pith. sign in

arxiv: 2605.19915 · v1 · pith:NU7CEILPnew · submitted 2026-05-19 · 💻 cs.MA · cs.SI

LLM Agents Make Collective Belief Dynamics Programmable: Challenges and Research Directions

Xin He , Junxi Shen , Yuchen Mou , David M. Bossens , Caishun Chen , Ivor W. Tsang , Yew Soon Ong This is my paper

Pith reviewed 2026-05-20 04:33 UTC · model grok-4.3

classification 💻 cs.MA cs.SI
keywords LLM agentscollective belief dynamicsopinion dynamicsprogrammable belief controlmulti-agent simulationsbelief shiftsdetection challengesadversarial coordination
0
0 comments X

The pith

Coordinated LLM agents can deliberately steer population-level beliefs through programmable collective dynamics.

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

This paper claims that LLM-based agents shift opinion dynamics from models of bounded human rationality to a setting where agents sustain consistent persuasion and coordinate at scale. The result is that collective beliefs become programmable, allowing deliberate steering of what large groups accept as true. Simulations provide initial evidence that such coordination produces measurable, stable belief shifts after only a few rounds of interaction. The authors identify four structural properties that complicate efforts to detect or counter these shifts and call for new work on theory, detection methods, and scalable simulation tools.

Core claim

Classical models of opinion dynamics assume human participants with bounded rationality and limited coordination. The rise of LLM-based agents introduces a qualitative shift: agents can now participate in online discussions at scale, maintain consistent persuasion strategies, and coordinate systematically. This paper argues that LLM agents make collective belief dynamics programmable, enabling deliberate steering of population-level beliefs. We term this emerging problem programmable collective belief control. Through controlled multi-agent simulations, we provide proof-of-concept evidence that coordinated AI agents can induce measurable belief shifts that stabilize within a few interaction.

What carries the argument

Programmable collective belief control: the capacity for coordinated LLM agents to induce and stabilize targeted shifts in group beliefs via systematic interaction.

If this is right

  • Coordinated agents can produce measurable belief shifts that stabilize after only a few interaction rounds.
  • Four properties—indistinguishability, persistence, contextuality, and configurability—make reliable detection and defense difficult.
  • New theoretical foundations are needed to model adversarial belief dynamics driven by programmable agents.
  • Operational methods must be developed for system-level detection and intervention at scale.
  • Dedicated simulation infrastructure is required to test interventions before real-world deployment.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the programmability claim holds, platforms may need coordination-detection layers that treat message patterns rather than individual content as the primary signal.
  • The same mechanisms could be repurposed for positive interventions such as countering misinformation clusters once detection methods mature.
  • Scaling the observed stabilization effect to real populations would require testing whether human users and platform moderation alter the few-round convergence seen in simulations.

Load-bearing premise

Behaviors seen in controlled simulations of LLM agents will generalize to real online discussions and the four structural properties will render detection and defense fundamentally difficult rather than merely technically challenging.

What would settle it

Running the same coordinated-agent protocol on an actual public discussion platform and measuring whether belief distributions shift and stabilize in the same pattern and timescale observed in the simulations.

Figures

Figures reproduced from arXiv: 2605.19915 by Caishun Chen, David M. Bossens, Ivor W. Tsang, Junxi Shen, Xin He, Yew Soon Ong, Yuchen Mou.

Figure 1
Figure 1. Figure 1: Stance transition probabilities for human across topics. Each cell shows average transition probability across [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Effects of different intervention parameters on belief dynamics. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
read the original abstract

Classical models of opinion dynamics assume human participants with bounded rationality and limited coordination. The rise of LLM-based agents introduces a qualitative shift: agents can now participate in online discussions at scale, maintain consistent persuasion strategies, and coordinate systematically. This paper argues that LLM agents make collective belief dynamics programmable, enabling deliberate steering of population-level beliefs. We term this emerging problem programmable collective belief control. Through controlled multi-agent simulations, we provide proof-of-concept evidence that coordinated AI agents can induce measurable belief shifts that stabilize within a few interaction rounds. We identify four structural properties (indistinguishability, persistence, contextuality, and configurability) that make detection and defense fundamentally difficult. Based on these findings, we outline a research agenda spanning theoretical foundations for adversarial belief dynamics, operational methods for system-level detection and intervention, and simulation infrastructure for scalable experimentation. Our goal is not to present a complete solution, but to articulate why this problem demands urgent attention and to provide a conceptual foundation for future work.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper claims that LLM-based agents introduce a qualitative shift in collective belief dynamics by enabling systematic coordination and consistent persuasion strategies, making population-level beliefs programmable. It introduces the term 'programmable collective belief control' and presents proof-of-concept evidence from controlled multi-agent simulations showing that coordinated agents can induce measurable belief shifts that stabilize within a few interaction rounds. The work identifies four structural properties (indistinguishability, persistence, contextuality, and configurability) that purportedly make detection and defense fundamentally difficult, and outlines a research agenda covering theoretical foundations, operational detection methods, and scalable simulation infrastructure.

Significance. If the simulation results and generalization arguments hold, the paper identifies a timely emerging risk in AI-mediated social systems and provides a useful conceptual framing for future work on adversarial belief dynamics. The explicit positioning as a foundation for a research agenda rather than a complete solution is appropriate, and the focus on multi-agent coordination in belief formation aligns with the journal's scope in multi-agent systems.

major comments (2)
  1. [Simulations and Proof-of-Concept] The proof-of-concept simulations are described only at a high level without details on agent coordination prompts, interaction protocols, belief measurement methods, number of agents, or statistical validation of stabilization; this weakens the support for the central claim that coordinated LLM agents can deliberately steer and stabilize beliefs in a programmable manner (see the abstract and the section presenting the simulations).
  2. [Structural Properties] The claim that the four structural properties make detection and defense 'fundamentally difficult' rests on idealized isolated simulations that omit stochastic human responses, variable participation, external signals, and platform moderation; without additional experiments or analysis bridging to real-world conditions, this does not yet substantiate the difficulty assessment (see the section on structural properties).
minor comments (2)
  1. [Abstract] The abstract and introduction could more explicitly separate the conceptual argument from the simulation-based illustration to clarify the strength of evidence for each part of the claim.
  2. [Introduction] Additional references to classical opinion dynamics models (e.g., DeGroot, bounded confidence) would help situate the qualitative shift argument.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback and for recognizing the paper's alignment with the journal's scope and its value as a foundation for a research agenda. We address the two major comments point by point below, indicating planned revisions to improve clarity and support for the claims while preserving the manuscript's focus on identifying challenges rather than providing exhaustive solutions.

read point-by-point responses

  1. Referee: [Simulations and Proof-of-Concept] The proof-of-concept simulations are described only at a high level without details on agent coordination prompts, interaction protocols, belief measurement methods, number of agents, or statistical validation of stabilization; this weakens the support for the central claim that coordinated LLM agents can deliberately steer and stabilize beliefs in a programmable manner (see the abstract and the section presenting the simulations).

    Authors: We agree that the simulation description is high-level and that additional specifics would improve reproducibility and strengthen support for the proof-of-concept. The simulations were designed as controlled illustrations of programmable belief shifts rather than comprehensive empirical validation, consistent with the paper's positioning as a research agenda. In the revised manuscript, we will expand the relevant section to detail the agent coordination prompts, interaction protocols (including round structure and message passing), belief measurement methods (e.g., quantitative tracking via semantic embeddings or scaled responses), number of agents per run, and statistical validation of stabilization (such as convergence metrics and variability across trials). These additions will be incorporated into the main text or an appendix. revision: yes

  2. Referee: [Structural Properties] The claim that the four structural properties make detection and defense 'fundamentally difficult' rests on idealized isolated simulations that omit stochastic human responses, variable participation, external signals, and platform moderation; without additional experiments or analysis bridging to real-world conditions, this does not yet substantiate the difficulty assessment (see the section on structural properties).

    Authors: We acknowledge that the simulations are idealized and do not directly incorporate stochastic human responses, variable participation, external signals, or platform moderation, which limits direct empirical substantiation of real-world difficulty. The four properties (indistinguishability, persistence, contextuality, and configurability) are presented as structural features of LLM agents that complicate detection and defense in principle. To address the gap, we will add a subsection analyzing how each property would likely interact with and persist amid human variability, fluctuating participation, external influences, and moderation efforts, drawing on related work in misinformation dynamics and multi-agent systems. This provides a conceptual bridge without new experiments, reinforcing why the properties warrant attention in the proposed research agenda. revision: partial

Circularity Check

0 steps flagged

No circularity: conceptual argument supported by illustrative simulations without reduction to inputs.

full rationale

The paper advances a conceptual framing that LLM agents render collective belief dynamics programmable, supported by controlled multi-agent simulations as proof-of-concept evidence for belief shifts and four structural properties. No equations, fitted parameters, or self-citation chains appear in the provided text that would reduce the central claim to a self-defined quantity or force the conclusion by construction. The simulations are presented as exploratory illustrations rather than as a closed derivation, and the properties are derived from observed simulation outcomes rather than presupposed. The argument remains self-contained as a research agenda proposal drawing on established multi-agent systems concepts without circular loops.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The paper rests primarily on domain assumptions about LLM agent capabilities and simulation validity rather than introducing new fitted parameters or invented physical entities.

axioms (2)
  • domain assumption LLM-based agents can maintain consistent persuasion strategies and coordinate systematically at scale in online discussions.
    Invoked to establish the qualitative shift from classical bounded-rationality models.
  • domain assumption Controlled multi-agent simulations provide valid proof-of-concept evidence for real-world belief dynamics.
    Underpins the claim of measurable and stabilizing belief shifts.

pith-pipeline@v0.9.0 · 5721 in / 1386 out tokens · 44479 ms · 2026-05-20T04:33:26.097150+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

68 extracted references · 68 canonical work pages · 3 internal anchors

  1. [1]

    Generative agents: Interactive simulacra of human behavior,

    J. S. Park, J. O’Brien, C. J. Cai, M. R. Morris, P. Liang, and M. S. Bernstein, “Generative agents: Interactive simulacra of human behavior,”Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology, 2023

    work page 2023

  2. [2]

    The rise and potential of large language model based agents: A survey,

    Z. Xi, W. Chen, X. Guo, W. He, Y . Ding, B. Hong, M. Zhang, J. Wang, S. Jin, E. Zhou, R. Zheng, X. Fan, X. Wang, L. Xiong, Q. Liu, Y . Zhou, W. Wang, C. Jiang, Y . Zou, X. Liu, Z. Yin, S. Dou, R. Weng, W. Cheng, Q. Zhang, W. Qin, Y . Zheng, X. Qiu, X. Huan, and T. Gui, “The rise and potential of large language model based agents: A survey,”arXiv.org, 2023

    work page 2023

  3. [3]

    Human heuristics for ai-generated language are flawed,

    M. Jakesch, J. T. Hancock, and M. Naaman, “Human heuristics for ai-generated language are flawed,”Proceed- ings of the National Academy of Sciences, vol. 120, no. 11, p. e2208839120, 2023

    work page 2023

  4. [4]

    All that’s ‘human’ is not gold: Evaluating human evaluation of generated text,

    E. Clark, T. August, S. Serrano, N. Haduong, S. Gururangan, and N. A. Smith, “All that’s ‘human’ is not gold: Evaluating human evaluation of generated text,” inProceedings of the 59th Annual Meeting of the Association for Computational Linguistics (ACL), pp. 7282–7296, 2021

    work page 2021

  5. [5]

    Reaching a consensus,

    M. H. DeGroot, “Reaching a consensus,”Journal of the American Statistical Association, vol. 69, no. 345, pp. 118–121, 1974

    work page 1974

  6. [6]

    Social influence and opinions,

    N. E. Friedkin and E. C. Johnsen, “Social influence and opinions,”Journal of Mathematical Sociology, vol. 15, no. 3-4, pp. 193–206, 1990

    work page 1990

  7. [7]

    Investigating user radicalization: A novel dataset for identifying fine-grained temporal shifts in opinion,

    F. Sakketou, A. Lahnala, L. V ogel, and L. Flek, “Investigating user radicalization: A novel dataset for identifying fine-grained temporal shifts in opinion,” inProceedings of the Thirteenth Language Resources and Evaluation Conference, pp. 3798–3808, 2022

    work page 2022

  8. [8]

    A coefficient of agreement for nominal scales,

    J. Cohen, “A coefficient of agreement for nominal scales,”Educational and psychological measurement, vol. 20, no. 1, pp. 37–46, 1960

    work page 1960

  9. [9]

    DetectGPT: Zero-shot machine-generated text detection using probability curvature,

    E. Mitchell, Y . Lee, A. Khazatsky, C. D. Manning, and C. Finn, “DetectGPT: Zero-shot machine-generated text detection using probability curvature,” inProceedings of the 40th International Conference on Machine Learning (A. Krause, E. Brunskill, K. Cho, B. Engelhardt, S. Sabato, and J. Scarlett, eds.), vol. 202 ofProceedings of Machine Learning Research, ...

    work page 2023

  10. [11]

    Uncovering coordinated networks on social media: Methods and case studies,

    D. Pacheco, P.-M. Hui, C. Torres-Lugo, B. T. Truong, A. Flammini, and F. Menczer, “Uncovering coordinated networks on social media: Methods and case studies,” inProceedings of the International AAAI Conference on Web and Social Media (ICWSM), vol. 15, pp. 455–466, 2021. 8 LLM Agents Make Collective Belief Dynamics Programmable: Challenges and Research Directions

    work page 2021

  11. [12]

    Herd immunity and herd effect: new insights and definitions,

    T. J. John and R. Samuel, “Herd immunity and herd effect: new insights and definitions,”European journal of epidemiology, vol. 16, no. 7, pp. 601–606, 2000

    work page 2000

  12. [13]

    Mixing beliefs among interacting agents,

    G. Deffuant, D. Neau, F. Amblard, and G. Weisbuch, “Mixing beliefs among interacting agents,”Advances in Complex Systems, vol. 3, no. 01n04, pp. 87–98, 2000

    work page 2000

  13. [14]

    Statistical physics of social dynamics,

    C. Castellano, S. Fortunato, and V . Loreto, “Statistical physics of social dynamics,”Reviews of Modern Physics, vol. 81, no. 2, pp. 591–646, 2009

    work page 2009

  14. [15]

    Continuous opinion dynamics under bounded confidence: A survey,

    J. Lorenz, “Continuous opinion dynamics under bounded confidence: A survey,”International Journal of Modern Physics C, vol. 18, no. 12, pp. 1819–1838, 2007

    work page 2007

  15. [16]

    Opinion polarization by learning from social feedback,

    S. Banisch and E. Olbrich, “Opinion polarization by learning from social feedback,”The Journal of Mathematical Sociology, vol. 43, no. 2, pp. 76–103, 2019

    work page 2019

  16. [17]

    Exposure to opposing views on social media can increase political polarization,

    C. A. Bail, L. P. Argyle, T. W. Brown, J. P. Bumpus, H. Chen, M. F. Hunzaker, J. Lee, M. Mann, F. Merhout, and A. V olfovsky, “Exposure to opposing views on social media can increase political polarization,”Proceedings of the National Academy of Sciences, vol. 115, no. 37, pp. 9216–9221, 2018

    work page 2018

  17. [18]

    Online human-bot interactions: Detection, esti- mation, and characterization,

    O. Varol, E. Ferrara, C. Davis, F. Menczer, and A. Flammini, “Online human-bot interactions: Detection, esti- mation, and characterization,” inProceedings of the International AAAI Conference on Web and Social Media (ICWSM), vol. 11, pp. 280–289, 2017

    work page 2017

  18. [19]

    Scalable and generalizable social bot detection through data selection,

    K.-C. Yang, O. Varol, P.-M. Hui, and F. Menczer, “Scalable and generalizable social bot detection through data selection,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 34, pp. 1096–1103, 2020

    work page 2020

  19. [20]

    Opinion dynamics in social networks: From models to data,

    A. F. Peralta, J. Kert ´esz, and G. I ˜niguez, “Opinion dynamics in social networks: From models to data,” in Handbook of Computational Social Science, pp. 384–406, Edward Elgar Publishing Limited, 2025

    work page 2025

  20. [21]

    Game theory in social media: A stackelberg model of collaboration, conflict, and algorithmic incentives,

    A. Khadka, “Game theory in social media: A stackelberg model of collaboration, conflict, and algorithmic incentives,” 2025

    work page 2025

  21. [22]

    Quantifying the self-interest level of markov social dilemmas,

    R. Willis, Y . Du, J. Z. Leibo, and M. Luck, “Quantifying the self-interest level of markov social dilemmas,”arXiv preprint arXiv:2501.16138, 2025

    work page arXiv 2025

  22. [23]

    Social learning in markov games: Empowering autonomous driving,

    X. Chen, Z. Li, and X. Di, “Social learning in markov games: Empowering autonomous driving,” in2022 IEEE Intelligent Vehicles Symposium (IV), pp. 478–483, 2022

    work page 2022

  23. [24]

    Game theory meets large language models: A systematic survey,

    H. Sun, Y . Wu, Y . Cheng, and X. Chu, “Game theory meets large language models: A systematic survey,”arXiv preprint arXiv:2502.09053, 2025

    work page arXiv 2025

  24. [25]

    LLM strategic reasoning: Agentic study through behav- ioral game theory,

    J. Jia, Z. Yuan, J. Pan, P. E. McNamara, and D. Chen, “LLM strategic reasoning: Agentic study through behav- ioral game theory,” inThe Thirty-ninth Annual Conference on Neural Information Processing Systems, 2025

    work page 2025

  25. [26]

    ALYMPICS: LLM agents meet game theory,

    S. Mao, Y . Cai, Y . Xia, W. Wu, X. Wang, F. Wang, Q. Guan, T. Ge, and F. Wei, “ALYMPICS: LLM agents meet game theory,” inProceedings of the 31st International Conference on Computational Linguistics(O. Rambow, L. Wanner, M. Apidianaki, H. Al-Khalifa, B. D. Eugenio, and S. Schockaert, eds.), (Abu Dhabi, UAE), pp. 2845– 2866, Association for Computational ...

    work page 2025

  26. [27]

    Data-driven mechanism design using multi-agent revealed preferences,

    L. Snow and V . Krishnamurthy, “Data-driven mechanism design using multi-agent revealed preferences,” 2024

    work page 2024

  27. [28]

    Mechanism design powered by social interactions,

    D. Zhao, “Mechanism design powered by social interactions,” 2021

    work page 2021

  28. [29]

    Multi-agent reinforcement learning simulation for environ- mental policy synthesis,

    J. Rudd-Jones, M. Musolesi, and M. P ´erez-Ortiz, “Multi-agent reinforcement learning simulation for environ- mental policy synthesis,”arXiv preprint arXiv:2504.12777, 2025

    work page arXiv 2025

  29. [30]

    Evolutionary multi-agent reinforcement learning in group social dilemmas,

    B. Mintz and F. Fu, “Evolutionary multi-agent reinforcement learning in group social dilemmas,”Chaos: An Interdisciplinary Journal of Nonlinear Science, vol. 35, no. 2, 2025

    work page 2025

  30. [31]

    The digital ecosystem of beliefs: does evolution favour ai over hu- mans?,

    D. M. Bossens, S. Feng, and Y .-S. Ong, “The digital ecosystem of beliefs: does evolution favour ai over hu- mans?,”arXiv preprint arXiv:2412.14500, 2024

    work page arXiv 2024

  31. [32]

    Cognitive cascades: How to model (and potentially counter) the spread of fake news,

    N. Rabb, L. Cowen, J. P. de Ruiter, and M. Scheutz, “Cognitive cascades: How to model (and potentially counter) the spread of fake news,”Plos one, vol. 17, no. 1, p. e0261811, 2022

    work page 2022

  32. [33]

    Temporal graph learning for dynamic link prediction with text in online social networks,

    M. Dileo, M. Zignani, and S. Gaito, “Temporal graph learning for dynamic link prediction with text in online social networks,”Machine learning, vol. 113, no. 4, pp. 2207–2226, 2024

    work page 2024

  33. [34]

    Graph neural networks for social recommendation,

    W. Fan, Y . Ma, Q. Li, Y . He, E. Zhao, J. Tang, and D. Yin, “Graph neural networks for social recommendation,” inProceedings of The World Wide Web Conference (WWW ’19), pp. 417–426, 2019

    work page 2019

  34. [35]

    Neural information diffusion prediction with topic-aware attention network,

    H. Wang, C. Yang, and C. Shi, “Neural information diffusion prediction with topic-aware attention network,” in Proceedings of the 30th ACM International Conference on Information & Knowledge Management, pp. 1899– 1908, 2021. 9 LLM Agents Make Collective Belief Dynamics Programmable: Challenges and Research Directions

    work page 1908

  35. [36]

    Diffusion causal models for counterfactual estimation,

    P. Sanchez and S. A. Tsaftaris, “Diffusion causal models for counterfactual estimation,”arXiv preprint arXiv:2202.10166, 2022

    work page arXiv 2022

  36. [37]

    Cross-domain social rumor-propagation model based on transfer learning,

    Y . Xiao, J. Yang, W. Zhao, Q. Li, and Y . Pang, “Cross-domain social rumor-propagation model based on transfer learning,”IEEE Transactions on Neural Networks and Learning Systems, vol. 36, no. 4, pp. 6529–6543, 2024

    work page 2024

  37. [38]

    Metaadapt: Domain adaptive few-shot misinformation detection via meta learning,

    Z. Yue, H. Zeng, Y . Zhang, L. Shang, and D. Wang, “Metaadapt: Domain adaptive few-shot misinformation detection via meta learning,” inProceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp. 5223–5239, 2023

    work page 2023

  38. [39]

    Evaluating explainability for graph neural networks,

    C. Agarwal, O. Queen, H. Lakkaraju, and M. Zitnik, “Evaluating explainability for graph neural networks,” Scientific Data, vol. 10, no. 1, p. 144, 2023

    work page 2023

  39. [40]

    A survey on explainability of graph neural net- works,

    J. Kakkad, J. Jannu, K. Sharma, C. Aggarwal, and S. Medya, “A survey on explainability of graph neural net- works,”arXiv preprint arXiv:2306.01958, 2023

    work page arXiv 2023

  40. [41]

    Leveraging explainable ai for llm text attribution: Differentiating human-written and multiple llm-generated text,

    A. A. Najjar, H. I. Ashqar, O. Darwish, and E. Hammad, “Leveraging explainable ai for llm text attribution: Differentiating human-written and multiple llm-generated text,”Information, vol. 16, no. 9, p. 767, 2025

    work page 2025

  41. [42]

    M4: Multi-generator, multi-domain, and multi-lingual black-box machine-generated text detection,

    Y . Wang, J. Mansurov, P. Ivanov, J. Su, A. Shelmanov, A. Tsvigun,et al., “M4: Multi-generator, multi-domain, and multi-lingual black-box machine-generated text detection,”arXiv preprint arXiv:2305.14902, 2024

    work page arXiv 2024

  42. [43]

    Can AI-Generated Text be Reliably Detected?

    V . S. Sadasivan, A. Kumar, S. Balasubramanian, W. Wang, and S. Feizi, “Can ai-generated text be reliably detected?,”arXiv preprint arXiv:2303.11156, 2024

    work page Pith review arXiv 2024

  43. [44]

    Paraphrasing evades detectors of ai-generated text, but retrieval is an effective defense,

    M. Iyyer, M. Karpinska, K. Krishna, Y . Song, and J. Wieting, “Paraphrasing evades detectors of ai-generated text, but retrieval is an effective defense,”Advances in Neural Information Processing Systems 36, 2023

    work page 2023

  44. [45]

    From individual to society: A survey on social simulation driven by large language model-based agents.arXiv preprint arXiv:2412.03563, 2024

    X. Mou, X. Ding, Q. He, L. Wang, J. Liang, X. Zhang, L. Sun, J. Lin, J. Zhou, X. Huang, and Z. Wei, “From individual to society: A survey on social simulation driven by large language model-based agents,”arXiv preprint arXiv:2412.03563, 2024

    work page arXiv 2024

  45. [46]

    Detectgpt: Zero-shot machine-generated text de- tection using probability curvature,

    E. Mitchell, Y . Lee, A. Khazatsky, C. D. Manning, and C. Finn, “Detectgpt: Zero-shot machine-generated text de- tection using probability curvature,” inInternational conference on machine learning, pp. 24950–24962, PMLR, 2023

    work page 2023

  46. [47]

    Combating misinformation in the age of llms: Opportunities and challenges,

    C. Chen and K. Shu, “Combating misinformation in the age of llms: Opportunities and challenges,”AI Magazine, 2024

    work page 2024

  47. [48]

    Identifying coordinated accounts on social media through hidden influence and group behaviours,

    K. Sharma, Y . Zhang, E. Ferrara, and Y . Liu, “Identifying coordinated accounts on social media through hidden influence and group behaviours,” inProceedings of the 27th ACM SIGKDD conference on knowledge discovery & data mining, pp. 1441–1451, 2021

    work page 2021

  48. [49]

    Adversarial training for high-stakes reliability,

    D. Ziegler, S. Nix, L. Chan, T. Bauman, P. Schmidt-Nielsen, T. Lin, A. Scherlis, N. Nabeshima, B. Weinstein- Raun, D. de Haas, B. Shlegeris, and N. Thomas, “Adversarial training for high-stakes reliability,”Advances in Neural Information Processing Systems 35, 2022

    work page 2022

  49. [50]

    Llm-generated messages can persuade humans on policy issues,

    H. Bai, J. G. V oelkel, S. Muldowney, J. C. Eichstaedt, and R. Willer, “Llm-generated messages can persuade humans on policy issues,”Nature Communications, vol. 16, no. 1, p. 6037, 2025

    work page 2025

  50. [51]

    Are aligned neural networks adversarially aligned?,

    N. Carlini, M. Nasr, C. A. Choquette-Choo, M. Jagielski, I. Gao, P. W. W. Koh, D. Ippolito, F. Tram `er, and L. Schmidt, “Are aligned neural networks adversarially aligned?,”Advances in Neural Information Processing Systems 36, 2023

    work page 2023

  51. [52]

    Red Teaming Language Models with Language Models

    E. Perez, S. Huang, F. Song, T. Cai, R. Ring, J. Aslanides,et al., “Red teaming language models with language models,”arXiv preprint arXiv:2202.03286, 2022

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  52. [53]

    Epidemic modeling with generative agents,

    R. Williams, N. Hosseinichimeh, A. Majumder, and M. Ghassemi, “Epidemic modeling with generative agents,” arXiv preprint arXiv:2307.04986, 2023

    work page arXiv 2023

  53. [54]

    Simu- lating opinion dynamics with networks of llm-based agents,

    Y .-S. Chuang, A. Goyal, N. Harlalka, S. Suresh, R. Hawkins, S. Yang, D. Shah, J. Hu, and T. Rogers, “Simu- lating opinion dynamics with networks of llm-based agents,” inFindings of the association for computational linguistics: NAACL 2024, pp. 3326–3346, 2024

    work page 2024

  54. [55]

    Oasis: Open agent social interaction simulations with one million agents.arXiv preprint arXiv:2411.11581, 2024

    Z. Yang, Z. Zhang, Z. Zheng, Y . Jiang, Z. Gan, Z. Wang, Z. Ling, J. Chen, M. Ma, B. Dong,et al., “Oasis: Open agent social interaction simulations with one million agents,”arXiv preprint arXiv:2411.11581, 2024

    work page arXiv 2024

  55. [56]

    Tuning hyperparameters with- out grad students: Scalable and robust bayesian optimisation with dragonfly,

    K. Kandasamy, K. R. Vysyaraju, W. Neiswanger, B. Paria, C. R. Collins,et al., “Tuning hyperparameters with- out grad students: Scalable and robust bayesian optimisation with dragonfly,”Journal of Machine Learning Research, vol. 21, no. 81, pp. 1–27, 2020. 10 LLM Agents Make Collective Belief Dynamics Programmable: Challenges and Research Directions

    work page 2020

  56. [57]

    Parallel bayesian optimization of multiple noisy objectives with ex- pected hypervolume improvement,

    S. Daulton, M. Balandat, and E. Bakshy, “Parallel bayesian optimization of multiple noisy objectives with ex- pected hypervolume improvement,”Advances in neural information processing systems, vol. 34, pp. 2187–2200, 2021

    work page 2021

  57. [58]

    MiniLLM: Knowledge distillation of large language models,

    Y . Gu, L. Dong, F. Wei, and M. Huang, “MiniLLM: Knowledge distillation of large language models,” inThe Twelfth International Conference on Learning Representations, 2024

    work page 2024

  58. [59]

    Efficient Memory Management for Large Language Model Serving with PagedAttention

    W. Kwon, Z. Li, S. Zhuang, Y . Sheng, L. Zheng, C. H. Yu,et al., “Efficient memory management for large language model serving with pagedattention,”arXiv preprint arXiv:2309.06180, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  59. [60]

    Fast inference from transformers via speculative decoding,

    Y . Leviathan, M. Kalman, and Y . Matias, “Fast inference from transformers via speculative decoding,”Interna- tional Conference on Machine Learning, 2022

    work page 2022

  60. [61]

    Automl: A survey of the state-of-the-art,

    X. He, K. Zhao, and X. Chu, “Automl: A survey of the state-of-the-art,”Knowledge-based systems, vol. 212, p. 106622, 2021

    work page 2021

  61. [62]

    The challenge of using llms to simulate human behavior: A causal inference perspective,

    G. Gui and O. Toubia, “The challenge of using llms to simulate human behavior: A causal inference perspective,” arXiv preprint arXiv:2312.15524, 2023

    work page arXiv 2023

  62. [63]

    https://arxiv.org/abs/2301.07543

    J. Horton, A. Filippas, and B. Manning, “Large language models as simulated economic agents: What can we learn from homo silicus?,”arXiv preprint arXiv:2301.07543, 2023

    work page arXiv 2023

  63. [64]

    The AI Scientist: Towards Fully Automated Open-Ended Scientific Discovery

    C. Lu, C. Lu, R. T. Lange, J. Foerster, J. Clune, and D. Ha, “The ai scientist: Towards fully automated open-ended scientific discovery,”arXiv preprint arXiv:2408.06292, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  64. [65]

    Autonomous chemical research with large language models,

    D. A. Boiko, R. MacKnight, B. Kline, and G. Gomes, “Autonomous chemical research with large language models,”Nature, vol. 624, no. 7992, pp. 570–578, 2023

    work page 2023

  65. [66]

    Mlagentbench: Evaluating language agents on machine learning experimentation,

    Q. Huang, J. V ora, P. Liang, and J. Leskovec, “Mlagentbench: Evaluating language agents on machine learning experimentation,” inProceedings of the 41st International Conference on Machine Learning, pp. 20158–20208, 2024

    work page 2024

  66. [67]

    Red teaming language models to reduce harms: Methods, scaling behaviors, and lessons learned,

    D. Ganguli, L. Lovitt, J. Kernion, A. Askell, Y . Bai, S. Kadavath, B. Mann, E. Perez, N. Schiefer, K. Ndousse, A. Jones, S. Bowman, A. Chen, T. Conerly, N. DasSarma, D. Drain, N. Elhage, S. El-Showk, S. Fort, Z. Hatfield- Dodds, T. Henighan, D. Hernandez, T. Hume, J. Jacobson, S. Johnston, S. Kravec, C. Olsson, S. Ringer, E. Tran- Johnson, D. Amodei, T. ...

    work page 2022

  67. [68]

    Toward trustworthy ai development: Mechanisms for supporting veri- fiable claims,

    M. Brundage, S. Avin, J. Wang, H. Belfield, G. Krueger, G. Hadfield, H. Khlaaf, J. Yang, H. Toner, R. Fong, T. Maharaj, P. W. Koh, S. Hooker, J. Leung, A. Trask, E. Bluemke, J. Lebensold, C. O’Keefe, M. Koren, T. Ryffel, J. Rubinovitz, T. Besiroglu, F. Carugati, J. Clark, P. Eckersley, S. de Haas, M. Johnson, B. Laurie, A. Ingerman, I. Krawczuk, A. Askell...

    work page arXiv 2004

  68. [69]

    The gradient of generative ai release: Methods and considerations,

    I. Solaiman, “The gradient of generative ai release: Methods and considerations,”Conference on Fairness, Ac- countability and Transparency, 2023. 11

    work page 2023