Pith. sign in

REVIEW 3 major objections 5 minor 28 references

Validating Generative Agent-Based Models of Social Norm Enforcement: From Replication to Novel Predictions

T0 review · 3 major / 5 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 Personas and theory of mind let LLM agents reproduce, then extend, human norm-enforcement experiments.

desk verdict A real validation effort with honest partial replication, but the load-bearing architecture claim sits on an unreported persona parameter that must be released. read the letter →

arxiv 2507.22049 v1 pith:AHYQXZAX submitted 2025-07-29 cs.MA

Logan Cross , Nick Haber , Daniel L.K. Yamins This is my paper
classification cs.MA
keywords generativeagent-basedmodelingthird-partypunishmentpublicgoodsgametheoryofmindcostlysignalingsocialnormenforcementLLMagentscooperation
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to establish a two-stage recipe for making generative LLM agents scientifically useful in the social sciences: first, find the minimal set of cognitive components that lets agents reproduce, with the same statistical tests, known behavioral effects from human experiments on social norm enforcement; second, keep the validated architecture and change only the environment to produce novel, testable predictions. It argues that persona-based individual differences and theory of mind are both necessary to replicate third-party punishment as a costly signal of trustworthiness, and that an additional strategic-reflection component is needed to replicate the extra cooperation produced by gossip plus ostracism. If the recipe works, generative agent-based modeling becomes a way to disambiguate competing theories and to preview interventions before running expensive human studies.

What carries the argument

The load-bearing mechanism is the modular cognitive-component architecture: agents get prompts that add components one at a time, and only components that improve replication of human effects are kept. The components are persona (stable individual differences in traits and demographic background), theory of mind (explicit reasoning about what interaction partners are like and how they may respond), and strategic reflection (explicit long-term payoff maximization). The architecture is validated by running the exact human experimental protocol as natural-language prompts and comparing agent behavior to published human data with identical statistical tests. The minimal validated architecture, persona plus theory of mind for Study 1 and with strategic reflection added for Study 2, is then held fixed while the environment is varied to generate novel predictions.

What would settle it

Reanalyze the agent-level data from Study 2 with group membership and round as random effects; if the gossip-versus-basic and gossip-with-ostracism-versus-gossip differences are no longer significant, the architecture comparison that justifies adding strategic reflection fails.

Watch

Extended reading notes

Core claim

The central claim is that a systematically validated generative agent architecture can both match known human results and generate genuinely new predictions about social norm enforcement. Using the third-party punishment trust game from Jordan et al. (2016), the authors show that agents with only observation and situation assessment punish uniformly; adding persona prompts produces individual variation, and adding theory of mind lets choosers trust punishers more and lets punishers actually return more, matching the human findings. In the public goods game of Feinberg et al. (2014), this same architecture captures the gossip effect but cannot distinguish gossip with ostracism from gossip alone; adding a strategic-reflection component restores the distinction and the correct temporal trends. The validated models then predict that making punishment anonymous lowers punishment rates by 16 percentage points but leaves a substantial baseline, implying mixed reputational and intrinsic motives, and that a pre-round discussion period raises contributions beyond gossip with ostracism.

Load-bearing premise

The validation's significance tests treat each agent decision as an independent observation, even though agents interact in fixed groups across multiple rounds, so correlated decisions could make the replicated effects appear stronger than they are.

Editorial extensions

If this is right

  • Third-party punishment studies with LLM agents require both persona-based variation and theory-of-mind reasoning; removing either component abolishes the trust-signal effect, so any generative agent model of costly punishment needs both.
  • The gossip and ostracism effects in public goods games require a strategic-reflection component on top of the social architecture; the emotional-reflection component actively reverses the ostracism effect.
  • The validated models predict that anonymity reduces third-party punishment by about 16 percentage points but does not eliminate it, attributing the drop to reputational motives and the remainder to intrinsic motives.
  • The models predict that an open discussion period before each round of a public goods game with gossip and ostracism further raises group contributions, especially in the first round.

Reading between the lines

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

  • An unstated consequence of the 16-point anonymity gap is a quantitative decomposition: if punishment in the private condition is the intrinsic baseline, the reputational share of third-party punishment is roughly one quarter of the public-condition rate, a number a human experiment could directly check.
  • The discussion-period result suggests a family of testable interventions: varying whether discussion is structured, whether promises are recorded, or whether discussion is secret would show whether the mechanism is norm creation, commitment, or coordination.
  • The same minimal-component search could be run on more complex social environments to see whether additional components become necessary when agents track partners over time or when status hierarchies exist.
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 / 5 minor

Summary. The paper proposes a two-stage framework for validating generative agent-based models (GABMs): first, systematically vary cognitive components (persona, theory of mind, strategic reflection, emotion) to replicate known human results from two landmark social dilemma papers (Jordan et al. 2016 on third-party punishment; Feinberg et al. 2014 on gossip and ostracism); second, use the validated architecture to generate novel predictions (anonymous punishment and pre-round discussion). In Study 1, the authors report that personas plus theory of mind are necessary to reproduce the effect that punishers are trusted more, while personas alone produce the effect that punishers are actually more trustworthy. In Study 2, they find that adding a strategic reflection component is needed to reproduce most (5 of 6) target effects, including the gossip-plus-ostracism advantage and the declining trend in the basic condition. The novel predictions are that anonymous punishment substantially reduces but does not eliminate third-party punishment, and that pre-round discussion further increases cooperation in the public goods game. The paper argues that this constitutes a validation template for GABMs and demonstrates their capacity for generating testable hypotheses.

Significance. The paper addresses an important and timely problem: how to validate generative agent-based models against human experimental data. The two-stage design is clear, the code is released, and the comparison of multiple cognitive architectures is a useful methodological template. If the central claims hold, this work would provide one of the more systematic demonstrations that LLM-based agents can reproduce known social-psychological effects and generate novel, falsifiable predictions. The authors are explicit that their validation criterion is directional replication with statistical significance, which is a reasonable starting point. However, the load-bearing conclusions about which cognitive components are 'essential' currently rest on an unreported parameter in the persona prompt and on a potentially confounded implementation of the strategic component; these issues must be resolved before the claims can be fully credited.

major comments (3)
  1. [Prompts, Personality Reflection] The Personality Reflection prompt for Study 1 contains the literal placeholder '{persona.cooperation tendency}', and the paper nowhere reports the numeric values (or any instantiation) of this parameter, nor a sensitivity analysis. Because the Base agent (without personas) punishes 100% of the time while persona-based agents punish at 64.8% (public) and 48.9% (private), this parameter is a direct calibration channel for the very punishment rates that drive both the replication and the novel prediction. Consequently, the conclusion that persona-based individual differences are 'essential' is not identifiable from the reported experiment: different draws of this parameter could account for the observed architecture differences. The authors must report the exact values used for all personas (e.g., in a table or supplementary file), justify their range, and show that the architecture ranking (e.g., the contrast between the Social architecture and the Personas-only ablation) is robust to reasonable variations of this parameter.
  2. [Study 2, Prompts (Strategy Reflection)] The strategic reflection component in Study 2 is implemented with the instruction: 'Focus on selfishly maximizing your own earnings.' This is a strong experimental demand that explicitly biases agents toward free-riding in a public goods game, and it could by itself produce the declining contributions in the basic condition and the larger contrast between conditions that the paper attributes to a 'strategic component.' The Methods section describes the component more neutrally as 'What's the best strategy for maximizing your long-term earnings in this game?' The authors should either use the neutral wording in the simulations, or justify why the explicitly selfish instruction is not a confound. Without such a control, the claim that strategic reasoning (rather than the demand to be selfish) is the critical cognitive component is not supported.
  3. [Study 2, Results and Table 2] The statistical analyses in Study 2 treat agent observations as independent despite the clear nested structure: agents play in groups of four over six rounds, and the manuscript reports that each condition line consists of 5 experiments of 24 agents. The reported F statistics (e.g., F(1,119) for pairwise comparisons) do not account for clustering by group or run, so the p-values are inflated and the validation looks stronger than is warranted. The authors should reanalyze the data with mixed-effects models (random intercepts for group and simulation run, or including round as a repeated measure) or with cluster-robust standard errors, and then confirm that the architecture comparisons and the novel prediction remain significant under the corrected analysis.
minor comments (5)
  1. [Methods, Agent Architecture Configurations] The Prompts section for Study 2 introduces a 'Theory of Mind Analysis 2' prompt, but the architecture configurations described in Methods do not mention this as a separate component. Please clarify which architectures included this second ToM prompt and whether it was part of the Social architecture used in both studies.
  2. [Table 2 and Figure 3] The unit of analysis in Study 2 is ambiguous: the text says '5 experiments of 24 agents each' (120 agents total), but the pairwise comparisons are reported as F(1,119), which would be consistent with 120 subjects in a repeated-measures design. Please state explicitly whether the same agents played all conditions, how many independent simulation runs were performed, and how the degrees of freedom were obtained.
  3. [Methods, Base Agent Architecture] The sampling temperature is mentioned only in passing ('a relatively high LLM sampling temperature of 1.0') for the base agent. Please report the temperature (and any other sampling hyperparameters) used for every architecture and condition, since this affects the stochasticity and reproducibility of the results.
  4. [Study 1, Table 1] The Personas-only (No ToM) ablation still shows a significant effect for the return percentage (t = 3.82, p < 0.001), meaning that the 'actual trustworthiness' component of the TPP finding is replicated without theory of mind. The abstract and discussion claim that ToM is 'essential' for replicating TPP as a costly signal; this claim should be nuanced to specify that ToM is essential for the signaling component (trust toward punishers) but not for the consistency component (punishers' actual trustworthiness).
  5. [Figure 2] The caption 'TPP Results' is very brief; please expand it to identify panels (A), (B), and (C) with the specific measures and conditions, and include the statistical annotations that are currently only in the text.

Circularity Check

0 steps flagged · score 1.0 of 10

No circular derivation: novel predictions are out-of-sample outputs of architectures selected on replication data; self-citation is not load-bearing. Unreported persona parameter is a reproducibility concern, not circularity.

full rationale

The derivation chain is replication-then-prediction, not prediction-from-fit. Study 1 selects the Social architecture (Persona + ToM) by comparing ablations against Jordan et al. (2016)'s two key findings (amount sent: t=6.08 with ToM vs t=0.13 without; return percentage: t=9.83). Study 2 similarly adds Strategic Reflection only after the Social architecture fails the gossip+ostracism contrast (F(1,119)=1.55 without strategy vs F=53.29 with strategy). The novel conditions—private TPP and pre-round discussion—are new manipulations; no equation in the paper computes these outcomes from the replication statistics, and no parameter is fitted to those prediction targets. The self-citation to Cross et al. (2024) in the Discussion ('more complex architectures like Hypothetical Minds... our goal here was to identify the minimal set') is contrastive and not load-bearing; no uniqueness theorem is imported from the authors' prior work. The main caveat is identifiability, not circularity: the Prompts section contains the literal template 'with a cooperation tendency of {persona.cooperation tendency}', and the numeric values/distribution are never reported. If those values were tuned to produce human-level punishment rates, the private-condition baseline would be partly fitted; but the text gives no evidence of such fitting, and the public/private difference and the discussion-condition boost are not algebraically reducible to this parameter. Study 2's non-independence (repeated rounds within groups) is also a statistical-correctness risk, not a circularity. Overall, no step in the claimed derivation reduces by construction to its own input.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The central claims rest on the choice of prompt components (personas, theory of mind, strategy), on unspecified persona parameters, and on the assumption that LLM agents are a valid proxy for human decision-making. The novel predictions are model outputs, not fitted constants, so circularity burden is low. The main risk is that architecture selection on the replication data and uncontrolled simulation parameters inflate the apparent validity.

free parameters (2)
  • persona.cooperation tendency = not reported
    TPP personality prompt includes a numeric cooperation tendency; values or sampling distribution not given. If tuned to reproduce human punishment rates, the public/private difference may partly depend on it.
  • LLM sampling temperature = not reported for most runs
    Mentioned only that base model used temperature 1.0; final temperature settings for GPT-4o runs not specified, affecting variance and significance.
assumptions (4)
  • domain assumption LLM agents with these prompt components are a valid proxy for human social decision-making in these games.
    The entire validation strategy depends on this premise; the paper only shows directional alignment on a few effects.
  • domain assumption The human results from Jordan et al. (2016) and Feinberg et al. (2014) are accurate and reproducible.
    Used as ground truth for validation; any errors propagate.
  • standard math Standard statistical tests (t-tests, ANOVA) are appropriate for comparing agent and human effects.
    Assumes normality and independence; independence is questionable in grouped simulations.
  • domain assumption GPT-4o's stochastic behavior is stable across API calls.
    Simulation results depend on API randomness; no seed control is reported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Validating Generative Agent-Based Models of Social Norm Enforcement: From Replication to Novel Predictions." pith.science (2026). https://pith.science/paper/AHYQXZAX

@misc{pith2026250722049,
  author       = {Pith},
  title        = {Pith review of: Validating Generative Agent-Based Models of Social Norm Enforcement: From Replication to Novel Predictions},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AHYQXZAX}},
  note         = {Machine review of arXiv:2507.22049}
}
read the original abstract

As large language models (LLMs) advance, there is growing interest in using them to simulate human social behavior through generative agent-based modeling (GABM). However, validating these models remains a key challenge. We present a systematic two-stage validation approach using social dilemma paradigms from psychological literature, first identifying the cognitive components necessary for LLM agents to reproduce known human behaviors in mixed-motive settings from two landmark papers, then using the validated architecture to simulate novel conditions. Our model comparison of different cognitive architectures shows that both persona-based individual differences and theory of mind capabilities are essential for replicating third-party punishment (TPP) as a costly signal of trustworthiness. For the second study on public goods games, this architecture is able to replicate an increase in cooperation from the spread of reputational information through gossip. However, an additional strategic component is necessary to replicate the additional boost in cooperation rates in the condition that allows both ostracism and gossip. We then test novel predictions for each paper with our validated generative agents. We find that TPP rates significantly drop in settings where punishment is anonymous, yet a substantial amount of TPP persists, suggesting that both reputational and intrinsic moral motivations play a role in this behavior. For the second paper, we introduce a novel intervention and see that open discussion periods before rounds of the public goods game further increase contributions, allowing groups to develop social norms for cooperation. This work provides a framework for validating generative agent models while demonstrating their potential to generate novel and testable insights into human social behavior.

Figures

Figures reproduced from arXiv: 2507.22049 by the authors.

Figure 1
Figure 1. A. Two-stage approach of model validation and novel prediction generation. B. Iterative refinement [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 5
Figure 5. Without distinct personas, agents showed [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 2
Figure 2. TPP Results: (A) Average amount sent by choosers to signallers in Stage 2. (B) Percentage [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figures from the paper (3 more)
Figure 3
Figure 3. Figure 3: Gossip experiment results. Each line includes 5 experiments of 24 agents each. [PITH_FULL_IMAGE:figures/full_fig_p009_3.png]
Figure 4
Figure 4. Figure 4: Additional condition in Study 2, examin [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]
Figure 5
Figure 5. Figure 5: Personas needed to not punish. Percentage of agents who chose to pay a cost to punish selfish behavior in Stage 1. Error bars represent 95% confidence intervals [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

28 extracted references · 21 canonical work pages

  1. [1]

    write newline

    " write newline " cite write " FUNCTION editor.postfix editor num.names #1 > "( )" "( )" if FUNCTION editor.trans.postfix editor num.names #1 > "( )" "( )" if FUNCTION trans.postfix translator num.names #1 > "( )" "( )" if FUNCTION authors.editors.reflist.apa5 'field := 'dot := field num.names 'numnames := numnames 'format.num.names := format.num.names na...

  2. [2]

    , Liu, R

    anthis2025llm APACrefauthors Anthis, J R. , Liu, R. , Richardson, S M. , Kozlowski, A C. , Koch, B. , Evans, J. Bernstein, M. APACrefauthors \ 2025 . Llm social simulations are a promising research method Llm social simulations are a promising research method . arXiv preprint arXiv:2504.02234

    arXiv 2025

  3. [3]

    tragedy of the commons

    barclay2004trustworthiness APACrefauthors Barclay, P. APACrefauthors \ 2004 . Trustworthiness and competitive altruism can also solve the “tragedy of the commons” Trustworthiness and competitive altruism can also solve the “tragedy of the commons” . Evolution and Human Behavior 25 4 209--220

    work page 2004

  4. [4]

    , Dickhaut, J

    berg1995trust APACrefauthors Berg, J. , Dickhaut, J. \ McCabe, K. APACrefauthors \ 1995 . Trust, reciprocity, and social history Trust, reciprocity, and social history . Games and economic behavior 10 1 122--142

    work page 1995

  5. [5]

    , Wang, X

    chen2024persona APACrefauthors Chen, J. , Wang, X. , Xu, R. , Yuan, S. , Zhang, Y. , Shi, W. others APACrefauthors \ 2024 . From persona to personalization: A survey on role-playing language agents From persona to personalization: A survey on role-playing language agents . arXiv preprint arXiv:2404.18231

    arXiv 2024

  6. [6]

    , Canevello, A

    crocker2017social APACrefauthors Crocker, J. , Canevello, A. \ Brown, A A. APACrefauthors \ 2017 . Social motivation: Costs and benefits of selfishness and otherishness Social motivation: Costs and benefits of selfishness and otherishness . Annual review of psychology 68 1 299--325

    work page 2017

  7. [7]

    , Xiang, V

    cross2024hypothetical APACrefauthors Cross, L. , Xiang, V. , Bhatia, A. , Yamins, D L. \ Haber, N. APACrefauthors \ 2024 . Hypothetical minds: Scaffolding theory of mind for multi-agent tasks with large language models Hypothetical minds: Scaffolding theory of mind for multi-agent tasks with large language models . arXiv preprint arXiv:2407.07086

    arXiv 2024

  8. [8]

    \ G \"a chter, S

    fehr2000cooperation APACrefauthors Fehr, E. \ G \"a chter, S. APACrefauthors \ 2000 . Cooperation and punishment in public goods experiments Cooperation and punishment in public goods experiments . American Economic Review 90 4 980--994

    work page 2000

Show all 28 references
  1. [9]

    , Willer, R

    feinberg2014gossip APACrefauthors Feinberg, M. , Willer, R. \ Schultz, M. APACrefauthors \ 2014 . Gossip and ostracism promote cooperation in groups Gossip and ostracism promote cooperation in groups . Psychological science 25 3 656--664

    2014

  2. [10]

    , G \"a chter, S

    fischbacher2001people APACrefauthors Fischbacher, U. , G \"a chter, S. \ Fehr, E. APACrefauthors \ 2001 . Are people conditionally cooperative? Evidence from a public goods experiment Are people conditionally cooperative? evidence from a public goods experiment . Economics let...

    2001

  3. [11]

    , Tybur, J M

    griskevicius2007blatant APACrefauthors Griskevicius, V. , Tybur, J M. , Sundie, J M. , Cialdini, R B. , Miller, G F. \ Kenrick, D T. APACrefauthors \ 2007 . Blatant benevolence and conspicuous consumption: when romantic motives elicit strategic costly signals. Blatant benevole...

    2007

  4. [12]

    , McElreath, R

    henrich2006costly APACrefauthors Henrich, J. , McElreath, R. , Barr, A. , Ensminger, J. , Barrett, C. , Bolyanatz, A. others APACrefauthors \ 2006 . Costly punishment across human societies Costly punishment across human societies . Science 312 5781 1767--1770

    2006

  5. [13]

    , Hoffman, M

    jordan2016third APACrefauthors Jordan, J J. , Hoffman, M. , Bloom, P. \ Rand, D G. APACrefauthors \ 2016 . Third-party punishment as a costly signal of trustworthiness Third-party punishment as a costly signal of trustworthiness . Nature 530 7591 473--476

    2016

  6. [14]

    \ Rand, D G

    jordan2020signaling APACrefauthors Jordan, J J. \ Rand, D G. APACrefauthors \ 2020 . Signaling when no one is watching: A reputation heuristics account of outrage and punishment in one-shot anonymous interactions. Signaling when no one is watching: A reputation heuristics acco...

    2020

  7. [15]

    , Hadfield-Menell, D

    koster2022spurious APACrefauthors K \"o ster, R. , Hadfield-Menell, D. , Everett, R. , Weidinger, L. , Hadfield, G K. \ Leibo, J Z. APACrefauthors \ 2022 . Spurious normativity enhances learning of compliance and enforcement behavior in artificial agents Spurious normativity e...

    2022

  8. [16]

    ledyard1994public APACrefauthors Ledyard, J O. \ . APACrefauthors \ 1994 . Public goods: A survey of experimental research Public goods: A survey of experimental research . Division of the Humanities and Social Sciences, California Inst. of Technology

    1994

  9. [17]

    , Vezhnevets, A S

    leibo2024theory APACrefauthors Leibo, J Z. , Vezhnevets, A S. , Diaz, M. , Agapiou, J P. , Cunningham, W A. , Sunehag, P. others APACrefauthors \ 2024 . A theory of appropriateness with applications to generative artificial intelligence A theory of appropriateness with applica...

    2024 arXiv

  10. [18]

    mind" exploration of large scale language model society Camel: Communicative agents for

    li2023camel APACrefauthors Li, G. , Hammoud, H A A K. , Itani, H. , Khizbullin, D. \ Ghanem, B. APACrefauthors \ 2023 . Camel: Communicative agents for" mind" exploration of large scale language model society Camel: Communicative agents for" mind" exploration of large scale la...

    2023

  11. [19]

    \ Olsen, J P

    march2008logic APACrefauthors March, J G. \ Olsen, J P. APACrefauthors \ 2008 . The logic of appropriateness The logic of appropriateness

    2008

  12. [20]

    , Semmann, D

    milinski2002reputation APACrefauthors Milinski, M. , Semmann, D. \ Krambeck, H J. APACrefauthors \ 2002 . Reputation helps solve the ‘tragedy of the commons’ Reputation helps solve the ‘tragedy of the commons’ . Nature 415 6870 424--426

    2002

  13. [21]

    , Sommerfeld, R D

    milinski2008collective APACrefauthors Milinski, M. , Sommerfeld, R D. , Krambeck, H J. , Reed, F A. \ Marotzke, J. APACrefauthors \ 2008 . The collective-risk social dilemma and the prevention of simulated dangerous climate change The collective-risk social dilemma and the pre...

    2008

  14. [22]

    APACrefauthors \ 2014

    neisser2014cognitive APACrefauthors Neisser, U. APACrefauthors \ 2014 . Cognitive psychology: Classic edition Cognitive psychology: Classic edition . Psychology press

    2014

  15. [23]

    APACrefauthors \ 1990

    ostrom1990governing APACrefauthors Ostrom, E. APACrefauthors \ 1990 . Governing the commons: The evolution of institutions for collective action Governing the commons: The evolution of institutions for collective action . Cambridge University

    1990

  16. [24]

    , O'Brien, J

    park2023generative APACrefauthors Park, J S. , O'Brien, J. , Cai, C J. , Morris, M R. , Liang, P. \ Bernstein, M S. APACrefauthors \ 2023 . Generative agents: Interactive simulacra of human behavior Generative agents: Interactive simulacra of human behavior . Proceedings of th...

    2023

  17. [25]

    , Zou, C Q

    park2024generative APACrefauthors Park, J S. , Zou, C Q. , Shaw, A. , Hill, B M. , Cai, C. , Morris, M R. Bernstein, M S. APACrefauthors \ 2024 . Generative agent simulations of 1,000 people Generative agent simulations of 1,000 people . arXiv preprint arXiv:2411.10109

    2024 arXiv

  18. [26]

    , Joireman, J

    van2013psychology APACrefauthors Van Lange, P A. , Joireman, J. , Parks, C D. \ Van Dijk, E. APACrefauthors \ 2013 . The psychology of social dilemmas: A review The psychology of social dilemmas: A review . Organizational Behavior and Human Decision Processes 120 2 125--141

    2013

  19. [27]

    , Agapiou, J P

    vezhnevets2023generative APACrefauthors Vezhnevets, A S. , Agapiou, J P. , Aharon, A. , Ziv, R. , Matyas, J. , Du \'e \ n ez-Guzm \'a n, E A. Leibo, J Z. APACrefauthors \ 2023 . Generative agent-based modeling with actions grounded in physical, social, or digital space using C...

    2023 arXiv

  20. [28]

    , Yang, K

    zhang2024proagent APACrefauthors Zhang, C. , Yang, K. , Hu, S. , Wang, Z. , Li, G. , Sun, Y. others APACrefauthors \ 2024 . ProAgent: building proactive cooperative agents with large language models Proagent: building proactive cooperative agents with large language models . P...

    2024

Pith tools

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