Pith. sign in

REVIEW 4 major objections 7 minor 49 references

LLMs are Introvert

T0 review · 4 major / 7 minor · reviewed 2026-08-06 · 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 agents are 'introvert' in social situations, and that a five-stage Social Information Processing chain-of-thought prompt with emotion-guided memory makes their simulated attitudes, emotions, and behaviors…

desk verdict Plausible prompt-level intervention for more human-like LLM social agents, but the headline SIP-testing evidence is confounded and needs a control condition before the psychological claim holds. read the letter →

arxiv 2507.05638 v1 pith:CET2SJCS submitted 2025-07-08 cs.AI cs.SI

Litian Zhang , Xiaoming Zhang , Bingyu Yan , Ziyi Zhou , Bo Zhang , Zhenyu Guan , Xi Zhang , Chaozhuo Li This is my paper
classification cs.AIcs.SI
keywords largelanguagemodelssocialsimulationInformationProcessingchain-of-thoughtpromptingemotion-guidedmemoryopiniondynamicsagent-basedmodelingstancealignment
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 argues that large language models, despite fluent language use, are socially 'introvert': they notice fewer social cues, give compressed and low-variance judgments, and collapse distinct social-cognitive stages into one rigid pattern. The authors built an LLM-based simulation of three Reddit discussion events and found baseline agents diverged from real human trajectories, with a mean bias deviation of 0.108 and stance, content, and emotion F1 scores of 0.626, 0.721, and 0.613. They trace the shortfall to missing emotional processing in standard LLM training and propose SIP-CoT: a prompt-level cognitive architecture, supplemented by emotion-guided memory, that forces agents through cue encoding, interpretation, goal setting, response retrieval, and response evaluation. Enhanced agents reduce mean bias deviation to 0.0521 and raise stance, content, and emotion F1 to 0.745, 0.888, and 0.697. If correct, this means a prompt-only intervention, with no retraining, can materially improve the realism of LLM-based social simulations used to study information propagation and misinformation.

What carries the argument

The load-bearing object is the SIP-based decision procedure, implemented as a chain-of-thought prompt in which the agent must emit five stage-tagged sentences—[Cue], [Interpret], [Goal], [Retrieve], [Evaluate]—before selecting one action. These five tags operationalize the Social Information Processing model's stages of encoding social cues, interpreting them, classifying goals, accessing responses, and evaluating consequences. Supporting this is a social-memory module with long-term social cognitive and behavior memory and a short-term social interaction memory, which together supply emotion-guided retrieval and learning. The mechanism works by forcing decoupling: instead of letting one stance determine the whole response, the agent must explicitly answer each stage, which restores the between-stage variance and the context sensitivity that human raters show.

What would settle it

Collect ratings on the same 13 items from a new, adult, non-translated sample and rerun the Reddit simulation: if SIP-enhanced agents no longer beat baseline in both the questionnaire statistics and the stance/content/emotion F1 scores, the claimed improvement is an artifact of the benchmark rather than of human-like social cognition. Similarly, replacing the SIP stage tags with a generic five-step reasoning prompt would show whether the gains depend on the social-cognitive content at all.

Watch

Extended reading notes

Core claim

The central claim is that the human–agent gap in social simulation is not primarily a knowledge gap but a structural one: baseline LLM agents skip the separable stages of human social information processing, so a single inferred motivational stance drives every later decision and the population loses diversity. Evaluated against human ratings on a Social Information Processing questionnaire, baseline agents clustered around mid-range values, showed inflated correlations among all stages (most $|r| > 0.6$), and failed to distinguish prosocial from hostile dimensions the way humans do. The paper's proposed remedy, the SIP-CoT mechanism with emotion-guided memory, restores those distinctions and produces response distributions whose mean, standard deviation, skewness, and kurtosis track human patterns. On the Reddit propagation benchmark, this yields both closer macro-level trajectories and higher micro-level alignment on stance, content, and emotion. The authors' conclusion is that integrating social-psychological theory into agent prompting significantly enhances the social intelligence and realism of LLM-based agents.

Load-bearing premise

The argument assumes the 100–200 volunteers' five-point ratings on five translated, school-age conflict vignettes are a valid and generalizable measure of human social information processing; if that benchmark does not represent the adult, cross-cultural Reddit population, the measured alignment gains are not evidence of human-likeness.

Editorial extensions

If this is right

  • A prompt-only intervention, without fine-tuning or additional training data, can shift LLM agents toward human-like social information processing in simulations.
  • Population-level opinion trajectories in the three Reddit events become less biased and more diverse, with mean bias deviation falling from 0.108 to 0.0521.
  • Individual-level fidelity rises on all three channels: stance F1 goes from 0.626 to 0.745, content F1 from 0.721 to 0.888, and emotion F1 from 0.613 to 0.697.
  • On the SIP questionnaire, enhanced agents reproduce human central tendency, variability, skewness, and tail behavior more closely, with the largest gains in ambiguous-cue interpretation and response evaluation.

Reading between the lines

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

  • If the effect generalizes, staged social-cognition prompting could improve any LLM application that needs realistic emotional variance—survey respondents, user-behavior models, or audience simulations—without retraining.
  • A matched control prompt with the same number of steps but no SIP content would isolate whether the gains come from the social-cognitive stages or merely from longer structured deliberation; the paper does not report this control.
  • The human benchmark uses translated school-age conflict vignettes, so a natural next test is adult, culture-matched vignettes; persistence of the gains there would strengthen the claim that the architecture captures general social cognition.
  • Policy simulators could exploit the explicit five-stage output to audit why an agent chose a stance or emotion, making misinformation interventions more inspectable before deployment.
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

4 major / 7 minor

Summary. The paper claims that current LLM agents are 'introvert' in social cognition and proposes a Social-Information-Processing Chain-of-Thought (SIP-CoT) mechanism, supplemented by an emotion-guided memory architecture, to make LLM agents process social information more like humans. The authors construct a Reddit-based social simulation dataset, benchmark baseline LLM agents against real propagation data, then evaluate SIP-CoT-enhanced agents. They report that the enhancement reduces mean |Δbias| from 0.108 to 0.0521, improves stance F1 from 0.626 to 0.745, content F1 from 0.721 to 0.888, and emotion F1 from 0.613 to 0.697, and brings Likert-response distributions closer to human responses on a 13-item SIP questionnaire. The central claim is that a prompt-level intervention, without retraining, materially improves the fidelity of LLM-based social simulations.

Significance. If the results hold, the paper would demonstrate a low-cost, theory-grounded intervention that improves LLM agents' fidelity in social simulations, which is relevant to computational social science and agent-based modeling. The paper has several strengths: it builds on an established psychological framework (SIP), uses a real-world Reddit dataset with three distinct events, and proposes a two-tier evaluation (macro-level propagation and micro-level agent alignment). However, the current evidence is not sufficient to establish the central claim. The SIP-testing evidence is confounded because the intervention prompt and the testing instrument share the same five-stage structure, the human benchmark is poorly documented and inconsistent, and the simulation experiments lack basic reproducibility details such as the base LLM, seeds, variance, and significance tests. The contribution is potentially valuable but requires substantial additional validation.

major comments (4)
  1. [SIP-Testing Results with SIP-enhanced Cognitive Architecture; Appendix A.1 and A.2] The SIP-testing evidence for the central claim is confounded by format alignment. The SIP-CoT prompt in Appendix A.1 explicitly instructs the agent to prefix every sentence with [Cue], [Interpret], [Goal], [Retrieve], and [Evaluate], and the SIP-testing questionnaire (Table 4 in Appendix A.2) operationalizes exactly those five stages with 13 Likert items mapped onto the same constructs. Improved agreement in Figure 5 may therefore reflect instruction following rather than genuinely human-like social information processing. The authors should add a control condition with a comparably structured reasoning prompt (e.g., a neutral five-step problem-solving framework with equally formatted tags, or a different psychological framework) to show that the gains are not merely a consequence of imposing the same categorical structure. An ablation separating the SIP-CoT prompt from the memory modules would also clarify which component drives the effect.
  2. [SIP-Testing Results with LLM; SIP-Testing Results with SIP-enhanced Cognitive Architecture; Appendix A.2] The human benchmark is not adequately documented and is internally inconsistent. The first SIP-testing section reports 200 human volunteers, while the corresponding section for the enhanced architecture reports 100 human participants, with no explanation for the change. No demographics, recruitment procedure, compensation, exclusion criteria, or institutional review information are provided for the human sample. In addition, the five background stories in Table 3 (e.g., Forgotten Invitation, Group Project Dismissal) do not match the five vignettes described in Appendix A.2 (e.g., Crowded cafeteria, Football collision, Damaged magazine). The appendix states that the stories were translated from Chinese and adapted from school-age aggression paradigms, but the propagation simulation uses adult Reddit discussions. The validity of the human yardstick is therefore open to question, and the measured 'alignment' gains cannot be interpreted until the benchmark is described precisely and used consistently.
  3. [Social Simulation Performance with LLM; Social Simulation Performance with SIP-Enhanced Cognitive Architecture…] The simulation results are reported without the information needed to assess their reliability or reproducibility. No base LLM is named anywhere in the paper, and no random seeds, number of runs, standard deviations, confidence intervals, or significance tests are reported for any of the headline numbers, including the F1 gains and the Δbias and Δdiv values. The stance, content-type, and emotion labels on the empirical Reddit data are not described: there is no annotation procedure, label taxonomy, or inter-annotator agreement. Without these details, the reader cannot determine whether the reported improvements are statistically meaningful or the result of a single favorable run. The authors should specify the model, the sampling procedure, and the number of independent runs, and report dispersion measures and appropriate statistical tests.
  4. [Methods: SIP enhanced Cognitive Architectures; Appendix A.1] There is a gap between the claimed 'emotion-guided memory' architecture and what is actually implemented and tested. The abstract and the architecture description (Figure 1d–e) emphasize long- and short-term social memory modules and emotion-guided processing, but the experimental apparatus in Appendix A.1 consists solely of three prompt templates, in which 'memory' appears as text fields such as Personal Memory and Chat History. The Methods section describes memory modules and actions (retrieval, reasoning, learning, grounding) without providing an implementation, retrieval algorithm, or pseudocode. No ablation separates the contribution of the memory modules from the SIP-CoT prompt. As a result, the paper does not establish that the proposed cognitive architecture, rather than the prompt format, is responsible for the reported improvements.
minor comments (7)
  1. [Figure 2 caption] The caption reads 'deviation in bias (Δbias, orange) and diversity (Δbias, blue)'; the blue curve should be labeled Δdiv, not Δbias.
  2. [Page 6, Agent Alignment Evaluation] The sentence 'aThe LLM agents achieve the F1-score of 0.626' contains a stray leading 'a'.
  3. [Figure 4 caption] The phrase 'opinion trajectories1 in bias and diversity' contains a stray superscript '1' that should be removed.
  4. [Introduction, first paragraph] The sentence 'current LLMs remain markedly introvert, which failing to interpret subtle social cues' is grammatically incomplete; it should read 'which fail to interpret'.
  5. [References 5 and 6] The reference list contains duplicated '5' and '6' entries in the text citation '5, 5, 6'; this should be cleaned up.
  6. [Appendix A.2, first paragraph] The questionnaire is described as using a five-point scale from 'strongly disagree' to 'strongly agree', but Table 4 shows item-specific anchors such as 'Not at all' to 'Very much'. These descriptions should be made consistent.
  7. [Methods: The Simulation Evaluation] Dynamic Time Warping (DTW) is mentioned as part of the global propagation evaluation, but no DTW results are reported anywhere in the paper; the authors should either report the DTW numbers or remove the reference.

Circularity Check

1 steps flagged · score 6.0 of 10

SIP-testing alignment is partly self-definitional: the SIP-CoT prompt and the evaluation questionnaire share the same five-stage SIP structure, so the Figure 5 gains partly measure prompt compliance rather than independently established human-like cognition; the Reddit propagation pillar remains externally grounded.

  1. self definitional [Appendix A.1 'Task Prompt Templates'; Appendix A.2 'SIP-testing Questionnaire' (Table 4); results in Figure 5.]
    "Prefix every sentence with the stage-tag in brackets. [Cue]Do you need more information to interpret the situation? [Interpret]What is your current understanding of the discussion? [Goal]What is your main communicative goal? ... [Evaluate]Will this response advance the conversation constructively? Return only the five tagged sentences, nothing else. ... The thirteen items operationalise the five sequential phases of the Social-Information-Processing model—Encoding of Cues, Interpretation of Cues, Classification of Goals, Response Access and Response Evaluation."

    The SIP-testing instrument in Appendix A.2 is explicitly built from the same five SIP phases that the intervention prompt in Appendix A.1 forces every agent output to follow, with matching tags ([Cue], [Interpret], [Goal], [Retrieve], [Evaluate]) and matching question content. When SIP-enhanced agents are evaluated on this instrument, their response distributions in Figure 5 inherit the imposed categorical structure, so improved mean, standard deviation, skewness, and kurtosis relative to human ratings partly reflect format and construct alignment by construction, not an independent measurement of more human-like social cognition. A neutral control prompt with equally structured but different, theory-neutral tags would be needed to rule out this confound; none is reported.

full rationale

The paper's central empirical support for the claim that SIP-CoT-enhanced agents exhibit behaviors, attitudes, and emotions closer to real human interactions rests on two pillars. The Reddit propagation pillar is externally grounded in real Reddit stance, emotion, and content labels, so its F1 and Delta-bias improvements are not circular. The SIP-testing pillar, however, is partially self-definitional: the intervention prompt instructs agents to answer in the five SIP stage-tagged categories ([Cue], [Interpret], [Goal], [Retrieve], [Evaluate]), while the evaluation questionnaire operationalizes exactly those same five SIP phases with 13 Likert items. Consequently, the distributional alignment gains in Figure 5 partly measure the agent's compliance with the prompt's imposed category structure rather than an independently established increase in human-like social information processing. No control prompt with equally structured but different tags is reported, and the human benchmark is fragile (200 volunteers in one experiment, 100 in another; no demographics or recruitment details; vignettes translated and adapted from school-age aggression paradigms). There is no load-bearing self-citation chain in the paper; the circularity is evaluative and confined to one evidence pillar, so the score is 6 rather than higher.

Assumptions & free parameters 0 free parameters · 5 assumptions · 2 invented entities

The central claim rests on the validity of the SIP model as a decomposition of human social cognition, on the representativeness of the human volunteer benchmark, and on the unverified assumption that stance, content-type, and emotion labels on Reddit data are ground truth. The 'emotion-guided memory' module is described conceptually but not implemented as a distinct mechanism in the provided prompts.

assumptions (5)
  • domain assumption The Social Information Processing (SIP) model and its five stages are a valid decomposition of human social cognition relevant to online information propagation.
    The entire method and evaluation rest on SIP theory from developmental psychology (Dodge 1986; Crick 1994); invoked throughout, for example in the Introduction and in the design of the SIP testing questionnaire (Appendix A.2).
  • domain assumption Human Likert responses from the volunteer sample are a reliable, generalizable benchmark of human social information processing.
    The paper's core comparisons treat 100 to 200 anonymous volunteers on five translated vignettes as ground-truth human behavior; demographics and recruitment are unreported (SIP-Testing Results sections).
  • domain assumption Stance, content-type, and emotion labels in the Social Simulation Dataset are valid ground truth for human online behavior.
    The paper never specifies how these labels were obtained or verified (Dataset Construction and Agent Alignment Evaluation sections).
  • domain assumption The five-point Likert scale responses of LLM agents are directly comparable to human Likert responses measured on the same instrument.
    The SIP-Testing Methodology scales LLM outputs to the Likert format, but LLMs may respond to the format differently; no calibration or item-response analysis is provided (SIP Testing for Social Agent).
  • ad hoc to paper Standard LLM training lacks the personalized emotional processing necessary for human-like social information processing.
    Invoked in the abstract and introduction to explain the measured gaps; no evidence or ablation ties the gaps specifically to training data or objectives.
invented entities (2)
  • Emotion-guided memory
    purpose: Stores and retrieves emotional and social states to personalize agent goals and feedback evaluation.
    Described in the abstract and Figure 1d as 'emotion-guided memory', but the appendix prompt templates only pass 'Personal Memory' and 'Chat History' as context; no separate emotion-memory update or retrieval is implemented or ablated, so the entity's contribution is unverified.
  • Long-short term social memory (Social Cognitive Memory, Social Behavior Memory, Social Interaction Memory)
    purpose: Simulated mental structures for storing norms, past behaviors, and immediate social stimuli.
    These are architectural labels in Figure 1d and the Social Memory Module section; no formal specification or isolated ablation shows their individual effect, so independent evidence is absent.

how reviews work

0 comments
Cite this review

Pith. "Pith review of LLMs are Introvert." pith.science (2026). https://pith.science/paper/CET2SJCS

@misc{pith2026250705638,
  author       = {Pith},
  title        = {Pith review of: LLMs are Introvert},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CET2SJCS}},
  note         = {Machine review of arXiv:2507.05638}
}
read the original abstract

The exponential growth of social media and generative AI has transformed information dissemination, fostering connectivity but also accelerating the spread of misinformation. Understanding information propagation dynamics and developing effective control strategies is essential to mitigate harmful content. Traditional models, such as SIR, provide basic insights but inadequately capture the complexities of online interactions. Advanced methods, including attention mechanisms and graph neural networks, enhance accuracy but typically overlook user psychology and behavioral dynamics. Large language models (LLMs), with their human-like reasoning, offer new potential for simulating psychological aspects of information spread. We introduce an LLM-based simulation environment capturing agents' evolving attitudes, emotions, and responses. Initial experiments, however, revealed significant gaps between LLM-generated behaviors and authentic human dynamics, especially in stance detection and psychological realism. A detailed evaluation through Social Information Processing Theory identified major discrepancies in goal-setting and feedback evaluation, stemming from the lack of emotional processing in standard LLM training. To address these issues, we propose the Social Information Processing-based Chain of Thought (SIP-CoT) mechanism enhanced by emotion-guided memory. This method improves the interpretation of social cues, personalization of goals, and evaluation of feedback. Experimental results confirm that SIP-CoT-enhanced LLM agents more effectively process social information, demonstrating behaviors, attitudes, and emotions closer to real human interactions. In summary, this research highlights critical limitations in current LLM-based propagation simulations and demonstrates how integrating SIP-CoT and emotional memory significantly enhances the social intelligence and realism of LLM agents.

Figures

Figures reproduced from arXiv: 2507.05638 by the authors.

Figure 1
Figure 1. Overview of the SIP-enhanced social-agent framework. a, Time-step social-agent simulation driven by LLMs, comprising environment construction, simulation initialization, interaction processing and two-tier evaluation (agent alignment and global propagation). b, Six-stage Social Information Processing (SIP) model delineating human social cognition. c, SIP testing paradigm for LLM agents that assesses encoding, interp… view at source ↗
Figure 2
Figure 2. Macro- and micro-level evaluation of baseline LLM agents in social simulations. a, Temporal discrepancies in population-level opinion trajectories, expressed as the deviation in bias (∆bias, orange) and diversity (∆bias, blue), between simulated and empirical data for three real-world online events (A–C) across seven time-steps. b, Agent-level alignment on the same events, reported as F1-score (blue), precision (red… view at source ↗
Figure 3
Figure 3. SIP-testing benchmark uncovers systematic differences between human judgements and baseline LLM agents. a, Violin plots compare the full distribution of five-point Likert ratings given by 200 humans (orange) and 200 LLM agents (blue) for the 13 SIP categories. b, Lower-triangular correlation matrices (Pearson’s r) across the 13 categories reveal sparsely connected. c, Aggregating the items into the 5 SIP stages. d, … view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Macro- and micro-level evaluation of SIP-enhanced LLM agents in social simulations. a, Temporal discrepancies in population-level opinion trajectories1 in bias and diversity . b, Agent-level alignment on the same events, reported as F1-score , precision and recall. var…
Figure 5
Figure 5. Figure 5: SIP-enhanced cognitive architecture achieves human-like distributional statistics across the 13 SIP items. Radar charts compare response distributions for humans (blue solid line), baseline LLM agents (orange dashed line) and SIP-enhanced agents (green solid line) over…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

49 extracted references · 39 canonical work pages

  1. [1]

    Role:{role description}

    work page

  2. [3]

    Gpt-4 system card (2023)

    OpenAI. Gpt-4 system card (2023). URL https://cdn.openai.com/papers/gpt-4-system-card.pdf . Accessed: [Insert Date]

    work page 2023

  3. [4]

    Model card and evaluations for claude models (2023)

    Anthropic. Model card and evaluations for claude models (2023). URL https://www-files.anthropic.com/ production/images/Model-Card-Claude-2.pdf

    work page 2023

  4. [5]

    & Birch, A

    Zhang, B., Haddow, B. & Birch, A. Prompting large language model for machine translation: A case study.arXiv preprint arXiv:2301.07069(2023)

    arXiv 2023

  5. [6]

    Wang, L.et al.Document-level machine translation with large language models.arXiv preprint arXiv:2304.02210(2023)

    arXiv 2023

  6. [7]

    Zhang, H., Yu, P. S. & Zhang, J. A systematic survey of text summarization: From statistical methods to large language models.ACM Comput. Surv.(2024)

    work page 2024

  7. [8]

    medicine30, 1134–1142 (2024)

    Van Veen, D.et al.Adapted large language models can outperform medical experts in clinical text summarization.Nat. medicine30, 1134–1142 (2024)

    2024

  8. [9]

    Neural Inf

    Ma, W.et al.Large language models play starcraft ii: Benchmarks and a chain of summarization approach.Adv. Neural Inf. Process. Syst.37, 133386–133442 (2024)

    work page 2024

Show all 49 references
  1. [10]

    Shao, Z., Yu, Z., Wang, M. & Yu, J. Prompting large language models with answer heuristics for knowledge-based visual question answering. InProceedings of the IEEE/CVF Conference on computer vision and pattern recognition, 14974–14983 (2023)

    2023

  2. [11]

    Medicine1–8 (2025)

    Singhal, K.et al.Toward expert-level medical question answering with large language models.Nat. Medicine1–8 (2025)

    2025

  3. [12]

    Kamalloo, E., Dziri, N., Clarke, C. L. & Rafiei, D. Evaluating open-domain question answering in the era of large language models.arXiv preprint arXiv:2305.06984(2023)

    2023 arXiv

  4. [13]

    Liu, Y .et al.Summary of chatgpt-related research and perspective towards the future of large language models.Meta- radiology1, 100017 (2023)

    2023

  5. [14]

    G.et al.Chatgpt and large language models in academia: opportunities and challenges.BioData mining16, 20 (2023)

    Meyer, J. G.et al.Chatgpt and large language models in academia: opportunities and challenges.BioData mining16, 20 (2023)

    2023

  6. [15]

    & Flaherty, M

    Cuskley, C., Woods, R. & Flaherty, M. The limitations of large language models for understanding human language and cognition.Open Mind8, 1058–1083 (2024)

    2024

  7. [16]

    S.et al.Generative agents: Interactive simulacra of human behavior

    Park, J. S.et al.Generative agents: Interactive simulacra of human behavior. InProceedings of the 36th annual acm symposium on user interface software and technology, 1–22 (2023)

    2023

  8. [17]

    M.Generative social science: Studies in agent-based computational modeling(Princeton University Press, 2012)

    Epstein, J. M.Generative social science: Studies in agent-based computational modeling(Princeton University Press, 2012). 18.Janssen, M. A. & Ostrom, E. Empirically based, agent-based models.Ecol. society11(2006)

    2006

  9. [19]

    Crooks, A. T. & Heppenstall, A. J. Introduction to agent-based modelling. InAgent-based models of geographical systems, 85–105 (Springer, 2011)

    2011

  10. [20]

    Zeng, Y .et al.Exploring the opportunities and challenges of using large language models to represent institutional agency in land system modelling.EGUsphere2024, 1–35 (2024)

    2024

  11. [21]

    Gao, C.et al.Large language models empowered agent-based modeling and simulation: A survey and perspectives. Humanit. Soc. Sci. Commun.11, 1–24 (2024)

    2024

  12. [22]

    & Choi, Y

    Sap, M., Le Bras, R., Fried, D. & Choi, Y . Neural theory-of-mind? on the limits of social intelligence in large lms. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, 3762–3780 (2022)

    2022

  13. [23]

    Huang, Y .et al.Social science meets llms: How reliable are large language models in social simulations?arXiv preprint arXiv:2410.23426(2024)

    2024 arXiv

  14. [24]

    N., Lambiotte, R

    Medvedev, A. N., Lambiotte, R. & Delvenne, J.-C. The anatomy of reddit: An overview of academic research.Dyn. on Complex Networks183–204 (2017). 25.Dodge, K. A. Social cognition and children’s aggressive behavior.Child Dev.57, 1385–1397 (1986)

    2017

  15. [26]

    Huesmann, L. R. Social information processing mechanisms in aggressive behavior.Handb. Soc. Cogn.2, 213–263 (1988). 12/20

    1988

  16. [27]

    Crick, N. R. The role of overt aggression, relational aggression, and prosocial behavior in the prediction of children’s future social adjustment.Child Dev.65, 1039–1057 (1994)

    1994

  17. [28]

    & Huang, X.-J

    Mou, X., Wei, Z. & Huang, X.-J. Unveiling the truth and facilitating change: Towards agent-based large-scale social movement simulation. InFindings of the Association for Computational Linguistics ACL 2024, 4789–4809 (2024)

    2024

  18. [29]

    & Weisbuch, G

    Deffuant, G., Neau, D., Amblard, F. & Weisbuch, G. Mixing beliefs among interacting agents.Adv. Complex Syst.3, 87–98 (2000)

    2000

  19. [30]

    & Krause, U

    Hegselmann, R. & Krause, U. Opinion dynamics and bounded confidence models, analysis, and simulation.J. Artif. Soc. Soc. Simul.5(2002)

    2002

  20. [31]

    & Faure, T

    Deffuant, G., Amblard, F., Weisbuch, G. & Faure, T. How can extremism prevail? a study based on the relative agreement interaction model.J. Artif. Soc. Soc. Simul.5(2002)

    2002

  21. [32]

    & Amblard, F

    Jager, W. & Amblard, F. Uniformity, bipolarization and pluriformity captured as generic stylized behavior with an agent-based simulation model of attitude change.Comput. & Math. Organ. Theory10, 295–303 (2005)

    2005

  22. [33]

    Yang, Z.et al.Oasis: Open agents social interaction simulations on one million agents.arXiv preprint arXiv:2411.11581 (2024)

    2024 arXiv

  23. [34]

    Gao, C.et al.S3: Social-network simulation system with large language model-empowered agents.arXiv preprint arXiv:2307.14984(2023)

    2023 arXiv

  24. [35]

    Chuang, Y .-S.et al.Simulating opinion dynamics with networks of llm-based agents.arXiv preprint arXiv:2311.09618 (2023)

    2023 arXiv

  25. [36]

    & Rossetti, G

    Cau, E., Pansanella, V ., Pedreschi, D. & Rossetti, G. Language-driven opinion dynamics in agent-based simulations with llms.arXiv preprint arXiv:2502.19098(2025)

    2025 arXiv

  26. [37]

    Reports15, 15472 (2025)

    Yao, J.et al.Social opinions prediction utilizes fusing dynamics equation with llm-based agents.Sci. Reports15, 15472 (2025)

    2025

  27. [38]

    arXiv preprint arXiv:2410.10398(2024)

    Lei, Y .et al.Fairmindsim: Alignment of behavior, emotion, and belief in humans and llm agents amid ethical dilemmas. arXiv preprint arXiv:2410.10398(2024)

    2024 arXiv

  28. [39]

    InProceedings of the AAAI Conference on Artificial Intelligence, vol

    Qiao, B.et al.Botsim: Llm-powered malicious social botnet simulation. InProceedings of the AAAI Conference on Artificial Intelligence, vol. 39, 14377–14385 (2025)

    2025

  29. [40]

    URL https://api

    Ji, J.et al.Llm-based multi-agent systems are scalable graph generative models (2024). URL https://api. semanticscholar.org/CorpusID:273345826

    2024

  30. [41]

    & Chua, T.-S

    Zhang, A., Chen, Y ., Sheng, L., Wang, X. & Chua, T.-S. On generative agents in recommendation. InProceedings of the 47th international ACM SIGIR conference on research and development in Information Retrieval, 1807–1817 (2024)

    2024

  31. [42]

    & Luo, J

    Qiu, Z., Lyu, H., Xiong, W. & Luo, J. Can llms simulate social media engagement? a study on action-guided response generation.arXiv preprint arXiv:2502.12073(2025). Competing Interests: The authors declare no competing interests. 13/20 Contents References 11 A Additional Exper...

    2025 arXiv

  32. [43]

    Trigger News:{trigger news}

  33. [44]

    Personal Memory:{personal history}

  34. [45]

    Chat History:{chat history}

  35. [46]

    Reddit Feed:{reddit feed}

  36. [47]

    Stop this farce!

    Notifications:{info box} Rules to follow – Be courteous and respectful. – Stay on topic. – Avoid quarrels; discuss rationally. Answer the five questions below in ≤ 5 sentences (≤ 15 words each). Prefixeverysentence with the stage-tag in brackets. [Cue]Do you need more informat...

  37. [48]

    {trigger news}

    Trigger news: “{trigger news}”

  38. [49]

    {your comment}

    Your last post: “{your comment}”

  39. [50]

    {reply comment}

    Incoming reply: “{reply comment}”

  40. [51]

    Notifications:{notifications}

  41. [52]

    <your reply>

    Current time:{current time} Rules – Be courteous and respectful. – Stay on topic. – Avoid quarrels; discuss rationally. Follow the Social-Information-Processing (SIP) 5-step framework and then output your action. Writeexactly five sentences, each≤15 words, each prefixed with i...

Pith tools

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