Pith. sign in

REVIEW 5 major objections 6 minor 32 references

Static network structure cannot stabilize cooperation among Large Language Model agents

T0 review · 5 major / 6 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read The paper claims that current large language model agents do not cooperate more in static social networks than in well-mixed populations, so static network structure cannot stabilize cooperation among them as it does in humans.

desk verdict Useful cautionary data on LLM agents in networked prisoner's dilemma, but the title overclaims: their own figures show LLMs do respond to social context, just not like humans. read the letter →

arxiv 2411.10294 v1 pith:46CYRHLW submitted 2024-11-15 cs.SI cs.CYcs.GTphysics.soc-ph

Jin Han , Balaraju Battu , Ivan Romić , Talal Rahwan , Petter Holme This is my paper · ORCID
classification cs.SIcs.CYcs.GTphysics.soc-ph
keywords largelanguagemodelsprisoner'sdilemmanetworkreciprocitycooperationsocialdilemmasLLMagentsstaticnetworkshuman-AIcomparison
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 asks whether large language models, cast as agents in a repeated prisoner's dilemma—a game in which cooperating costs the player but benefits a partner—reproduce the human tendency to cooperate more in fixed social networks than in fluid, well-mixed populations. It reports that they do not: GPT-3.5 hovers near 50 percent cooperation regardless of network degree or benefit-to-cost ratio, while GPT-4 cooperates more in well-mixed settings and less on static networks, the opposite of the human pattern. The authors conclude that current LLMs lack the reciprocity norms and network inference that let humans stabilize cooperation when the benefit-to-cost ratio exceeds degree, and that claims of LLM-human equivalence in behavioral experiments do not generalize to networked social dilemmas.

What carries the argument

The load-bearing setup is the repeated prisoner's dilemma on circulant ring networks, where each of the $k$ nearest neighbors receives benefit $b$ if the agent cooperates, and cooperation costs $c$; the game is played for about 15 rounds with the same neighbors while the ratio $b/c$ is varied from 2 to 6. The human benchmark is the threshold $b/c > k$, under which cooperation becomes stable in static networks through assortment. The LLM agents are prompted with the human instructions from that experiment and asked to respond with C or D, and their cooperation trajectories are compared across well-mixed and networked conditions, with controlled perturbations of neighbor behavior serving as an additional probe of reciprocity.

What would settle it

If, when the prompt explicitly supplies the complete list or diagram of neighbors and their past actions, LLM cooperation rises above 50 percent and tracks the $b/c > k$ threshold across values of $k$, the claim that static network structure cannot stabilize LLM cooperation would be disproven.

Watch

Extended reading notes

Core claim

The central claim is that static network structure cannot stabilize cooperation among LLM agents as of 2024. Replicating the design of a classical human experiment—a repeated prisoner's dilemma on ring networks with degree $k=2,4,6$ and benefit-to-cost ratios $b/c=2,4,6$—the authors find that humans cooperate more in the network when $b/c > k$, while GPT-3.5's cooperation stays nearly flat and GPT-4's cooperation is higher in the well-mixed condition than on the network. They interpret this as evidence that LLMs do not strategically adjust to network structure or evolving social context, do not infer the neighborhood they are embedded in, and instead follow rigid, model-specific behavioral patterns. The paper's conclusion is that claims that LLMs can replace human participants in behavioral experiments are overgeneralized.

Load-bearing premise

The experiments assume that telling an LLM it will face the same neighbors each round and showing it their choices gives the model enough information to perceive the network, so its failure to cooperate more on the network is a behavioral limitation rather than a missing-information artifact.

Editorial extensions

If this is right

  • LLMs cannot currently be used as drop-in human substitutes for studying how network structure shapes cooperation.
  • Cooperation in LLM social dilemmas is driven more by the model's intrinsic behavior than by the structural parameters $b/c$ and $k$.
  • The human pattern of network reciprocity—cooperation rising when $b/c > k$—is absent, so future LLM design must explicitly encode social norms or network awareness to close the gap.
  • The GPT-4 result, higher cooperation in well-mixed than in networked settings, runs opposite to human behavior, so simply making LLMs more cooperative is not the fix.

Reading between the lines

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

  • The authors do not vary the information given to the model; a natural extension would present the full neighbor list or a network diagram and test whether cooperation then tracks $b/c > k$, which would separate missing information from missing capability.
  • Because the result is a snapshot of specific 2024 models, it should be treated as a benchmark rather than a law; future LLMs trained on more social-interaction data may behave differently.
  • The same controlled-stimulus protocol could be used to measure whether prompting LLMs to reason explicitly about reciprocity, or giving them a norm-rich backstory, restores network-stabilized cooperation.
  • If the pattern holds across architectures, the experimental template becomes a useful test for social intelligence in AI agents, complementing game-theoretic measures.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 6 minor

Summary. This paper reports experiments in which LLM agents (GPT-3.5, GPT-4, Claude, Mixtral) play repeated Prisoner's Dilemma games in both well-mixed and static ring-network conditions, closely following the design of Rand et al. (2014). The authors find that, unlike human participants, LLMs do not show higher cooperation in static networks relative to well-mixed populations: GPT-3.5 is largely insensitive to network parameters, GPT-4 cooperates more in well-mixed settings, and a controlled stimulus experiment shows that some models respond to the composition of neighbors. The paper concludes that static network structure cannot stabilize cooperation among LLM agents and that LLMs do not adapt to network structures or evolving social contexts.

Significance. If the findings are robust, they provide a cautionary data point for the growing use of LLMs as human surrogates in behavioral experiments. The paper includes a replication of a well-known human experiment, reports error bands and some inferential statistics, and makes its prompt content available, which are strengths. However, the significance is currently limited by overgeneralized claims, an internal inconsistency between the stated conclusion and the presented data, and several methodological ambiguities. The most defensible contribution is a negative result for the specific models and conditions tested; the universal 'cannot' claim is not supported.

major comments (5)
  1. [Abstract and Discussion] The abstract and Discussion state that 'LLMs do not alter their cooperative behavior in response to network structures or evolving social contexts,' but this is contradicted by the paper's own data. In Fig. 7, Claude and GPT-4 drop from near-total cooperation to near-zero when cooperative neighbors are replaced by defectors, and the Results explicitly say these models 'are sensitive to shifts in their social environment.' Moreover, GPT-4 shows higher cooperation in the well-mixed condition than in the fixed-ring condition (Fig. 3C vs. Fig. 3B; Fig. 4D vs. Fig. 4A), so it does distinguish the two structures, albeit in the opposite direction from humans. This internal inconsistency undermines the central claim as stated. The conclusion should be narrowed to 'LLMs do not show the human pattern of increased cooperation in static networks,' and the blanket 'do not alter' phrasing should be removed.
  2. [Title and Experimental Setup] The title and abstract assert that 'static network structure cannot stabilize cooperation among Large Language Model agents,' a universal negative, but the evidence comes from only four model families (GPT-3.5, GPT-4, Claude, Mixtral), with no model versions or sampling temperatures reported, and from only one network topology (circulant rings with k = 2, 4, 6). This is insufficient to support a claim about all LLMs and all static network structures. The claim should be restricted to the tested models and settings, and the title should be revised accordingly.
  3. [Methods, 'Prompting the LLMs'] The prompt tells the participant 'You will be connected to the same people every round' and shows neighbor choices, but it never provides the network structure, such as the ring topology, the node's degree, or the fact that neighbors are connected to each other. The Discussion attributes the absence of network-adaptive behavior to a lack of capacity to infer network structure, but an alternative explanation is that the models simply lack the information needed to infer the structure. The paper needs a control condition in which the graph structure is explicitly described to distinguish between incapacity and missing information; without this, the conclusion that LLMs 'lack the capacity to infer network structure' does not follow.
  4. [Methods, 'Experimental setup'] The well-mixed condition is severely underspecified. The paper contrasts 'well-mixed' with 'structured networks' but does not state how the LLM is paired with other players in each round (e.g., random rematching, a complete graph, or something else). In the human benchmark (Rand et al. 2014), well-mixed means random, one-shot interactions without repeated partners. For LLM agents, the implementation matters because the dialogue format may carry over information between rounds, making the comparison with the human condition ambiguous. The manuscript must specify the matching algorithm and verify its comparability to the human protocol.
  5. [General reproducibility] No model version (e.g., gpt-3.5-turbo vs. text-davinci-003), temperature, or other sampling parameters are reported for any of the four models. LLM behavior is known to vary substantially with these choices, and the paper's conclusion depends on a small number of model calls. This omission prevents independent replication and weakens the generality of the findings. At a minimum, the authors should report these details; ideally, they should probe sensitivity to them.
minor comments (6)
  1. [Results] In the paragraph beginning 'In contrast, GPT-4 shows a more dynamic approach,' the text says 'when the network structure is sparse, with fewer connections (such as k = 2k)'; this appears to be a typo for 'k = 2'.
  2. [Throughout] The name of the Mixtral model is written as 'Mistral' in the Discussion section; the authors should use one consistent spelling.
  3. [References] There are duplicate references: Ref. 8 and Ref. 15 are both Lorè, N. & Heydari, B. (2023), and Ref. 13 and Ref. 31 are both Rand et al. (2014). The reference list should be consolidated.
  4. [Figure 7] The caption says the neighborhood changes from four to three cooperators in panel A at 'timestep 5,' but the text describes 'a progressive replacement of cooperators with defectors in three stages' over the course of 25 rounds; the exact timing and mapping of panels to stages should be clarified for readability.
  5. [Figure captions] The figure captions contain apparent text-extraction artifacts of the form '/gid00002/gid00001/...' that make parts of the captions unreadable; these need to be cleaned up in the final version.
  6. [Experimental setup] The paper says experiments are repeated five times for averages, but Fig. 7 states values are averaged over 10 runs; this discrepancy should be clarified.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper is an empirical comparison against an external human benchmark with no fitted parameters and no self-citation chain.

full rationale

This is an empirical behavioral study, not a derivation, so there is no claimed derivation chain whose conclusions reduce to its inputs. The authors explicitly adapt the methods and settings of Rand et al. (Ref. 13) as an external human benchmark, and all LLM cooperation rates are measured directly from repeated prompted responses. No parameter is fitted to the LLM data, no quantity is defined in terms of the outcome it is used to predict, and no load-bearing uniqueness theorem or ansatz is imported from the authors' own prior work. The only apparent weakness is that the prompt tells the LLM it will be connected to the same people every round but never displays the graph structure; this is a construct-validity concern about whether the null result tests inference capacity, not a circular reduction. The claim that LLMs 'do not alter their cooperative behavior' is also internally inconsistent with Fig. 7, where agents sharply reduce cooperation when neighbors defect, but that inconsistency is a correctness problem rather than a circularity problem. Because the central empirical result is self-contained against an external benchmark and is not statistically forced by construction, the circularity score is 0.

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

No free parameters are fitted to data. The experimental variables k and b/c are taken from the human study of Rand et al. (2014). The paper introduces no new theoretical entities. It relies on the assumptions that the human protocol translates to prompts, that the human benchmark is comparable, and that parsed C/D choices reflect the models' strategies.

assumptions (4)
  • domain assumption The payoff structure and network protocol from Rand et al. (2014) are a valid translation for LLM agents, including cost c=10k and per-neighbor benefit b.
    The study assumes the human experiment design can be carried over to prompts without distorting the social dilemma. Invoked in Methods 'Prompting the LLMs' and 'Experimental setup'.
  • domain assumption The human cooperation benchmark from Ref. 13 is reliable and comparable despite differences in participant counts and procedure.
    Figures 3, 5, and 6 compare LLM cooperation levels directly to human data from Ref. 13, which had different sample sizes and no LLM-specific prompts.
  • domain assumption LLM utterances of 'C' or 'D' map unambiguously to cooperative or defective behavior, and dialogue failures can be repaired without biasing results.
    The method section describes semi-automated repair of non-compliant responses and interruption after three failures, which presumes that repair does not introduce systematic bias.
  • domain assumption Circulant ring networks represent static networks in general.
    Following Ref. 13, only cycle graphs are tested; the conclusion 'static network structure cannot stabilize' generalizes from this one topology.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Static network structure cannot stabilize cooperation among Large Language Model agents." pith.science (2026). https://pith.science/paper/46CYRHLW

@misc{pith2026241110294,
  author       = {Pith},
  title        = {Pith review of: Static network structure cannot stabilize cooperation among Large Language Model agents},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/46CYRHLW}},
  note         = {Machine review of arXiv:2411.10294}
}
read the original abstract

Large language models (LLMs) are increasingly used to model human social behavior, with recent research exploring their ability to simulate social dynamics. Here, we test whether LLMs mirror human behavior in social dilemmas, where individual and collective interests conflict. Humans generally cooperate more than expected in laboratory settings, showing less cooperation in well-mixed populations but more in fixed networks. In contrast, LLMs tend to exhibit greater cooperation in well-mixed settings. This raises a key question: Are LLMs about to emulate human behavior in cooperative dilemmas on networks? In this study, we examine networked interactions where agents repeatedly engage in the Prisoner's Dilemma within both well-mixed and structured network configurations, aiming to identify parallels in cooperative behavior between LLMs and humans. Our findings indicate critical distinctions: while humans tend to cooperate more within structured networks, LLMs display increased cooperation mainly in well-mixed environments, with limited adjustment to networked contexts. Notably, LLM cooperation also varies across model types, illustrating the complexities of replicating human-like social adaptability in artificial agents. These results highlight a crucial gap: LLMs struggle to emulate the nuanced, adaptive social strategies humans deploy in fixed networks. Unlike human participants, LLMs do not alter their cooperative behavior in response to network structures or evolving social contexts, missing the reciprocity norms that humans adaptively employ. This limitation points to a fundamental need in future LLM design -- to integrate a deeper comprehension of social norms, enabling more authentic modeling of human-like cooperation and adaptability in networked environments.

Figures

Figures reproduced from arXiv: 2411.10294 by the authors.

Figure 1
Figure 1. Illustrating the networked interactions in an eight-node network, with the degrees k = 2,4,6 in panels A, B, and C, respectively. The convention of using so-called circulant graphs to represent static networks follows Ref. 31. participants. Through these experiments, we aim to determine whether LLMs approach social dilemmas similarly to humans in analogous settings. In our setting, we have a community or social netw… view at source ↗
Figure 2
Figure 2. Two example experiments with GPT-3.5. The parameter settings are k = 2 and b/c = 6. The position around the circle corresponds to the identity of the node. In panel A, the interactions are well-mixed. In panel B, neighbors along the perimeter interact with each other. More rounds, fewer participants We explore how the cooperative dynamics change when LLMs like GPT-3.5 interact in different social structures, specifi… view at source ↗
Figure 3
Figure 3. Comparing networked and well-mixed topologies for humans (panel A, adapted from Ref.13), GPT-3.5 (panel B), and GPT-4 (panel C). All AI values are averaged over five independent runs of the experiments. The shaded regions represent the one-standard error confidence bands.              [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Average cooperation levels for GPT-3.5 (panels A, B, and C) and GPT-4 (panels D, E, and F). A and D give our results for the sparsest networks (k = 2), B and E show results for k = 4, and C and F give results for k = 6. The shaded regions represent one-standard-error c…
Figure 5
Figure 5. Figure 5: Cooperation level after 15 rounds as reported in Ref. 13 for humans A, compared to AI: GPT-3.5 and GPT-4, respectively. The colors represent the three theoretically distinct classes b/c < k, b/c = k, b/c > k, respectively. The first number on the bars indicates b/c, wh…
Figure 6
Figure 6. Figure 6: Comparison of relative payoff differences between human participants and GPT models. Panels A and B show the relative payoffs of defectors and cooperators for humans (adapted from Ref. 13), while panels C and D represent GPT-3.5, and panels E and F represent GPT-4. Pan…
Figure 7
Figure 7. Figure 7: The response of AI agents to changes in the neighbor’s behavior. In panel A, at timestep 5, the neighborhood of the AI player is changed from four to three cooperators. In panels B and C, the number of cooperators after the change is two and one, respectively. The chan…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

32 extracted references · 25 canonical work pages

  1. [1]

    Grossmann, I. et al. AI and the transformation of social science research. Science 380, 1108–1109 (2023)

    work page 2023

  2. [2]

    & Evans, J

    Sourati, J. & Evans, J. Accelerating science with human-aware artificial intelligence. Nat. Hum. Behav. 7, 1682–1696 (2023)

    work page 2023

  3. [3]

    ChatGPT broke the Turing test – the race is on for new ways to assess AI

    Biever, C. ChatGPT broke the Turing test – the race is on for new ways to assess AI. Nature 619, 686–689 (2023)

    work page 2023

  4. [4]

    Argyle, L. P. et al. Out of one, many: Using language models to simulate human samples. Polit. Anal. 31, 337–351 (2023)

    work page 2023

  5. [5]

    & Jackson, M

    Mei, Q., Xie, Y ., Yuan, W. & Jackson, M. O. A turing test of whether AI chatbots are behaviorally similar to humans.Proc. Natl. Acad. Sci. USA 121, e2313925121 (2024)

    work page 2024

  6. [6]

    & Matz, S

    Peters, H. & Matz, S. Large language models can infer psychological dispositions of social media users. PNAS Nexus pgae231 (2024)

    work page 2024

  7. [7]

    Wang, Z. et al. Large language models overcome the machine penalty when acting fairly but not when acting selfishly or altruistically (2024). Preprint arXiv:2410.03724

    arXiv 2024

  8. [9]

    J., Park, P

    Atari, M., Xue, M. J., Park, P. S., Blasi, D. & Henrich, J. Which humans? (2023). Working paper, Department of Human Evolutionary Biology, Harvard University. 9/12

    work page 2023

Show all 32 references
  1. [10]

    Shi, L. et al. Freedom of choice adds value to public goods. Proc. Natl. Acad. Sci. USA 117, 17516–17521 (2020)

    2020

  2. [11]

    & Baronchelli, A

    Centola, D. & Baronchelli, A. The spontaneous emergence of conventions: An experimental study of cultural evolution. Proc. Natl. Acad. Sci. USA 112, 1989–1994 (2015)

    2015

  3. [12]

    Airoldi, E. M. & Christakis, N. A. Induction of social contagion for diverse outcomes in structured experiments in isolated villages. Science 384, eadi5147 (2024)

    2024

  4. [14]

    Akata, E. et al. Playing repeated games with large language models (2023). Preprint arXiv:2305.16867

    2023 arXiv

  5. [15]

    & Heydari, B

    Lorè, N. & Heydari, B. Strategic behavior of large language models: Game structure vs. contextual framing (2023). Preprint arXiv:2309.05898

    2023 arXiv

  6. [16]

    & DeBacker, J

    Brookins, P. & DeBacker, J. M. Playing games with GPT: What can we learn about a large language model from canonical strategic games? (2023). Preprint SSRN:4493398

    2023

  7. [17]

    & Aiello, L

    Fontana, N., Pierri, F. & Aiello, L. M. Nicer than humans: How do large language models behave in the prisoner’s dilemma? arXiv preprint arXiv:2406.13605 (2024)

    2024 arXiv

  8. [18]

    & Fáth, G

    Szabó, G. & Fáth, G. Evolutionary games on graphs. Phys. Rep. 446, 97–216 (2007)

    2007

  9. [19]

    C., Pacheco, J

    Santos, F. C., Pacheco, J. M. & Lenaerts, T. Cooperation prevails when individuals adjust their social ties. PLoS Comp. Biol. 2, e140 (2006)

    2006

  10. [20]

    Connections: An Introduction to the Economics of Networks (Princeton University Press, Princeton NJ, 2007)

    Goyal, S. Connections: An Introduction to the Economics of Networks (Princeton University Press, Princeton NJ, 2007)

    2007

  11. [21]

    Camerer, C. F. Progress in behavioral game theory. J. Econ. Perspect. 11, 167–188 (1997)

    1997

  12. [22]

    & Krakauer, D

    Mitchell, M. & Krakauer, D. C. The debate over understanding in AI’s large language models. Proc. Natl. Acad. Sci. USA 120, e2215907120 (2023)

    2023

  13. [23]

    & Chiriatti, M

    Floridi, L. & Chiriatti, M. Gpt-3: Its nature, scope, limits, and consequences. Minds Mach. 30, 681–694 (2020)

    2020

  14. [24]

    Argyle, L. P. et al. Leveraging ai for democratic discourse: Chat interventions can improve online political conversations at scale. Proc. Natl. Acad. Sci. 120, e2311627120 (2023)

    2023

  15. [25]

    AI emerges as the frontier in behavioral science

    Meng, J. AI emerges as the frontier in behavioral science. Proc. Natl. Acad. Sci. USA 121, e2401336121 (2024)

    2024

  16. [26]

    R., Bai, X

    McKee, K. R., Bai, X. & Fiske, S. T. Humans perceive warmth and competence in artificial intelligence. iScience 26 (2023)

    2023

  17. [27]

    Large language models (llms) and empathy-a systematic review (2023)

    Sorin, V .et al. Large language models (llms) and empathy-a systematic review (2023). Preprint medRxiv:23293769

    2023

  18. [28]

    Grossmann, I. et al. Ai and the transformation of social science research. Science 380, 1108–1109 (2023)

    2023

  19. [29]

    Bail, C. A. Can generative AI improve social science? Proc. Natl. Acad. Sci. USA 121, e2314021121 (2024)

    2024

  20. [30]

    & Sun, Y

    Wang, H., Li, J., Wu, H., Hovy, E. & Sun, Y . Pre-trained language models and their applications.Engineering 25, 51–65 (2023)

    2023

  21. [31]

    G., Nowak, M

    Rand, D. G., Nowak, M. A., Fowler, J. H. & Christakis, N. A. Static network structure can stabilize human cooperation. Proc. Natl. Acad. Sci. USA 111, 17093–17098 (2014)

    2014

  22. [32]

    Nowak, M. A. Five rules for the evolution of cooperation. Science 314, 1560–1563 (2006)

    2006

  23. [33]

    Rand, D. G. & Nowak, M. A. Human cooperation. Trends Cogn. Sci. 17, 413–425 (2013)

    2013

  24. [34]

    & Fazelpour, S

    Gao, Y ., Lee, D., Burtch, G. & Fazelpour, S. Take caution in using LLMs as human surrogates: Scylla ex machina (2024). Preprint arXiv:2410.19599. Acknowledgements P.H. was supported by JSPS KAKENHI Grant No. JP 21H04595. Author contributions statement P.H. and I.R. conceived ...

    2024 arXiv

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.