Pith. sign in

REVIEW 3 major objections 4 minor 74 references

Modeling Social Dynamics with an LLM-Enabled Agent Based Network-Dynamic (LAND) Model

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read The paper claims that coordination and influence in online social systems emerge from a recursive loop between network structure and narrative exchange, and that a 314,244-agent LLM simulation reproduces this emergence with dynamics compara

desk verdict A large-scale LLM+ABM simulation with real engineering value, but the validation evidence is internally contradictory and the emergence claim is undercut by heavy scripting. read the letter →

arxiv 2608.00929 v1 pith:CZA5H2HB submitted 2026-08-02 cs.AI

classification cs.AI
keywords agent-basedmodelinglargelanguagemodelssocialmediadynamicsego-networktopologiessemanticnetworkscoordinationinfluencesimulation
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

The paper sets out to show that large-scale social simulation can generate realistic emergent social dynamics by combining agent-based modeling with large language models. It builds the AuraSight scenario, in which 314,244 heterogeneous agents exchange 529,327 messages over 30 days around a fictional international song contest, and examines the resulting ego-network evolution, semantic network change, coordination, and influence maneuvers. The central claim is that coordination and influence do not come from individual agent cognition but from the recursive loop between network topology and narrative exchange: structure channels discourse, discourse reshapes structure, and this loop produces the observed dynamics. If correct, this gives social scientists a controlled testbed for ethical counterfactual experiments on influence, coordination, and narrative manipulation at scale.

What carries the argument

The load-bearing mechanism is the recursive coupling of network topology and narrative exchange, realized through the GhostField architecture—a hybrid LLM-enabled agent-based network-dynamic (LAND) model with two subsystems: AESOP, which converts a storyline into parameterized scenario templates (events, topics, agent personas, stances, behaviors, narrative seeds), and SynTelX, which generates the actual messages conditioned on agent personas, topic selection, and topic popularity. Agent activation follows a geometric-distribution stochastic rhythm; interaction is constrained by network-science principles like preferential attachment and key influencers. The architecture defines the interact

What would settle it

Run the AuraSight scenario under multiple random seeds (or with different LLM sampling temperatures) and compare the four-week trajectories of the key actors' eigenvector and betweenness centrality, the Week 3-4 semantic-network QAP correlation, and the $\chi^2$ statistics for the Dismay and Neglect maneuvers. If these quantities swing widely across seeds—e.g., Oliver's betweenness centrality dropping below 0.9 or the Week 3-4 semantic correlation moving outside 0.3–0.5—then the claim that macro-level dynamics are reproducible, and hence the stylized-fact validation, would fail.

Watch

Extended reading notes

Core claim

The core discovery is an existence proof at scale: a hybrid LLM-enabled agent-based network-dynamic model, instantiated as AuraSight, produces emergent coordination and influence dynamics that resemble real online ecosystems. In the simulation, Oliver, Ella, and Ezekiel develop different ego-network centrality trajectories—Oliver's centrality stays high and stable across all four weeks, while Ella's rises sharply only when the legal narrative implicates her—without any scripted centrality outcome. Semantic networks grow and reorganize around each event, with the sharpest structural break occurring when the narrative shifts from personal competition to legal conflict. Coordination splits alon

Load-bearing premise

The load-bearing premise is that one stochastic run of the simulation represents the model's macro-level dynamics: the paper states that runs differ but macro structure is reproducible, yet reports no repeated-seed variance analysis, so if a different random draw changed centrality or coordination patterns, the stylized-fact validation would not be established.

Editorial extensions

If this is right

  • If the central claim holds, a single LAND simulation provides a controlled, repeatable environment for counterfactual social-cybersecurity experiments—altering event sequences, actor stances, or bot presence without the ethical problems of manipulating real platforms.
  • Models that attribute social dynamics to agent attributes alone (follower counts, posting frequency, bot likelihood) will miss the structural dimension; models that ignore narrative content will miss content-driven topology changes—both must be modeled jointly.
  • The stylized-fact matches imply that coordination and influence patterns in simulated data can serve as proxies for real-world detection: bot-intensive semantic coordination and human-dominated social coordination are reproducible signatures.
  • Influence maneuver portfolios shift in response to narrative events even with no scripted strategy change, suggesting adversarial narrative management can be detected as a temporal signature rather than a single metric.

Reading between the lines

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

  • A natural extension the paper leaves implicit is quantitative validation: match the simulated data against empirical corpora on degree distributions, retweet cascade shapes, or synchronization-index distributions, rather than qualitative stylized facts.
  • Because the architecture is parameterized by event salience and bot archetypes, it could isolate causal contributions: remove bot archetypes or delay an event and measure how much of the Dismay shift or centrality consolidation disappears.
  • The paper states that runs differ but macro structure is reproducible; a direct test is to report seed sensitivity and cross-run variance, which would turn the framework from a single illustrative run into a robust experimental platform.
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

3 major / 4 minor

Summary. The paper introduces the LAND (LLM-Enabled Agent Based Network-Dynamic) model, implemented via the GhostField architecture, and instantiates it as the AuraSight scenario: a 30-day simulated social-media environment with 314,244 agents and 529,327 messages. The authors analyze four layers of social dynamics—ego-network topology, semantic network evolution, agent-agent coordination, and BEND influence maneuvers—and claim that the observed dynamics emerge from recursive interactions between network topology and narrative exchange, not from individual agent cognition. They further claim that the simulation reproduces several stylized facts of real-world social media dynamics, thereby providing a controlled testbed for social-cybersecurity research.

Significance. If the central claims held, this would be a noteworthy contribution: a large-scale LLM-ABM that generates linguistically realistic, network-structured social-media data in a standard API format, with a multi-layer analysis protocol. The use of the GhostField architecture and the BEND framework connects the work to an established social-cybersecurity literature, and the stylized-fact validation approach is a sensible way to benchmark generative simulations against empirical regularities. However, the current manuscript contains internal inconsistencies that directly undermine a key empirical stylized fact, and the claimed 'emergence without scripted directions' is contradicted by the scenario parameterization. These issues need to be resolved before the contribution can be accepted.

major comments (3)
  1. [Section 4.3, Table 9] The text states that 'For semantic coordination, Bot-Bot dyads consistently exhibit the highest mean edge weights across all snapshots (e.g. Day 17 μ_bot=4.11, μ_human=1.74).' Appendix C, Table 9 directly contradicts this: in every weekly snapshot of semantic-coordination edge weights, the Human-Human mean is the highest (Week 1: 2.08 vs 1.76; Week 2: 2.51 vs 2.36; Week 3: 1.74 vs 1.53; Week 4: 1.84 vs 1.74), and Bot-Bot is the lowest. The cited example values are also misattributed: 4.11 appears in Table 10 for social coordination (Week 3 Bot-Bot), and 1.74 appears in Table 9 as the Week 3 Human-Human mean. This internal inconsistency invalidates the stylized fact 'Bot coordination is more intensive than human coordination' in Table 3 and the corresponding statement in Figure 7's caption. The authors must either correct the analysis, or revise the text and the stylized-fact table to mat
  2. [Sections 3.1, 3.2, 4.4, and 6] The paper repeatedly claims that the observed dynamics are 'emergent' and 'without any scripted directions' (e.g., Section 4.4: 'The GhostField architecture reproduces this sequencing organically and without any scripted directions'). However, Section 3.1 states that AESOP parameterizes the scenario with a storyline, event timeline, topics and pro/anti stances, agent rosters with assigned stances, narrative seeds, and behavior archetypes; Section 3.2 describes fifteen bot archetypes with behavior characteristics; and Section 4.1 itself attributes Oliver's bot-heavy network to 'the behavioral scripting of bot archetypes that amplify the narratives of high-centrality human actors.' These statements are in direct tension. The macro-level outcomes—who supports or opposes whom, when events occur, and how bots behave—are at least partly specified by the scenario designer. The claim that dynami
  3. [Section 3, first paragraph; Section 4, overall] The manuscript asserts that 'although the system is stochastic, each run produces a slightly different realization of agent communications, but crucially, the structural and behavioral constraints ensure that the macro-level social dynamics remain reproducible over time.' No evidence for this reproducibility is provided. All reported results—centrality values, QAP correlations, coordination edge-weight means, and BEND chi-square statistics—come from a single stochastic realization. The differences that ground some stylized facts are small (e.g., Table 9 semantic-coordination means of 1.76 vs 2.08 in Week 1), and without multiple seeds or runs, it is impossible to know whether the conclusions are stable or artifacts of one draw. The authors should report seed sensitivity, confidence intervals, or replicate runs, and at minimum refrain from claiming macro-level reproducibility without supp
minor comments (4)
  1. [Abstract and Section 1] There are repeated typos and grammatical slips, e.g., 'uses the the GhostField architecture' in the abstract, 'clsuters' in Section 4.2, 'intreractions' in Section 6, and 'the converge of edge weight distributions' in Section 4.3. A careful proofreading pass is needed.
  2. [Figure 7 caption and Section 4.3] The caption states that 'Bot-Bot and Bot-Human agent dyads have more coordination between them than the Human-Human dyad.' This is not supported by Table 9 for semantic coordination, where Human-Human dyads have the highest mean edge weights in all four snapshots. The caption should be revised to be consistent with the reported quantitative results.
  3. [Section 4.3] The text refers to 'Day 17' when discussing the week-3 snapshot, but the analysis is framed in terms of weekly events. Day numbering is not introduced until the results are discussed; please align the terminology (Day 17 vs. Week 3) throughout the coordination section and appendices.
  4. [Table 2 and Section 4.4] The chi-square tests are performed on a single simulated dataset, and several maneuvers report identical chi-square values (Distract, Narrow, Enhance all χ²=18.34). This is not inherently impossible, but the authors should clarify the degrees of freedom and cell counts, and address the dependence of the tests on the single-run realization discussed in the major comments.

Circularity Check

3 steps flagged · score 7.0 of 10

Validation stylized facts partly reduce to scripted inputs: Back/Negate by stance definitions, bot coordination and Oliver centrality by scripted bot behavior.

  1. self definitional [Section 3.2 / Section 4.4 / Section 5, Table 3]
    "Back maneuvers are consistently performed by pro-Oliver, pro-Ella, pro-Ezekiel actors throughout the scenario. ... Negate shows zero temporal variation and is exclusively produced by anti-Oliver agents across all time windows. [Table 3] The BEND framework consists of sixteen adverse and affirmative maneuvers such as Back, Bridge, Negate or Dismiss. [4.4]"

    The AESOP scenario input assigns agents to topics and stances ('a roster of agents ... assigned to topics, stances towards topics, and narratives that they will engage in', Sec 3.1), and the BEND framework labels Back as affirmative and Negate as adverse. Thus 'people who like an actor perform Back' and 'people who dislike an actor perform Negate' are true by construction: the stance is an input and the maneuver label encodes the same valence. Reporting these as validated stylized facts is a renaming of the input labeling as an emergent finding.

  2. fitted input called prediction [Section 4.1 / Section 6]
    "Such reflects the behavioral scripting of bot archetypes that amplify the narratives of high-centrality human actors. [4.1] Oliver’s structural centrality was not prescribed but emerged from the recursive interaction process. [6]"

    The storyline input has Oliver win Ethal's finals (Sec 3.2: 'Oliver ... enters and wins Ethal’s national finals'), and bot archetypes are scripted to amplify high-centrality human actors. The observed stable dominance of Oliver is therefore an expected consequence of the seeded event and the amplification rule, not an independent prediction. Calling it 'not prescribed but emerged' contradicts the paper's own scripting description and reduces the claimed emergence to a re-description of the input.

1 more flagged steps
  1. fitted input called prediction [Section 4.3 / Section 5, Table 3 / Appendix C, Table 9]
    "The core of the networks are typically dominated by bot agents (red nodes), which exhibit significantly higher Combined Synchronization Index values than human agents. This reflects the behavioral scripting of bot archetypes to deploy coordinated hashtag campaigns. ... For semantic coordination, Bot-Bot dyads consistently exhibit the highest mean edge weights across all snapshots (e.g. Day 17 μ_bot=4.11, μ_human=1.74)."

    Coordinated hashtag deployment is a scripted behavior of bot archetypes, i.e., a design input, so the resulting bot dominance in semantic coordination is the script's direct output rather than an emergent stylized fact that independently validates realism. In addition, Appendix Table 9 contradicts the 'highest mean edge weights' claim in every snapshot (e.g., Week 1 Human-Human 2.08 vs Bot-Bot 1.76; Week 3 1.74 vs 1.53), and the illustrative 4.11 appears in Table 10 (social coordination), not Table 9. This is a correctness problem that further undermines the validation, though the circularity is the scripted-input-as-prediction pattern.

full rationale

The paper is a simulation study rather than an axiomatic derivation, and the central mechanism (GhostField/AESOP/SynTelX, BEND analysis) is not machine-checked or presented as a uniqueness theorem; no self-citation chain is the main issue. The circularity is concentrated in the validation-by-stylized-facts section. Three Table 3 stylized facts are not independent evidence: (a) Back/Negate are defined by stance and valence, so 'people who like an actor use Back' is true by construction; (b) bot coordination is scripted as coordinated hashtag campaigns, so bot-heavy semantic coordination is an output of the script; (c) Oliver's stable centrality is explained by bot archetypes scripted to amplify high-centrality humans, contradicting the later claim that his centrality was 'not prescribed.' The paper itself says 'the storyline constraints the alignment of the social and semantic networks' and describes 'behavioral scripting of bot archetypes,' so the claimed emergence is partly a re-description of inputs. Separately, not circularity but a serious internal inconsistency: Section 4.3's claimed Bot-Bot semantic-coordination dominance is contradicted by Appendix Table 9 in every snapshot, and the illustrative 4.11/1.74 values come from different tables. This eliminates the one stylized fact that was not already scripted. The semantic-network and BEND temporal shifts are also driven by the scripted event timeline, so they are weaker independent support than the paper claims. Overall, the headline 'dynamics emerge from network-topology/narrative recursion' is not fully forced by definition, but its main validation facts reduce to scripted inputs and definitional maneuver labels, giving partial-to-substantial circularity.

Assumptions & free parameters 5 free parameters · 6 assumptions · 0 invented entities

The paper introduces no new scientific entities. Oliver, Ella, Ezekiel, and the bot archetypes are fictional scenario elements or classifications from prior literature, not postulated mechanisms with independent falsifiable handles. The LAND model is an architecture assembled from existing parts (GhostField, SynTelX, GPT-4.1) rather than a new entity claiming independent evidence.

free parameters (5)
  • Agent persona and stance roster = Hand-authored (values not disclosed)
    The three protagonists, 50 hand-instantiated agents, and their topics/stances determine the narrative roles; Oliver is scripted as the central figure.
  • Event timeline = 4 weeks, 4 key events
    The events (national finals win, statement, court proceedings) are chosen to create the observed narrative arc.
  • Bot behavior archetypes = 15 archetypes from ref [51]
    Amplifier bots are restricted to retweets, thus mechanically boosting high-centrality actors; this constrains the coordination and influence outcomes.
  • Coordination thresholds = 95th percentile; 5-minute window
    Edges in coordination graphs are defined by the 95th-percentile threshold, which controls the density and composition of coordination networks.
  • Number of LLM-generated personas = 50 (generated from 50 manual)
    The expansion from manual to LLM-generated personas is a design choice affecting heterogeneity but not disclosed in full.
assumptions (6)
  • domain assumption Agent activation follows a geometric distribution reflecting posting rhythm
    Section 3.1; assumption that human-like posting timing is geometric, with no validation on the simulation's output.
  • domain assumption The behavioral constraints of agent classes are derived from empirically observed online behaviors
    Section 3.2; this transfers real-world categories onto synthetic agents without demonstrating equivalence.
  • domain assumption The 5-minute coordination window from offline literature applies to this synthetic X-like environment
    Section 4.3; borrowed threshold, no sensitivity analysis.
  • domain assumption A single stochastic run is representative of macro-level dynamics
    Section 3 claims macro reproducibility despite stochasticity, but no multi-run variance analysis is provided.
  • standard math Network analysis methods (eigenvector/betweenness centrality, QAP, modularity) are valid on the constructed graphs
    Standard toolbox; but QAP is described while Pearson correlations are reported.
  • ad hoc to paper The storyline constrains the alignment of social and semantic networks
    Section 3.1; this is the circular core: the scenario explicitly defines the narrative alignment that the paper later 'discovers'.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Modeling Social Dynamics with an LLM-Enabled Agent Based Network-Dynamic (LAND) Model." pith.science (2026). https://pith.science/paper/CZA5H2HB

@misc{pith2026260800929,
  author       = {Pith},
  title        = {Pith review of: Modeling Social Dynamics with an LLM-Enabled Agent Based Network-Dynamic (LAND) Model},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CZA5H2HB}},
  note         = {Machine review of arXiv:2608.00929}
}
read the original abstract

Social dynamics encode the process in which individual network and discourse interactions aggregate into collective influence, narrative dominance and coordinate behavior. This paper uses the the GhostField architecture, a hybrid LLM-Enabled Agent Based Network-Dynamic (LAND) model as a social simulation framework to build the AuraSight scenario. In the AuraSight scenario, 314,244 heterogeneous cyber social agents and human actors exchange 529,327 messages over 30 days surrounding a fictional international song-writing contest. We methodologically examine emergent social dynamics across four analytical layers: ego-network topology, semantic network evolution, coordination dynamics and influence dynamics. Our results show how generated social simulations do also produce social dynamics, and how the dynamics of coordination and influence emerge not from individual agents but from the recursive interaction between network topology and narrative exchange.

Figures

Figures reproduced from arXiv: 2608.00929 by the authors.

Figure 1
Figure 1. AuraSight support diagram of the key actors, support factions and news agencies. Green arrows from [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Two-hop Ego Networks of key actors over Time. Nodes represent actors, which are sized by eigenvector [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. QAP heatmap comparison of ego-networks of key actors across events. ** indicates significant [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Semantic Networks. Nodes are hashtags, edges represent co-occurrence of hashtags in the same tweet. [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]
Figure 5
Figure 5. Figure 5: QAP heatmap comparison of semantic networks across events. ** indicates significant difference at [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]
Figure 6
Figure 6. Figure 6: Coordination Networks. Nodes represent agents, which are sized by Combined Synchronization Index. [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]
Figure 7
Figure 7. Figure 7: Comparison of coordination between dyad pairs of agent types. Coordination between two agents is [PITH_FULL_IMAGE:figures/full_fig_p013_7.png]
Figure 8
Figure 8. Figure 8: Influence dynamics over time measured by the BEND metrics. ** indicates [PITH_FULL_IMAGE:figures/full_fig_p014_8.png]
Figure 9
Figure 9. Figure 9: Centrality values of the main characters Oliver, Ella and Ezekiel across the event timestamps. [PITH_FULL_IMAGE:figures/full_fig_p023_9.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

74 extracted references · 65 canonical work pages

  1. [1]

    Carlo Adornetto, Adrian Mora, Kai Hu, Leticia Izquierdo Garcia, Parfait Atchade-Adelomou, Gianluigi Greco, Luis Alberto Alonso Pastor, and Kent Larson. 2025. Generative agents in agent-based modeling: Overview, validation, and emerging challenges.IEEE Transactions on Artificial Intelligence(2025)

  2. [2]

    Iuliia Alieva and Kathleen M Carley. 2026. Dynamics of Russian anti-war discourse on X (Twitter): a computational analysis using NLP and network methods.npj Complexity3, 1 (2026), 2

  3. [3]

    Iuliia Alieva, JD Moffitt, and Kathleen M Carley. 2022. How disinformation operations against Russian opposition leader Alexei Navalny influence the international audience on Twitter.Social Network Analysis and Mining12, 1 (2022), 80

  4. [4]

    Neal Altman, Kathleen M Carley, and Jeffrey Reminga. 2020. ORA user’s guide 2020.Carnegie-Mellon Univ. Pittsburgh PA Inst of Software Research International, Tech. Rep2 (2020), 2

  5. [5]

    Keyvan Azadbakht, Nikolaos Bezirgiannis, and Frank S de Boer. 2017. Distributed network generation based on preferential attachment in ABS. InInternational Conference on Current Trends in Theory and Practice of Informatics. Springer, 103–115

  6. [6]

    Christopher Andrew Bail. 2016. Combining natural language processing and network analysis to examine how advocacy organizations stimulate conversation on social media.Proceedings of the National Academy of Sciences113, 42 (2016), 11823–11828

  7. [7]

    Albert-László Barabási and Réka Albert. 1999. Emergence of scaling in random networks.science286, 5439 (1999), 509–512

  8. [8]

    David M Beskow and Kathleen M Carley. 2019. Agent based simulation of bot disinformation maneuvers in Twitter. In 2019 Winter simulation conference (WSC). IEEE, 750–761

Show all 74 references
  1. [9]

    John M Betts and Ana-Maria Bliuc. 2022. The effect of influencers on societal polarization. In2022 Winter Simulation Conference (WSC). IEEE, 370–381

  2. [10]

    Danah Boyd. 2010. Social network sites as networked publics: Affordances, dynamics, and implications. InA networked self. Routledge, 47–66

  3. [11]

    Ulrik Brandes. 2001. A faster algorithm for betweenness centrality.Journal of mathematical sociology25, 2 (2001), 163–177

  4. [12]

    Anna D Broido and Aaron Clauset. 2019. Scale-free networks are rare.Nature communications10, 1 (2019), 1017

  5. [13]

    Kathleen M Carley. 1990. Group stability: A socio-cognitive approach.Advances in group processes7, 1 (1990), 44

  6. [14]

    Kathleen M Carley. 2020. Social cybersecurity: an emerging science.Computational and mathematical organization theory26, 4 (2020), 365–381

  7. [15]

    Kathleen M. Carley. 2026.BEND: Operationalizing Influence Campaigns. Technical Report Technical Report 20260115-

  8. [16]

    Netanomics Technical Report

  9. [17]

    Kathleen M Carley, Michael K Martin, and Brian R Hirshman. 2009. The etiology of social change.Topics in Cognitive Science1, 4 (2009), 621–650

  10. [18]

    Carley, Jonathan H

    Kathleen M. Carley, Jonathan H. Morgan, and Mihovil Bartulovic. 2026.GhostField. Technical Report Technical Report 20260609-001. Netanomics Technical Report

  11. [19]

    Francesco Cauteruccio, Salvatore Citraro, Andrea Failla, Giulio Rossetti, et al. 2026. Structure and dynamics of temporal hypergraphs via multi-rooted ego networks.Social Network Analysis and Mining(2026), 1–24

  12. [20]

    Kamer Cekini, Elisabetta Biondi, Chiara Boldrini, Andrea Passarella, and Marco Conti. 2026. The Impact of COVID-19 on Twitter Ego Networks: Structure, Sentiment, and Topics.Machine Learning115, 3 (2026), 36

  13. [21]

    Alberto Ceria, Shlomo Havlin, Alan Hanjalic, and Huijuan Wang. 2022. Topological–temporal properties of evolving networks.Journal of Complex Networks10, 5 (2022), cnac041

  14. [22]

    Matteo Cinelli, Gianmarco De Francisci Morales, Alessandro Galeazzi, Walter Quattrociocchi, and Michele Starnini

  15. [23]

    2009.Marketing communications: A brand narrative approach

    Micael Dahlen, Fredrik Lange, and Terry Smith. 2009.Marketing communications: A brand narrative approach. John Wiley & Sons

  16. [24]

    Iain J Cruickshank and Kathleen M Carley. 2020. Characterizing communities of hashtag usage on twitter during the 2020 COVID-19 pandemic by multi-view clustering.Applied Network Science5, 1 (2020), 66

  17. [25]

    James A Danowski. 1993. Network analysis of message content.Progress in communication sciences12 (1993), 198–221

  18. [26]

    Adya Danaditya, Lynnette Hui Xian Ng, and Kathleen M Carley. 2022. From curious hashtags to polarized effect: profiling coordinated actions in indonesian twitter discourse.Social Network Analysis and Mining12, 1 (2022), 105

  19. [27]

    Jana Diesner and Kathleen M Carley. 2005. Revealing social structure from texts: meta-matrix text analysis as a novel method for network text analysis. InCausal mapping for research in information technology. IGI Global Scientific Publishing, 81–108. J. ACM, Vol. 37, No. 4, Ar...

  20. [28]

    Matthew R DeVerna, Rachith Aiyappa, Diogo Pacheco, John Bryden, and Filippo Menczer. 2024. Identifying and characterizing superspreaders of low-credibility content on Twitter.Plos one19, 5 (2024), e0302201

  21. [29]

    Rui Fan, Ke Xu, and Jichang Zhao. 2018. An agent-based model for emotion contagion and competition in online social media.Physica a: statistical mechanics and its applications495 (2018), 245–259

  22. [30]

    Robert M Entman et al . 1993. Framing: Towards clarification of a fractured paradigm.McQuail’s reader in mass communication theory390 (1993), 397

  23. [31]

    Noah E Friedkin and Eugene C Johnsen. 1990. Social influence and opinions.Journal of mathematical sociology15, 3-4 (1990), 193–206

  24. [32]

    Yujie Feng, Zexin Lu, Bo Liu, Liming Zhan, and Xiao-Ming Wu. 2023. Towards LLM-driven dialogue state tracking. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. 739–755

  25. [33]

    Sharad Goel, Ashton Anderson, Jake Hofman, and Duncan J Watts. 2016. The structural virality of online diffusion. Management science62, 1 (2016), 180–196

  26. [34]

    David R Gibson. 2005. Taking turns and talking ties: Networks and conversational interaction.American journal of sociology110, 6 (2005), 1561–1597

  27. [35]

    Mark Granovetter. 1985. Economic action and social structure: The problem of embeddedness.American journal of sociology91, 3 (1985), 481–510

  28. [36]

    Timothy Graham, Sam Hames, and Elizabeth Alpert. 2024. The coordination network toolkit: a framework for detecting and analysing coordinated behaviour on social media.Journal of Computational Social Science7, 2 (2024), 1139–1160

  29. [37]

    Binxuan Huang and Kathleen M Carley. 2019. A hierarchical location prediction neural network for twitter user geolocation. InProceedings of the 2019 conference on empirical methods in natural language processing and the 9th international joint conference on natural language pr...

  30. [38]

    2024.AI-enabled social cyber maneuver detection and creation

    Matthew Hicks. 2024.AI-enabled social cyber maneuver detection and creation. Ph. D. Dissertation. Thesis]. https://mhicks3. github. io/Hicks_thesis_proposal. pd f

  31. [39]

    Tuja Khaund, Baris Kirdemir, Nitin Agarwal, Huan Liu, and Fred Morstatter. 2021. Social bots and their coordination during online campaigns: a survey.IEEE Transactions on Computational Social Systems9, 2 (2021), 530–545

  32. [40]

    Hawoong Jeong, Zoltan Néda, and Albert-László Barabási. 2003. Measuring preferential attachment in evolving networks.EPL (Europhysics Letters)61, 4 (2003), 567–572

  33. [41]

    Jiazhen Liu, Shengda Huang, Nathaniel M Aden, Neil F Johnson, and Chaoming Song. 2023. Emergence of polarization in coevolving networks.Physical review letters130, 3 (2023), 037401

  34. [42]

    Maik Larooij and Petter Törnberg. 2025. Do large language models solve the problems of agent-based modeling? a critical review of generative social simulations.arXiv preprint arXiv:2504.03274(2025)

  35. [43]

    Thomas Magelinski, Lynnette Ng, and Kathleen Carley. 2022. A synchronized action framework for detection of coordination on social media.Journal of Online Trust and Safety1, 2 (2022)

  36. [44]

    Mei Liu, Jinyuan Zhang, Xiuchun Xiao, and Long Jin. 2024. Opinion dynamics considering Matthew effect under switching topology in social networks.IEEE Transactions on Network Science and Engineering11, 4 (2024), 3866–3875

  37. [45]

    Xinyi Mou, Xuanwen Ding, Qi He, Liang Wang, Jingcong Liang, Xinnong Zhang, Libo Sun, Jiayu Lin, Jie Zhou, Xuanjing Huang, et al . 2024. From individual to society: A survey on social simulation driven by large language model-based agents.arXiv preprint arXiv:2412.03563(2024)

  38. [46]

    Rebecca Marigliano, Lynnette Hui Xian Ng, and Kathleen M Carley. 2024. Analyzing digital propaganda and conflict rhetoric: a study on Russia’s bot-driven campaigns and counter-narratives during the Ukraine crisis.Social Network Analysis and Mining14, 1 (2024), 170

  39. [47]

    Lynnette Hui Xian Ng and Kathleen M Carley. 2022. Online coordination: methods and comparative case studies of coordinated groups across four events in the united states. InProceedings of the 14th ACM Web Science Conference 2022. 12–21

  40. [48]

    Zachary P Neal, Kathleen Carley, Noshir Contractor, Manlio De Domenico, James Hollway, Giuseppe Joe Labianca, Nicholas Landry, David Lazer, Miranda Lubbers, Jennifer Watling Neal, et al. 2026. Guidelines for Reporting About Network Data (GRAND). (2026)

  41. [49]

    Lynnette Hui Xian Ng and Kathleen M Carley. 2025. Are LLM-Powered Social Media Bots Realistic?. InInternational Conference on Social Computing, Behavioral-Cultural Modeling and Prediction and Behavior Representation in Modeling and Simulation. Springer, 14–23

  42. [50]

    Lynnette Hui Xian Ng and Kathleen M Carley. 2023. A combined synchronization index for evaluating collective action social media.Applied network science8, 1 (2023), 1

  43. [51]

    Lynnette Hui Xian Ng and Kathleen M Carley. 2026. 35 The Dual Personas of Social Media Bots. InOnline Trust and Safety: Tools to Combat Online Harms, Misinformation and Malicious Content. CRC Press Boca Raton and London, 35–54

  44. [52]

    Lynnette Hui Xian Ng and Kathleen M Carley. 2025. A global comparison of social media bot and human characteristics. Scientific Reports15, 1 (2025), 10973

  45. [53]

    Lynnette Hui Xian Ng, Iain J Cruickshank, Adrian Xuan Wei Lim, and Kathleen M Carley. 2026. Social Theory Should Be a Structural Prior for Agentic AI: A Formal Framework for Multi-Agent Social Systems.arXiv preprint arXiv:2605.07069(2026)

  46. [54]

    Lynnette Hui Xian Ng and Kathleen M Carley. 2026. BotSim: Mitigating The Formation Of Conspiratorial Societies with Useful Bots.Journal of Artificial Societies and Social Simulation29, 1 (2026), 4. J. ACM, Vol. 37, No. 4, Article 111. Publication date: August 2018. 111:22 Ng & Carley

  47. [55]

    Diogo Pacheco, Pik-Mai Hui, Christopher Torres-Lugo, Bao Tran Truong, Alessandro Flammini, and Filippo Menczer

  48. [56]

    Lynnette Hui Xian Ng, Ian Kloo, Samantha Clark, and Kathleen M Carley. 2024. An exploratory analysis of COVID bot vs human disinformation dissemination stemming from the Disinformation Dozen on Telegram.Journal of Computational Social Science7, 1 (2024), 695–720

  49. [57]

    Steve Rathje and Jay J Van Bavel. 2025. The psychology of virality.Trends in Cognitive Sciences(2025)

  50. [58]

    InProceedings of the international AAAI conference on web and social media, Vol

    Uncovering coordinated networks on social media: methods and case studies. InProceedings of the international AAAI conference on web and social media, Vol. 15. 455–466

  51. [59]

    Joon Sung Park, Joseph O’Brien, Carrie Jun Cai, Meredith Ringel Morris, Percy Liang, and Michael S Bernstein. 2023. Generative agents: Interactive simulacra of human behavior. InProceedings of the 36th annual acm symposium on user interface software and technology. 1–22

  52. [60]

    Michael D Slater. 2007. Reinforcing spirals: The mutual influence of media selectivity and media effects and their impact on individual behavior and social identity.Communication theory17, 3 (2007), 281–303

  53. [61]

    Steve Rathje, Jay J Van Bavel, and Sander Van Der Linden. 2021. Out-group animosity drives engagement on social media.Proceedings of the national academy of sciences118, 26 (2021), e2024292118

  54. [62]

    2010.The Matthew effect: How advantage begets further advantage

    Daniel Rigney. 2010.The Matthew effect: How advantage begets further advantage. Columbia University Press

  55. [63]

    1984.Social dynamics models and methods

    Nancy Brandon Tuma. 1984.Social dynamics models and methods. Elsevier

  56. [64]

    Mark Steyvers and Joshua B Tenenbaum. 2005. The large-scale structure of semantic networks: Statistical analyses and a model of semantic growth.Cognitive science29, 1 (2005), 41–78

  57. [65]

    Patrick Taillandier, Jean Daniel Zucker, Arnaud Grignard, Benoit Gaudou, Nghi Quang Huynh, and Alexis Drogoul

  58. [66]

    Lilian Weng, Alessandro Flammini, Alessandro Vespignani, and Fillipo Menczer. 2012. Competition among memes in a world with limited attention.Scientific reports2, 1 (2012), 335

  59. [67]

    Evan M Williams, Peter Carragher, Kyle Herdrich, Luke Prakarsa, and Kathleen M Carley. 2025. Extending the BEND Framework to Webgraphs. InInternational Conference on Social Computing, Behavioral-Cultural Modeling and Prediction and Behavior Representation in Modeling and Simul...

  60. [68]

    Eero Vaara, Scott Sonenshein, and David Boje. 2016. Narratives as sources of stability and change in organizations: Approaches and directions for future research.Academy of management annals10, 1 (2016), 495–560

  61. [69]

    Soroush Vosoughi, Deb Roy, and Sinan Aral. 2018. The spread of true and false news online.science359, 6380 (2018), 1146–1151

  62. [70]

    Ke Zhou, Marios Constantinides, Daniele Quercia, and Sanja Šćepanović. 2023. How circadian rhythms extracted from social media relate to physical activity and sleep. InProceedings of the international AAAI conference on web and social media, Vol. 17. 948–959. A Agent Ego-Netwo...

  63. [72]

    Ziyi Yang, Zaibin Zhang, Zirui Zheng, Yuxian Jiang, Ziyue Gan, Zhiyu Wang, Zijian Ling, Martin Ma, Bowen Dong, Prateek Gupta, et al. 2024. OASIS: Open Agents Social Interaction Simulations on One Million Agents. InNeurIPS 2024 Workshop on Open-World Agents

  64. [73]

    Kun Zhao, Juliette Stehlé, Ginestra Bianconi, and Alain Barrat. 2011. Social network dynamics of face-to-face interactions.Physical Review E—Statistical, Nonlinear, and Soft Matter Physics83, 5 (2011), 056109

  65. [2021]

    The echo chamber effect on social media.Proceedings of the national academy of sciences118, 9 (2021), e2023301118

  66. [2025]

    Integrating llm in agent-based social simulation: Opportunities and challenges.arXiv preprint arXiv:2507.19364 (2025)

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.