Do LLMs Beat Nash? Testing Decentralized Coordination in Self-Play Multi-Agent Games

Deborah Sinishaw; Edwin Meriaux; Gregory Dudek; Qile Zhu

REVIEW 3 major objections 5 minor 24 references

Do LLMs Beat Nash? Testing Decentralized Coordination in Self-Play Multi-Agent Games

T0 review · 3 major / 5 minor · reviewed 2026-08-16 · deepseek-v4-flash

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

Pith's one-line read Two frontier LLMs consistently beat the Nash baseline in silent one-shot games.

desk verdict A genuinely new no-communication self-play benchmark with a plausible but unverified Nash baseline for three-action games; the headline 'beat Nash' claim needs the baseline fixed before it fully lands. read the letter →

arxiv 2608.12547 v1 pith:MEVN6O6A submitted 2026-08-12 cs.MA cs.RO

Deborah Sinishaw , Qile Zhu , Edwin Meriaux , Gregory Dudek This is my paper · ORCID

classification cs.MAcs.RO

keywords LLMagentsself-playno-communicationcoordinationNashequilibriummatrixgamesteamsuperrationalitymulti-agent

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 independent copies of the same large language model can coordinate in one-shot games with no communication, relying only on knowing that the counterpart is the same model. Its central claim is that two frontier-hosted models consistently exceed the Nash equilibrium baseline and approach the optimal joint outcome in several two-player archetypes, while most open-weight models achieve only partial gains that vary sharply with game structure. The paper also claims that this coordination ability largely fails in team games with four or more interchangeable agents, especially as the action space grows, because matching behavior helps in common-interest games but provides no mechanism for breaking symmetry when agents must choose different roles. The result matters for deploying LLM agents without a central controller or communication channel: some coordination is possible, but it is archetype-dependent and does not scale cleanly to larger teams.

What carries the argument

The load-bearing machinery is a parameterized benchmark of one-shot, no-communication self-play games with verified Nash benchmarks. Two-player games use seven archetypes plus controls, with cyclic rotation of action labels to remove position bias and a two-sided normalized score $s=(a-N)/(O-N)$ for $a\ge N$ and $s=(a-N)/(N-W)$ for $a<N$, mapping Nash to 0, the optimal joint payoff to $+1$, and the worst outcome to $-1$. Team games compress payoffs to the composition of each team, meaning the count of agents choosing each action, which reduces a $k^n$ profile space to $\binom{n+k-1}{k-1}$ compositions and keeps higher-tier games tractable; pure Nash equilibria are found by exhaustive search and mixed equilibria by the Lemke--Howson algorithm. The mechanism supposed to carry coordination is each agent's knowledge that its counterpart is the same model, acting as a shared-reasoning device in place of a communication channel.

What would settle it

For each three-action team game, recompute the mixed Nash equilibria directly on the original unperturbed payoff matrix, without the perturbation trick, and compare them with the stored benchmark values; if any reported solution fails the no-profitable-deviation condition, the baseline shifts and the below-Nash collapse needs to be re-scored. Alternatively, run the two-player anti-coordination games with a visible player index but still no communication; a large score increase would show that symmetry-breaking, not reasoning, is the missing capability.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that shared identity is a usable but narrow coordination device. In two-player matrix games across seven archetypes and two to ten actions, two frontier-hosted models consistently score above their Nash benchmark, reaching the optimal joint outcome in several archetypes, while open-weight models cluster at partial gains or below Nash depending on the archetype. The authors interpret this as evidence that LLMs can exploit knowledge that the other player is the same model when success means matching actions, as in coordination, stag hunt, and public goods, but not when it requires differentiation, as in anti-coordination and, for some models, battle of the sexes. In team games the capability evidently does not transfer: with four or more interchangeable agents, most models score below Nash in the three-action tier, and the strongest two-action team performers retain a positive signal only in the threshold archetype.

Load-bearing premise

The load-bearing premise is that the mixed Nash equilibria reported for the three-action games, computed after adding a small random perturbation to payoffs during solving and then discarding it, are genuine equilibria of the original unperturbed payoff matrices; if they are not, the above-Nash scores are measured against a baseline that may not be the true Nash value.

Editorial extensions

If this is right

If the central claim holds, independent copies of the same LLM can exceed the Nash baseline without communication in dyadic common-interest games, so explicit negotiation is not always necessary for LLM-agent coordination.
Archetype structure matters more than model scale: anti-coordination games remain a failure mode even for models that do well in matching games, and performance does not track parameter count.
Team coordination is the harder regime: most open-weight models fall below Nash in three-action team games, and even the strongest team performers keep a positive signal only in threshold games.
The stability of two-player scores across action counts, contrasted with the collapse of team scores from two to three actions, implies that the limiting factor in teams is not action-space size per se but the combinatorial spread of teammate compositions.
Model selection for decentralized LLM systems should therefore be guided by the strategic structure of the task, not by general-purpose capability rankings.

Reading between the lines

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

If shared-identity reasoning is the active mechanism, a direct test is to vary how explicitly the prompt states the counterpart's identity: revealing it should raise common-interest scores and leave or lower anti-coordination scores, without any communication.
The perturbation-then-discard method for three-action mixed Nash suggests the team baselines may be approximate; re-solving the original payoff matrices exactly could move several models from below-Nash to at-Nash, narrowing the reported degradation.
The hosted-local gap cannot be cleanly attributed to the models themselves, since backend and trial count differ; some of the frontier advantage could come from provider-side inference behavior rather than a coordination capability.
A combinatorial explanation for team failure predicts that giving models access to the team-composition belief distribution, or asking them to enumerate teammate splits explicitly, should recover part of the lost signal even without communication.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

The paper does something new and useful: it tests one-shot, no-communication, same-model self-play, with rotated action labels to kill position bias and a Nash-normalized score. The empirical finding that frontier hosted models exceed their Nash benchmark in matching-oriented archetypes but fail in anti-coordination, and that this ability degrades sharply in larger team games, is a solid contribution. The composition-based payoff reduction for teams and the breadth of models are also good choices.

The soft spot is real and load-bearing. Section III.C says that for the three-action suite they add a small random perturbation to payoffs during solving, then discard it and report the mixed solution against the original payoffs. No verification is described. A Lemke-Howson output for a perturbed game need not be an equilibrium of the original game, and if it is not, the score's zero point is a phantom. This matters because the headline claim spans action counts up to ten and the team tiers use three actions. The fix is easy: check the reported mixed profiles for approximate equilibrium against the original payoff matrices, or use a solver that handles degeneracy directly.

A second, softer issue: Eq. 1 sets the zero point to the worst symmetric equilibrium, not the full Nash set. That is a more forgiving baseline than the abstract's 'the Nash equilibrium' suggests. It is a defensible choice, but it should be stated explicitly so readers know the claim is 'beats the worst symmetric equilibrium,' not 'beats every Nash equilibrium.'

The trial counts for team games are low (8–9 per game), so the team-level conclusions are noisier than the two-player ones. That is a minor concern given the total of ~53k instances. Lack of code and prompts until acceptance is also common, but here it makes the baseline harder to audit.

Overall, the core idea and most of the empirical work hold up. The perturbation issue is addressable and should not sink the paper. This is a useful benchmark for anyone working on LLM agents in decentralized or communication-constrained settings, and it deserves a serious refereeing. I'd recommend sending it to peer review with the clear expectation that the authors verify their Nash baselines against original payoffs and clarify the baseline selection.

Referee Report

3 major / 5 minor

Summary. This paper introduces a benchmark for one-shot, no-communication multi-agent games in which independent copies of the same LLM choose actions simultaneously. The benchmark spans two-player matrix games (seven archetypes, k=2\u201310) and team games with two or three actions per agent, scoring each model's expected joint payoff relative to a Nash baseline normalized to zero. The central empirical claims are that two frontier-hosted models consistently exceed the Nash benchmark and approach optimal joint outcomes in dyadic games, while open-weight models show partial, archetype-dependent gains, and that team performance collapses when the action space grows. The paper also concludes that shared-identity reasoning helps in matching games but hurts in anti-coordination games, and that the small self-play gains in dyadic settings do not transfer to larger teams.

Significance. The question is well posed and the benchmark design has several strengths: externally computed game-theoretic benchmarks rather than fitted labels, explicit rotation of action labels to control position bias, separate normalization of upside and downside relative to the Nash zero point, and a model set spanning multiple families and scales. If the headline results hold, the finding that some models coordinate above the Nash baseline without communication\u2014and that this depends on whether the game requires matching or differentiating\u2014would be a useful empirical contribution for multi-agent LLM deployment. The paper's central quantitative contribution, however, rests on the correctness of the Nash baselines, and that correctness is currently not established for the three-action games.

major comments (3)

[III.C] The mixed-Nash benchmarks for the three-action suite are not verified as equilibria of the original games. The manuscript states that "for the three-action suite we add a small random perturbation to payoffs only during solving, then discard it and report the mixed solution against the original payoffs." A Lemke\u2013Howson output for a perturbed game need not be an equilibrium, or even an approximate equilibrium, of the unperturbed payoff matrix. Because Eq. (1) sets the zero point of every score to N, an invalid mixed equilibrium would make the "above Nash" scores meaningless for those games, and Section I's claim of "verified Nash benchmarks" would be unsupported. The authors should verify each reported mixed profile by computing the maximum unilateral deviation gain on the original payoffs, recompute equilibria with an exact method (e.g., support enumeration) on the original payoffs, or explicitly restrict the headline claims to pure equilibria and to k=2 games. The perturbation magnitude and random seed must also be reported.
[IV, Eq. (1)] The choice of baseline N as the worst (lowest-welfare) symmetric equilibrium is more forgiving than the "Nash equilibrium" language used in the abstract and Introduction. With this definition, a model can score above zero while being below other Nash equilibria. It is also unclear how the worst symmetric equilibrium is identified when Lemke\u2013Howson returns only one mixed equilibrium; if not all symmetric equilibria are enumerated, N may not actually be the worst. Given that the central claim is that models "exceed their Nash benchmark," the authors should either use the full set of Nash equilibria (e.g., report scores against the worst, mean, and best equilibrium welfare) or clearly define N as the worst symmetric equilibrium throughout the abstract and methodology, and show the sensitivity of the reported rankings to this baseline choice.
[V (team tiers)] The abstract's statement that "performance degrades substantially in team-based games with four or more interchangeable agents" is not directly supported for the two frontier-hosted models, because Section V states they were "omitted from the team tiers for cost reasons." The team results are therefore entirely from open-weight models, several of which were not the strongest dyadic performers, and the Conclusion's claim that the dyadic self-play capability "does not transfer to larger multi-agent teams" is an extrapolation. Please either run the hosted models on a representative subset of team games (especially the 3-action tier) or rephrase the abstract and conclusion to say that the degradation is documented for the open-weight models, and that the transfer claim for hosted models remains untested.

minor comments (5)

[IV] The exact prompt given to each agent is not quoted. The abstract says each model is "told only that its counterparts are running the same model," but the protocol describes only the payoff table and the instruction to maximize one's own reward; please include the full prompt (or an appendix) so that the shared-identity information is verifiable.
[Figures 2\u20135] The manuscript does not state whether the plotted scores include the 95% confidence intervals described in Section IV; please add error bars or state in the captions that intervals are omitted, and specify how the archetype averages over game sizes are computed.
[V] The claim that performance is "largely stable across action count" is supported only by visual inspection; provide a quantitative summary (e.g., the mean/median score change per additional action or a regression slope) for the k=2\u201310 range.
[IV] The sentence "Given this setup roughly 53,000 tests are run for the sake of this paper" is vague; specify the exact trial counts per (model, archetype, action count) and state whether the 53,000 figure refers to tests or game instances, since both terms are used elsewhere.
[Table II] The status column entries "Zero-sum" and "General-sum" are not mutually exclusive with the mechanism descriptions (e.g., zero-sum hybrids are later said to be non-constant-sum); please clarify the taxonomy or rename the statuses.

Circularity Check

0 steps flagged · score 1.0 of 10

No circular derivation: the Nash baselines are external game-theoretic computations and LLM scores are measured, not fitted; mild benchmark-selection choices raise correctness, not circularity.

full rationale

The paper's central claim, that some LLMs score above the Nash baseline in one-shot no-communication games, rests on a measurement pipeline rather than on a derivation that re-imports its own conclusion. The Nash benchmarks are produced by exhaustive pure-strategy search and the Lemke-Howson algorithm on the payoff matrices, independently of the LLM outputs, and the LLM scores in Eqs. (1)-(2) are computed from collected model distributions against those fixed baselines. No parameter of any model is fitted to the benchmark, and no equilibrium value is defined in terms of the model's behavior. The three-action mixed-Nash computation does add a random perturbation during solving and then discards it (Section III.C), which is a correctness threat for the baseline but not a circular step: the reported equilibrium is not defined as the LLM's output, and the LLM behavior is not used to construct it. Likewise, choosing the worst symmetric equilibrium as the normalization zero in Eq. (1) makes 'beat Nash' a more forgiving standard than the abstract's singular 'Nash equilibrium' suggests, but that is benchmark selection rather than a self-definitional reduction. The only self-citation, reference [3] in the robotics-motivation paragraph, is contextual and carries no load in the evaluation chain. The paper is therefore self-contained against external benchmarks; the noted baseline-verification gap should be treated as a correctness risk, not as circularity.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The central claim rests on standard game theory (Nash existence, Lemke-Howson), the interpretation of LLM output distributions as mixed strategies, and the choice of the worst symmetric equilibrium as the baseline. No new entities are postulated. The only hand-set numerical parameter is the payoff perturbation used in Nash solving, whose magnitude is unreported.

free parameters (1)

Payoff perturbation magnitude for Nash solving
Added to payoffs during Lemke-Howson solving for three-action games (Section III.C) to avoid ties; the magnitude is unreported and the resulting equilibrium is not verified against original payoffs.

assumptions (4)

standard math Every finite game has at least one mixed-strategy Nash equilibrium (Nash 1950).
Relied on to define the benchmark for all games in the suite.
standard math Lemke-Howson computes a mixed equilibrium for non-degenerate bimatrix games.
Used in Section III.C to compute mixed Nash benchmarks; correctness is assumed.
domain assumption An LLM's returned probability distribution over actions represents its mixed strategy in the one-shot game.
Expected joint payoff is calculated from the model output distributions (Section IV), so the measurement assumes the distribution is a faithful expression of the agent's strategy.
domain assumption The worst symmetric equilibrium is the appropriate uncoordinated-play baseline.
Section IV, Eq. 1 defines N as the lowest-welfare symmetric equilibrium; this choice biases the benchmark and is not the standard full Nash comparison.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Do LLMs Beat Nash? Testing Decentralized Coordination in Self-Play Multi-Agent Games." pith.science (2026). https://pith.science/paper/MEVN6O6A

@misc{pith2026260812547,
  author       = {Pith},
  title        = {Pith review of: Do LLMs Beat Nash? Testing Decentralized Coordination in Self-Play Multi-Agent Games},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MEVN6O6A}},
  note         = {Machine review of arXiv:2608.12547}
}

read the original abstract

Large language model agents deployed without a central controller are often assumed to require communication to coordinate their actions. We ask what remains possible without it: when independent instances of the same model cannot communicate, can they still reason about their counterparts well enough to exceed the standard game-theoretic baseline for uncoordinated play? We introduce a benchmark of one-shot, no-communication games in which each of thirteen language models is told only that its counterparts are running the same model and is evaluated against the Nash equilibrium of the underlying game. In two-player matrix games spanning seven archetypes and two to ten actions per player, two frontier-hosted models consistently exceed their Nash benchmark, approaching the optimal joint outcome in several archetypes, while most open-weight models achieve only partial gains that vary sharply by game structure. Performance degrades substantially in team-based games with four or more interchangeable agents, particularly as the action space grows, suggesting that whatever capability drives self-play gains in dyadic games does not transfer to larger multi-agent teams.

Figures

Figures reproduced from arXiv: 2608.12547 by the authors.

**Figure 2.** Tier 1 (two-player) games across action count ( [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Tier 1 (two-player) across archetypes, averaged over game sizes. [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗

Figures from the paper (1 more)

**Figure 5.** Figure 5: Tier 3 (3-action team) across archetypes. [PITH_FULL_IMAGE:figures/full_fig_p004_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

24 extracted references · 11 canonical work pages

[1]

Cooperative mobile robotics: Antecedents and directions,

Y . U. Cao, A. S. Fukunaga, and A. B. Kahng, “Cooperative mobile robotics: Antecedents and directions,”Autonomous Robots, vol. 4, no. 1, pp. 7–27, 1997, doi: 10.1023/A:1008855018923

work page doi:10.1023/a:1008855018923 1997
[2]

A formal analysis and taxonomy of task allocation in multi-robot systems,

B. P. Gerkey and M. J. Matari ´c, “A formal analysis and taxonomy of task allocation in multi-robot systems,”The International Jour- nal of Robotics Research, vol. 23, no. 9, pp. 939–954, 2004, doi: 10.1177/0278364904045564

doi:10.1177/0278364904045564 2004
[3]

Decentralized multi- agent goal assignment for path planning using large language models,

M. Ismayilov, E. Meriaux, S. Wen, and G. Dudek, “Decentralized multi- agent goal assignment for path planning using large language models,” in2025 IEEE MIT Undergraduate Research Technology Conference (URTC), 2025, pp. 1–5

work page 2025
[4]

Playing repeated games with large language models,

E. Akata, L. Schulz, J. Coda-Forno, S. J. Oh, M. Bethge, and E. Schulz, “Playing repeated games with large language models,”Nature Human Behaviour, vol. 9, pp. 1380–1390, 2025

2025
[5]

LLM-Coordination: Evaluating and analyzing multi-agent coordination abilities in large language models,

S. Agashe, Y . Fan, A. Reyna, and X. E. Wang, “LLM-Coordination: Evaluating and analyzing multi-agent coordination abilities in large language models,” inFindings of the Association for Compu- tational Linguistics: NAACL 2025, pp. 8053–8072, 2025, doi: 10.18653/v1/2025.findings-naacl.448

doi:10.18653/v1/2025.findings-naacl.448 2025
[6]

Equilibrium points inn-person games,

J. F. Nash, “Equilibrium points inn-person games,”Proceedings of the National Academy of Sciences, vol. 36, no. 1, pp. 48–49, 1950

1950
[7]

Subjectivity and correlation in randomized strategies,

R. J. Aumann, “Subjectivity and correlation in randomized strategies,” Journal of Mathematical Economics, vol. 1, no. 1, pp. 67–96, 1974

work page 1974
[8]

T. C. Schelling,The Strategy of Conflict. Cambridge, MA: Harvard University Press, 1960

1960

Show all 24 references

[9]

D. K. Lewis,Convention: A Philosophical Study. Cambridge, MA: Harvard University Press, 1969

1969
[10]

Dilemmas for superrational thinkers, leading up to a luring lottery,

D. R. Hofstadter, “Dilemmas for superrational thinkers, leading up to a luring lottery,”Scientific American, vol. 248, no. 6, Jun. 1983

1983
[11]

Program equilibrium,

M. Tennenholtz, “Program equilibrium,”Games and Economic Behav- ior, vol. 49, no. 2, pp. 363–373, 2004

2004
[12]

Other-play for zero- shot coordination,

H. Hu, A. Lerer, A. Peysakhovich, and J. Foerster, “Other-play for zero- shot coordination,” inProc. Int. Conf. Machine Learning (ICML), 2020, pp. 4399–4408

2020
[13]

GTBench: Uncovering the strategic reasoning capabilities of LLMs via game-theoretic evaluations,

J. Duan, R. Zhang, J. Diffenderfer, B. Kailkhura, L. Sun, E. Stengel- Eskin, M. Bansal, T. Chen, and K. Xu, “GTBench: Uncovering the strategic reasoning capabilities of LLMs via game-theoretic evaluations,” inAdvances in Neural Information Processing Systems (NeurIPS), 2024

2024
[14]

Equilibrium points of bimatrix games,

C. E. Lemke and J. T. Howson, Jr., “Equilibrium points of bimatrix games,”Journal of the Society for Industrial and Applied Mathematics, vol. 12, no. 2, pp. 413–423, 1964, doi: 10.1137/0112033

1964 doi
[15]

Gemini 3.5 Flash: Model card,

Google DeepMind, “Gemini 3.5 Flash: Model card,” May 2026. [Online]. Available: https://deepmind.google/models/model-cards/ gemini-3-5-flash/

2026
[16]

GPT-5.6 system card,

OpenAI, “GPT-5.6 system card,” Jul. 2026. [Online]. Available: https: //deploymentsafety.openai.com/gpt-5-6

2026
[17]

Mistral 7B,

A. Q. Jianget al., “Mistral 7B,” arXiv:2310.06825, 2023, doi: 10.48550/arXiv.2310.06825

2023 doi arXiv
[18]

Gemma 2: Improving open language models at a prac- tical size,

Gemma Team, “Gemma 2: Improving open language models at a prac- tical size,” arXiv:2408.00118, 2024, doi: 10.48550/arXiv.2408.00118

2024 doi arXiv
[19]

Gemma 3 technical report,

Gemma Team, “Gemma 3 technical report,” arXiv:2503.19786, 2025, doi: 10.48550/arXiv.2503.19786

2025 doi arXiv
[20]

The Llama 3 herd of models,

A. Grattafioriet al., “The Llama 3 herd of models,” arXiv:2407.21783, 2024, doi: 10.48550/arXiv.2407.21783

2024 doi arXiv
[21]

Llama 3.2: Revolutionizing edge AI and vision with open, customizable models,

Meta AI, “Llama 3.2: Revolutionizing edge AI and vision with open, customizable models,” Sep. 2024. [Online]. Available: https://ai.meta. com/blog/llama-3-2-connect-2024-vision-edge-mobile-devices/

2024
[22]

DeepSeek-R1: Incentivizing reasoning capability in LLMs via reinforcement learning,

DeepSeek-AI, “DeepSeek-R1: Incentivizing reasoning capability in LLMs via reinforcement learning,” arXiv:2501.12948, 2025, doi: 10.48550/arXiv.2501.12948

2025 doi arXiv
[23]

Phi-4-Mini technical report: Compact yet powerful mul- timodal language models via mixture-of-LoRAs,

Microsoft, “Phi-4-Mini technical report: Compact yet powerful mul- timodal language models via mixture-of-LoRAs,” arXiv:2503.01743, 2025, doi: 10.48550/arXiv.2503.01743

2025 doi arXiv
[24]

Gemma 4 technical report,

Gemma Team, “Gemma 4 technical report,” arXiv:2607.02770, 2026, doi: 10.48550/arXiv.2607.02770

2026 doi arXiv

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.

[1] [1]

Cooperative mobile robotics: Antecedents and directions,

Y . U. Cao, A. S. Fukunaga, and A. B. Kahng, “Cooperative mobile robotics: Antecedents and directions,”Autonomous Robots, vol. 4, no. 1, pp. 7–27, 1997, doi: 10.1023/A:1008855018923

work page doi:10.1023/a:1008855018923 1997

[2] [2]

A formal analysis and taxonomy of task allocation in multi-robot systems,

B. P. Gerkey and M. J. Matari ´c, “A formal analysis and taxonomy of task allocation in multi-robot systems,”The International Jour- nal of Robotics Research, vol. 23, no. 9, pp. 939–954, 2004, doi: 10.1177/0278364904045564

doi:10.1177/0278364904045564 2004

[3] [3]

Decentralized multi- agent goal assignment for path planning using large language models,

M. Ismayilov, E. Meriaux, S. Wen, and G. Dudek, “Decentralized multi- agent goal assignment for path planning using large language models,” in2025 IEEE MIT Undergraduate Research Technology Conference (URTC), 2025, pp. 1–5

work page 2025

[4] [4]

Playing repeated games with large language models,

E. Akata, L. Schulz, J. Coda-Forno, S. J. Oh, M. Bethge, and E. Schulz, “Playing repeated games with large language models,”Nature Human Behaviour, vol. 9, pp. 1380–1390, 2025

2025

[5] [5]

LLM-Coordination: Evaluating and analyzing multi-agent coordination abilities in large language models,

S. Agashe, Y . Fan, A. Reyna, and X. E. Wang, “LLM-Coordination: Evaluating and analyzing multi-agent coordination abilities in large language models,” inFindings of the Association for Compu- tational Linguistics: NAACL 2025, pp. 8053–8072, 2025, doi: 10.18653/v1/2025.findings-naacl.448

doi:10.18653/v1/2025.findings-naacl.448 2025

[6] [6]

Equilibrium points inn-person games,

J. F. Nash, “Equilibrium points inn-person games,”Proceedings of the National Academy of Sciences, vol. 36, no. 1, pp. 48–49, 1950

1950

[7] [7]

Subjectivity and correlation in randomized strategies,

R. J. Aumann, “Subjectivity and correlation in randomized strategies,” Journal of Mathematical Economics, vol. 1, no. 1, pp. 67–96, 1974

work page 1974

[8] [8]

T. C. Schelling,The Strategy of Conflict. Cambridge, MA: Harvard University Press, 1960

1960