 27

4 Supplementary Note 3: Self-inconsistency Problem in LLM Agents

Pairwise experiments reveal a left-leaning tendency in agents’ opinion transitions, which the proposed self-regulation strategy largely suppresses.

4 Supplementary Note 3: Self-inconsistency Problem in LLM Agents

  • Across most issues, agents more often shift into the left-leaning camp, and even right-camp-only experiments sometimes produce leftward shifts.The authors describe this as an inherent lean-learning tendency, also reported in prior LLM studies.Source: §4 · p. 27, Figure S3 · p. 29, Figure S4 · p. 30
  • The abortion-ban simulation transitions from consensus to polarization while homophilic agents form communities.The figure depicts a system of 2,000 self-regulated LLM agents.Source: Figure S2 · p. 28
  • Self-regulation largely suppresses agents’ tendency to adopt left-leaning opinions in pairwise interaction experiments.The experiments include partisan alignment, abortion ban, gun control, Obamacare, and education reform.Source: §4 · pp. 27–28, Figure S3 · p. 29, Figure S4 · p. 30

1 Self-expression

Self-inconsistency occurs across the agents’ three stages, with opinion updating the most error-prone and retry-intensive stage. Polarization-speed patterns remain highly consistent between the original and self-regulated systems.

1 Self-expression

  • Self-inconsistency errors occur in self-expression, communication, and opinion updating.The examples are presented in Figure S5 and Table S2.Source: Figure S5 · p. 31, Table S2 · p. 32

2 Communication

  • Opinion updating has the highest self-inconsistency frequency, above 0.6 across five issues, and averages more than 4 retries for partisan alignment.The authors identify opinion updating as the most difficult task in the overall system.Source: §2 · p. 33, Figure S6 · p. 34
  • Polarization-speed patterns remain highly consistent between original and self-regulated systems despite minor issue-specific differences.For partisan alignment, growth is slightly faster in the original system; for gun control, marginally faster in the debiased system; abortion-ban results vary over time.Source: §2 · p. 35, Figure S7 · p. 35

5 Supplementary Note 4: Observations on LLM agents’ Self-organized Social Network

LLM agents’ self-organized networks become more homophilic, while echo-chamber and backfire-like patterns occur selectively. Network structure also matters: static and random alternatives do not produce balanced polarization.

5 Supplementary Note 4: Observations on LLM agents’ Self-organized Social Network

  • Modularity increases in all three systems, while assortativity and homophily rise over time, indicating clustering among like-minded agents and avoidance of opponents.Gun-control and abortion-ban networks reach modularity scores above 0.3, indicating a clear division into two communities.Source: §5 · p. 35, Figure S8 · p. 36, Figure S9 · p. 37
  • Agents are more likely to interact with peers in the same political camp, a pattern consistent with echo chambers.The analysis defines echo chambers using the joint distribution of an agent’s opinion and the average opinion of its neighborhood, following Cinelli et al..Source: §5 · p. 35, Figure S10 · p. 38
  • Opposing interactions do not consistently reduce polarization and can occasionally intensify it; backfire-consistent effects occur in 0.82%–2.12% of agents.The authors note that agents are rarely in environments dominated by opposing opinions.Source: §5 · pp. 36, 38, Figure S11 · p. 38
  • Only 29.1% of agents receive more than 75% of interactions from like-minded sources; the median share is 48.2%.The authors report that echo-chamber-consistent patterns do not occur universally, but emerge selectively among agents.Source: §5 · pp. 36–37, Figure S12 · p. 38
  • The network’s in-degree distribution develops a power-law tail, and the top 20 nodes average 9.38 in-degree, 2.35 times the overall average.Of these top 20 nodes, 86.7% are left-leaning.Source: §5 · pp. 38–39, Figure S13 · p. 39
  • With static or random networks, no balanced polarization pattern forms and homophilic opinions dominate 75% of agents.The authors identify self-organized social networks as essential to polarization in their system.Source: §5 · p. 40, Figure S14 · p. 40

6 Supplementary Note 5: Results of Other Individual-level Behavioural Patterns

Simulations with increased prevalence of three specified behavioral patterns show higher polarization.

6 Supplementary Note 5: Results of Other Individual-level Behavioural Patterns

  • A higher proportion of agents prompted to exhibit exaggerated misperception, objective illusion, or stereotyping increases polarization.The simulations explicitly configured 50% of agents to exhibit these patterns, with other experimental settings unchanged.Source: §6 · p. 40

7 Supplementary Note 6: Experiments on Different LLMs

Across LLM backbones, agent interactions produce polarization and homophilic clustering, while interventions reduce polarization with varying effectiveness. Switching models mid-simulation does little to reverse established dynamics; intervention outcomes depend on strategy and network change.

7 Supplementary Note 6: Experiments on Different LLMs

  • Free-form interactions produce polarization across tested LLM systems; Llama-3 reverses the left-skewed tendency seen in other models, while Claude-3 was excluded after a 10% response rate.The tested backbones included GPT-3.5 Turbo, GPT-4o, Llama-3, and ChatGLM.Source: §7 · p. 40, Figure S16 · p. 42, Figure S17 · p. 43
  • Systems with more agents exhibiting exaggerated misperceptions, objective illusion, or stereotyping show increased polarization.The figure reports mean polarization for the last five timesteps with corresponding 95% confidence intervals.Source: Figure S15 · p. 41
  • Mid-simulation switches from GPT-3.5 to GPT-4o or DeepSeek-V3 leave polarization and opposing-view communities largely intact, consistent with path dependence after networks and opinions have formed.The authors attribute this persistence to accumulated interactions and social context reinforcing polarizing tendencies.Source: §7 · pp. 41, 43, Figure S20 · p. 45, Figure S21 · p. 45

8 Supplementary Note 7: Analyses on Different Intervention Strategies

  • Neutral elite signaling and no selective exposure reduce polarization most, while network interventions are less effective in the synthetic system.All tested strategies gradually reduce polarization.Source: §8 · p. 43, Figure S22 · p. 46, Figure S23 · p. 46
  • Interventions alter opinion distributions and make agents more likely to moderate, with neutral signaling increasing neutral opinions and no confirmation bias increasing moderate opinions.Random interaction reduces polarization but affects only the right camp in this system.Source: §8 · p. 44, Figure S24 · p. 47, Figure S25 · p. 48
  • All strategies reduce homophilic interactions, but no selective exposure changes the network least; its network remains stable and dominated by homophilic peers.Network-level interventions have the greatest structural impact, while individual-level interventions gradually shape the network.Source: §8 · p. 46, Figure S26 · p. 48, Figure S27 · p. 49
  • Open-mindedness alone yields a negligible, nonsignificant polarization reduction of about 1%, while adding opposing-camp exposure significantly reduces polarization versus the original and oppose-only conditions.The combined condition differs from the original (t = 4.18, p = .0019) and oppose-only (t = 2.32, p = .042); 95% of agents accepted the fictitious link between open-mindedness and life success.Source: §8 · p. 47, Figure S28 · p. 49

9 Supplementary Note 8: Analyses on Agents’ Social Perceptions and Argument Accumulation Process

Agents favor impressions of like-minded peers over opponents, and their descriptions and reported opinion-update processes show patterns associated with polarization. As opinions polarize, agents accumulate more supporting arguments across the examined topics.

9 Supplementary Note 8: Analyses on Agents’ Social Perceptions and Argument Accumulation Process

  • Agents give the best impressions to similar same-camp peers and the lowest impressions to opposing-camp agents across the examined issues.Ratings were collected for similar, opposing, and neutral agents.Source: §9 · p. 50, Figure S29 · p. 50, Figure S30 · p. 51
  • Descriptions are generally positive, but agents use negative terms such as “conflicting,” “rigid,” and “opinionated” for opponents.The reported sentiment distributions cover partisan alignment, abortion ban, and gun control.Source: §9 · p. 50, Figure S33 · p. 53, Figure S34 · p. 54
  • Opinion updates were not hard-coded to psychological mechanisms, yet self-reports frequently reflected confirmation bias, motivated reasoning, and moral conviction.The authors describe these patterns as closely mirroring mechanisms in human social psychology.Source: §9 · p. 51, Table S3 · p. 59
  • Agents who become more polarized accumulate significantly more supporting arguments across all three topics, with argument themes varying by issue.The authors report this as additional evidence consistent with argument-accumulation models.Source: §9 · pp. 51, 53, Figure S32 · p. 52
  • Among less-polarized agents, the listed processes included cognitive dissonance reduction (81%), value-based reasoning (32%), and motivated reasoning (22%), concerning conflicting beliefs, core values, and pre-existing motivations, respectively.Source: §9 · p. 60

10 Supplementary Note 9: Experiments on LLMs with Varying Temperatures

Polarization emerges across tested temperature settings, but lower temperature reduces its level and homophilic clustering. At 0.5, agents change communication partners less often, a pattern the authors connect to lower polarization.

10 Supplementary Note 9: Experiments on LLMs with Varying Temperatures

  • Lower temperature markedly reduces polarization and homophilic interactions, whereas higher temperature has no noticeable effect on these measures.The simulations also show polarization emerging at both lower and higher temperatures.Source: §10 · pp. 60, 63, Figure S39 · p. 61, Figure S40 · p. 62
  • At temperature 0.5, agents change only 1.32% of their edges per timestep on average, less than half the rate at 1.0 or 1.5.The authors associate maintaining initial relationships rather than approaching more homophilic peers with reduced polarization; the lower-temperature reduction in opinion updating was not statistically significant.Source: Figure S41 · p. 62, §10 · p. 63

11 Supplementary Note 10: Experiments on LLMs with Enhanced Prompting

Seeding 10% of agents with chain-of-thought prompting is associated with more moderate or neutral opinions, lower polarization, and weaker network division. Enhanced prompting can also model subgroup profiles with distinct behavioral patterns.

11 Supplementary Note 10: Experiments on LLMs with Enhanced Prompting

  • Seeding 10% of agents with chain-of-thought prompting raises the moderate-or-neutral share from 8% to 38%, lowers polarization by 15%, and reduces modularity from 0.43 to 0.25.The seeded agents were randomly distributed at simulation start; the authors present this as a synthetic-testbed result.Source: §11 · p. 63, Figure S43 · p. 64
  • Enhanced prompting with subgroup-inspired profiles lets LLM agents simulate two U.S. subgroups and capture distinctive behavioral patterns, while main experiments omit subgroup-specific prompt information.The authors describe this as an extension illustrating system flexibility.Source: §11 · p. 64

12 Supplementary Note 11: Experiments on the Effect of Increasing Severing Costs

Across altered tie costs, initial conditions, and issue domains, LLM-agent systems still develop polarized communities on divisive topics, while agents converge on a fact-based issue. Small severing costs substantially reduce silo formation, and polarization persists across varied starts.

12 Supplementary Note 11: Experiments on the Effect of Increasing Severing Costs

  • A 10% increase in severing cost marginally mitigates polarization but substantially suppresses ideological silo formation across four network-level metrics.The metrics are modularity, assortativity, homophilic interactions, and homophily index.Source: §12 · p. 64

13 Supplementary Note 12: Results of Simulation under Different Initial Conditions

  • Starting with 80% neutral agents delays but does not prevent polarization and the gradual split into two opposing communities.The initial opinion distribution is [0,0.1,0.8,0.1,0].Source: §13 · p. 65
  • Despite substantial differences in initial network structure, agents form two opinion-based communities and converge to stable polarization across the tested network models.The tested starts include Erdős–Rényi, Barabási–Albert, and Watts–Strogatz networks.Source: §13 · p. 65

14 Supplementary Note 13: Results of Simulation on Different Issues

  • On immigration restrictions, agents also develop polarized collective opinions, supporting use of the system to explore opinion-dynamics hypotheses across issue domains.Source: §14 · p. 65
  • On flat Earth theory, agents reach consensus within ten timesteps rather than polarizing, strongly opposing the theory.The authors present this contrast as evidence that the system is not inherently predisposed to generate polarization.Source: §14 · p. 65

15 Supplementary Note 14: Supplementary Notes on Statistical Testing and Robustness Checks

The authors checked assumptions underlying frequentist tests and used parametric tests when assumptions permitted. They also report non-parametric robustness checks where appropriate.

15 Supplementary Note 14: Supplementary Notes on Statistical Testing and Robustness Checks

  • Normality and equal variances were assessed with Shapiro–Wilk and Levene’s tests for two-sample comparisons.Source: §15 · p. 65
  • Student’s t-tests were used when equal-variance assumptions held; otherwise, the authors used Welch’s t-tests.Source: §15 · p. 65
  • Where appropriate, robustness checks used Mann–Whitney U or Wilcoxon signed-rank tests; large-sample and proportion tests relied on the central limit theorem.Source: §15 · p. 65

Related papers

Loading 2501.05171v3…