Pith. sign in

REVIEW 5 major objections 7 minor 2 cited by

Simulating Rumor Spreading in Social Networks using LLM Agents

T0 review · 5 major / 7 minor · reviewed 2026-08-09 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper claims that LLM-driven agents with personas and post feeds can simulate rumor spreading in social networks, with spread ranging from no agents affected to 83% of agents, depending on network structure, seeding, and personality.

desk verdict A clean, runnable LLM-agent rumor simulator with released code, but the quantitative network comparisons rest on single unreplicated runs and need error bars before the claims hold. read the letter →

arxiv 2502.01450 v1 pith:AO6GPWCM submitted 2025-02-03 cs.SI cs.AI

Tianrui Hu , Dimitrios Liakopoulos , Xiwen Wei , Radu Marculescu , Neeraja J. Yadwadkar This is my paper · ORCID
classification cs.SIcs.AI
keywords rumorspreadingLLMagentsagent-basedmodelingsocialnetworksmisinformationnetworkstructurepersonassimulation
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish that large language models can stand in for human users in agent-based simulations of rumor spreading, and that the simulated dynamics respond to the same factors that shape real misinformation cascades. The authors build social networks of more than one hundred ChatGPT-4o-mini agents, each with a persona, a post history, and a list of rumors, and run iterative rounds in which selected agents post and update their belief in each rumor. Across four network types and multiple seeding and activation schemes, the reported spread ranges from no agents affected to 83% of agents affected. If correct, this gives social-network researchers a way to study rumors with far more behavioral texture than equation-based models, and a testbed for interventions before they are tried in the wild.

What carries the argument

The central object is the LLM-driven agent: one ChatGPT-4o-mini call per network node, with a prompt that bundles a persona (name, age, job, traits, and explicit willingness to accept and forward rumors), the posts visible to that agent, and the complete list of rumors under test. Each agent response contains a new post and an updated belief $\langle b_{ij}\rangle \in [0,1]$ for each rumor $r_j$. The process runs on an iterative loop: at each time step one agent is selected randomly or by degree, it posts, the post is appended to its own history and to the histories of all neighbors, and the agent's beliefs are refreshed. This loop is what carries the argument because it couples network topology (who sees what) with individual psychology (who believes and forwards what) in a single simulation.

What would settle it

Run one rumor through one network configuration 50 times with different random seeds and nonzero temperature; if the maximum percentage of affected agents swings from near zero to above 80% across runs, the reported network, persona, and seeding effects are not stable enough to support the paper's conclusions.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that LLM-driven agents reproduce differentiated rumor dynamics: network structure, agent personas, and spreading schemes each change whether and how far a rumor travels. The Small-World network, with sparse connectivity and moderate clustering, is the most susceptible, reaching up to 50% of nodes, while the denser and more clustered Facebook-derived network suppresses spread. Degree-based seeding and degree-based activation spread every tested rumor, whereas random strategies leave most rumors confined to a few agents. Agents also behave like flexible readers rather than fixed carriers: some who initially accept a rumor later reject it as they see new posts, and rumors that are easy to disprove from pretrained knowledge are dismissed while less familiar health and nature rumors propagate. The headline number is the range of outcomes, from no spread to 83% of agents affected, showing that the framework is sensitive to the variables it claims to model.

Load-bearing premise

Everything rests on the premise that ChatGPT-4o-mini agents, given a persona and a post feed, accept and forward rumors the way real people would, a premise supported here by a single 500-iteration run with no replication and no comparison to human rumor-spreading data.

Editorial extensions

If this is right

  • The framework runs at practical scale: more than one hundred agents and thousands of edges, so LLM-agent rumor simulations are not limited to toy graphs.
  • Network topology alone can change a rumor's fate: the sparse Small-World network spreads rumors to more agents than the dense, highly clustered Facebook-derived network.
  • Seeding and activity rules are decisive: rumors that start at high-degree agents and are posted by high-degree agents spread across the whole network, while random seeding and activation leave most rumors localized.
  • Agent personality shifts the outcome monotonically: moving all agents from 'likely to accept' to 'unlikely to accept' produces a clear decline in the maximum percentage of nodes affected.
  • Rumor content matters: agents tend to dismiss rumors they can disprove from pretrained knowledge and accept unfamiliar health and nature rumors, and some agents reverse their belief over time.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A direct comparison against classical threshold or compartmental rumor models on the same networks would show whether the LLM agents add predictive value beyond a parameterized curve.
  • Because the model's pretraining data is entangled with rumor content, running the framework on fictional, time-stamped rumors would isolate the effects of network structure and persona from the model's prior knowledge.
  • The same machinery is an intervention testbed the paper does not fully exploit: injecting fact-checker agents, deleting high-degree nodes, or rewiring the graph mid-simulation would show how the 83% ceiling moves under mitigation.
  • Re-running each configuration across many seeds and temperatures would reveal how much of the reported spread differences is stable signal versus sample noise.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 7 minor

Summary. The paper presents a framework in which each node of a social network is an LLM-driven agent (ChatGPT-4o-mini) with a persona, a post history, and a stated tendency to accept and forward rumors, and simulates rumor propagation over four network types: Erdős-Rényi, Scale-Free, Small-World, and a real Facebook network. Each simulation runs for 500 iterations, with agents selected randomly or by degree for posting; the agent's belief in each rumor is updated by the LLM. Three experiments investigate the effect of network structure, of initialization/activation strategy, and of agent personas on the maximum percentage of affected nodes. The paper claims that network structure, personas, and spreading schemes significantly influence rumor dissemination, with outcomes ranging from no spread to 83% of agents affected, and concludes that the framework provides a realistic simulation of rumor spread.

Significance. If the central claims were established, the framework would be a useful tool for generating hypotheses about rumor dynamics and for demonstrating LLM-based agent-based modeling at nontrivial network scale. Creditable strengths include that the code is publicly released, the framework is described precisely enough to reimplement, and no fitted parameters are used to produce the network-level differences. However, as presented, the quantitative conclusions are not supported by the evidence: every experimental condition is a single unreplicated stochastic run, no variance or significance information is reported, and the abstract's 83% figure is not reconciled with the 50% maximum reported for the network-structure experiment. The persona experiment largely confirms the manipulation built into the prompt rather than discovering an emergent property of personas. The 'realistic simulation' claim is also not validated against human rumor-spreading data or against classical rumor models. The contribution is therefore more a demonstration of feasibility than a validated finding about rumor dynamics.

major comments (5)
  1. [Experiments / Effect of Network Structure] The paper's central quantitative claim is not supported by the reported evidence because all reported conditions are single 500-iteration runs. Algorithm 1 selects agents stochastically (fully random or degree-proportional activation), and ChatGPT-4o-mini responses are sampled without any reported temperature, seed, or replication. Figures 2 and 4 each plot the maximum percentage of affected nodes from one run per condition, so the observed ordering across networks (e.g., Small-World above Erdős-Rényi) could be due to run-to-run LLM variance or the particular activation sequence rather than network topology. To support the abstract's statement that network structure, personas, and spreading schemes 'can significantly influence' rumor dissemination, the authors should provide multiple replications per condition, report means and confidence intervals, and perform significance tests for the reported orderings.
  2. [Abstract / Figure 2 and Figure 4] The abstract's headline range 'from no spread to affecting 83% of agents' is not reconciled with Figure 2, where the Small-World network's maximum is stated as 50% of nodes affected. The 83% figure appears to come from the initialization/activation experiment in Figure 4, but the text does not explicitly connect the abstract's number to any specific condition or table value, and no variance or replication context is given. As written, a reader cannot determine whether the 83% outcome is a single run, a stable finding, or an outlier. The abstract and the experimental reporting should be aligned with explicit quantitative reporting, or the headline claim should be tempered.
  3. [Appendix / Effect of Agent's Personas] The persona experiment is described as showing that agents' personality configurations 'significantly influence' rumor spread, but the manipulation is built directly into the prompt: the values 'agent rumors acc' and 'agent rumors spread' (1-4) are set by the experimenters and the prompt explicitly instructs the LLM whether to accept or forward rumors. The observed decline from 'high acceptance' to 'random' to 'low acceptance' is therefore a manipulation check, not an emergent or discovered property of LLM agents. Furthermore, Figure 8 apparently reports a single run per condition without error bars or statistical tests. The conclusion should be reframed as confirming that the prompt-based trait controls affect the simulation, and the limitations of this design should be acknowledged.
  4. [Table 1 / Effect of Network Structure] The comparison across network types is confounded because the networks differ in several structural properties simultaneously. The Small-World network has only 200 edges and average degree 4.0, while the Erdős-Rényi and Scale-Free networks have roughly twice the edges (396 and 390) and average degrees 7.92 and 7.80; the Facebook network has 1656 edges and average degree 19.71. The paper attributes the higher spread in Small-World to 'sparse connectivity and moderate clustering,' but this is indistinguishable from a simple density effect, and the clustering-coefficient explanation is not tested independently. To support a claim about network structure per se, the authors should use matched networks that vary one property at a time, or at least perform a controlled ablation.
  5. [Abstract / Conclusion] The claim that the framework offers 'a realistic simulation of rumor spread in social networks' is not established. The paper does not compare the simulated dynamics against human rumor-spreading data, existing empirical studies, or classical rumor models such as SIR-like or compartmental models. Without such a benchmark, the realism claim is unsupported. The authors should either add a validation experiment or substantially soften the realism language to 'illustrative' or 'qualitatively plausible' until such validation is provided.
minor comments (7)
  1. [Table 1] Table 1 contains a stray '#686' entry in the header row that appears to be an artifact; it should be removed or explained.
  2. [Algorithm 1] Algorithm 1 line 2 writes '{rj}L i=j' but the bound should likely be 'j=1'; please correct the subscript.
  3. [Rumor list] The first rumor is typeset as 'Nicolae Ceaus, escu is not dead!' with an apparent comma inside the name; the typo should be fixed.
  4. [Figure 3] Figure 3 is said to show 'an almost linear relationship' between rumor spread and time, but the axes are not described and no quantitative slope or confidence information is given; please add axis labels and units or clarify what is plotted.
  5. [Prompt Template] The example prompts include rumors such as 'COVID-19 now named as COVID-114514' that are not in the actual rumor list; the authors should clarify whether these are illustrative only and avoid confusing the reader about the experimental rumors.
  6. [References] Several references are incomplete (e.g., 'OpenAI. 2024. ChatGPT' gives only a URL, and some arXiv entries lack version or venue information); please complete them according to the journal style.
  7. [Figure 8] Figure 8 is described as showing a decline in rumor spread as receptivity decreases, but no numerical axis values or error bars are visible in the description; please provide a readable figure with explicit quantitative results.

Circularity Check

1 steps flagged · score 2.0 of 10

Minor self-definitional persona manipulation check; network-structure results remain emergent.

  1. self definitional [Experiments, 'Effect of Agent's Personas'; Appendix, 'Agent Personas' and 'Prompt Template']
    "You {likely to accept rumors [agent rumors acc]}, and you {likely to forward rumors [agent rumors spread]} ... likely to accept rumors: 4: will easily accept any rumors, even if there are doubts or criticisms. As expected, the agents’ personality configurations significantly influence the spread of rumors. The results ... reveal a clear decline in rumor propagation as the agents’ likelihood of accepting rumors transitions from highly receptive to highly resistant."

    The independent variable in the persona experiment is the numeric 'agent rumors acc' and 'agent rumors spread' values, which are inserted directly into the agent's system prompt as instructions about how readily the agent should accept and forward rumors. The dependent variable is the agent's CHECK belief and post behavior, i.e., the same construct of rumor acceptance and spread. Thus the observed decline in spread as receptivity decreases is a prompt-compliance check, true by construction, rather than an emergent finding. This does not bear on the network-structure or scalability claims, which are emergent simulation outcomes.

full rationale

The paper contains no fitted parameters, no derived equations, and no self-citations. The network-structure comparison (Figures 2 and 3) is emergent: the four networks differ in degree, clustering, and path length, and the LLM agents' beliefs are not directly set by those metrics, so the reported ordering is an empirical simulation result. Similarly, the initialization/activation scheme experiment varies algorithmic inputs and observes spread; although hub-targeting effects are algorithmically expected, the LLM must still choose whether to post and believe rumors, so it is not purely circular. The one definitional step is the persona experiment: 'agent rumors acc' and 'agent rumors spread' are inserted verbatim into the prompt as instructions about how readily the agent should accept and forward rumors, and the measured outcome is exactly rumor acceptance and forwarding. Figure 8 therefore demonstrates prompt compliance rather than an independent discovery. Because the persona claim is one of three headline effects and is not load-bearing for the framework's scalability or network-structure findings, the paper has only minor, partial self-definitionality, not significant circularity.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The central quantitative outcomes are driven by hand-set persona scores and strategy choices; the only external input is the proprietary LLM, whose behavior is not independently validated. This makes the 'realistic simulation' claim depend on an unverified proxy assumption.

free parameters (6)
  • Agent rumor acceptance score (1-4) = 3 or 4 for high-acceptance config; random 1-4 for random config; 1 for low-acceptance config
    Hand-assigned in the prompt template (Appendix 'Agent Personas'); directly controls how readily agents accept rumors, so Experiment 3 mostly reflects this setting.
  • Agent rumor spread score (1-4) = 3 for baseline personas; 1-4 across configs
    Hand-assigned forwarding tendency; controls whether agents share rumors seen in posts.
  • Activation strategy = random or degree-based
    Experimental condition; when set to degree-based, popular agents post more, driving spread.
  • Initialization strategy = random or degree-based
    Determines which agents first receive the rumor; degree-based initialization strongly increases spread.
  • Rumor set = four hand-picked rumors (Ceaușescu, dinosaur, LLM, ale/cancer)
    Chosen by authors; rumor content influences spread through the LLM's pretrained knowledge, a confound for network-structure comparisons.
  • Number of iterations T = 500
    Chosen for all runs; no ablation on T.
assumptions (4)
  • domain assumption LLM agents' outputs are a valid proxy for human rumor-spreading behavior.
    The entire framework and conclusions rely on this; the paper provides no comparison with human rumor-spread data or traditional rumor models.
  • domain assumption The self-reported belief check (True/False per rumor) accurately reflects the agent's actual belief.
    The prompt asks the LLM to output CHECK values, and the belief matrix B is built from these; there is no grounding against behavior.
  • ad hoc to paper A single 500-iteration run is representative of the stochastic process.
    No seeds, repeated runs, or confidence intervals are reported; every figure shows maxima over one run per condition.
  • standard math Standard network generative models are implemented correctly.
    ER, SF, and SW models are standard; results depend on their parameters but exact generation parameters are not fully specified.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Simulating Rumor Spreading in Social Networks using LLM Agents." pith.science (2026). https://pith.science/paper/AO6GPWCM

@misc{pith2026250201450,
  author       = {Pith},
  title        = {Pith review of: Simulating Rumor Spreading in Social Networks using LLM Agents},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AO6GPWCM}},
  note         = {Machine review of arXiv:2502.01450}
}
read the original abstract

With the rise of social media, misinformation has become increasingly prevalent, fueled largely by the spread of rumors. This study explores the use of Large Language Model (LLM) agents within a novel framework to simulate and analyze the dynamics of rumor propagation across social networks. To this end, we design a variety of LLM-based agent types and construct four distinct network structures to conduct these simulations. Our framework assesses the effectiveness of different network constructions and agent behaviors in influencing the spread of rumors. Our results demonstrate that the framework can simulate rumor spreading across more than one hundred agents in various networks with thousands of edges. The evaluations indicate that network structure, personas, and spreading schemes can significantly influence rumor dissemination, ranging from no spread to affecting 83\% of agents in iterations, thereby offering a realistic simulation of rumor spread in social networks.

Figures

Figures reproduced from arXiv: 2502.01450 by the authors.

Figure 1
Figure 1. Design of LLM-based multi-agent network. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Maximum percentage of affected nodes across [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Propagation of Rumor #2. The Small-World net [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: All rumors are spread when they originate from [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: Visualization of Erdos-R ˝ enyi random network. ´ [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Visualization of Scale-Free network [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 7
Figure 7. Figure 7: Visualization of Small World network. Agent Personas In all experiments, each agent’s personas are randomly gen￾erated by ChatGPT-4, following this structure: id: 3 agent name: Leo agent age: 35 agent job: Software Developer agent traits: Analytical, Persistent agent r…
Figure 8
Figure 8. Figure 8: As the personality of the agents shifts from being [PITH_FULL_IMAGE:figures/full_fig_p007_8.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Graphs Meet AI Agents: Taxonomy, Progress, and Future Opportunities

    cs.AI 2025-06 conditional novelty 5.0 of 10

    A survey that groups graph-empowered AI agent research into planning, execution, memory, and multi-agent coordination, plus agents-for-graphs and applications.

  2. Public Discourse Sandbox: Facilitating Human and AI Digital Communication Research

    cs.CY 2025-05 conditional novelty 4.0 of 10

    The Public Discourse Sandbox is a Django-based research platform with human-AI and AI-AI accounts, IRB-based experiment management, and open-source deployment.

Reference graph

Works this paper leans on

20 extracted references · 14 canonical work pages · cited by 2 Pith papers

  1. [1]

    J.; and Geller, A

    Alam, S. J.; and Geller, A. 2011. Networks in Agent-Based Social Simulation. In Agent-Based Models of Geographical Systems, 199--216. Springer

    work page 2011

  2. [2]

    Barab \'a si, A.-L.; and Bonabeau, E. 2003. Scale-free networks. Scientific american, 288(5): 50--9

    work page 2003

  3. [3]

    Chen, X.; and Wang, N. 2020. Rumor spreading model considering rumor credibility, correlation and crowd classification based on personality. Scientific reports, 10(1): 5887

    work page 2020

  4. [4]

    Chen, Y.; Arkin, J.; Zhang, Y.; Roy, N.; and Fan, C. 2024. Scalable Multi-Robot Collaboration with Large Language Models: Centralized or Decentralized Systems? arXiv:2309.15943

    arXiv 2024

  5. [5]

    Chuang, Y.-S.; Goyal, A.; Harlalka, N.; Suresh, S.; Hawkins, R.; Yang, S.; Shah, D.; Hu, J.; and Rogers, T. T. 2024. Simulating Opinion Dynamics with Networks of LLM-based Agents . arXiv:2311.09618

    arXiv 2024

  6. [6]

    Erd \"o s, P.; and R \'e nyi, A. 1959. On random graphs I. Publicationes Mathematicae (Debrecen), 6: 290--297. Dedicated to O. Vargo on the occasion of his 50th birthday

    work page 1959

  7. [7]

    Hamidian, S.; and Diab, M. T. 2019. Rumor Detection and Classification for Twitter Data . arXiv:1912.08926

    work page Pith review arXiv 2019

  8. [8]

    Kaligotla, C.; Y \"u cesan, E.; and Chick, S. E. 2015. An agent based model of spread of competing rumors through online interactions on social media. In 2015 winter simulation conference (WSC), 3985--3996. IEEE

    work page 2015

Show all 20 references
  1. [9]

    Leskovec, J.; and Mcauley, J. 2012. Learning to discover social circles in ego networks. Advances in neural information processing systems, 25

    2012

  2. [10]

    Li, J.; Zhang, Q.; Yu, Y.; Fu, Q.; and Ye, D. 2024. More Agents Is All You Need. arXiv:2402.05120

    2024 arXiv

  3. [11]

    Liu, Q.; Tao, X.; Wu, J.; Wu, S.; and Wang, L. 2024. Can Large Language Models Detect Rumors on Social Media? arXiv:2402.03916

    2024 arXiv

  4. [12]

    OpenAI. 2024. ChatGPT . https://www.openai.com

    2024

  5. [13]

    S.; O'Brien, J.; Cai, C

    Park, J. S.; O'Brien, J.; Cai, C. J.; Morris, M. R.; Liang, P.; and Bernstein, M. S. 2023. Generative Agents: Interactive Simulacra of Human Behavior. In Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology. ISBN 9798400701320

    2023

  6. [14]

    Radford, A.; Narasimhan, K.; Salimans, T.; and Sutskever, I. 2018. Improving Language Understanding by Generative Pre-Training. Available: OpenAI website

    2018

  7. [15]

    J.; and Strogatz, S

    Watts, D. J.; and Strogatz, S. H. 1998. Collective dynamics of ‘small-world’ networks. Nature, 393: 440--442. Published: 04 June 1998

    1998

  8. [16]

    Wilensky, U. 1999. NetLogo . http://ccl.northwestern.edu/netlogo

    1999

  9. [17]

    N.; Out, C.; and Khelejan, S

    Zehmakan, A. N.; Out, C.; and Khelejan, S. H. 2023. Why Rumors Spread Fast in Social Networks, and How to Stop It. arXiv:2305.08558

    2023 arXiv

  10. [18]

    E.; et al

    Zheng, L.; Yin, L.; Xie, Z.; Huang, J.; Sun, C.; Yu, C.; Cao, S.; Kozyrakis, C.; Stoica, I.; Gonzalez, J. E.; et al. 2023. Efficiently Programming Large Language Models using SGLang

    2023

  11. [19]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all...

  12. [20]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.