Pith. sign in

REVIEW 3 major objections 6 minor 35 references

Cross-game transfer of strategic behaviour in fine-tuned language models is predictable from two equilibrium statistics, not payoff geometry.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 06:02 UTC pith:ZUBGLX3U

load-bearing objection Strict leave-one-game-out protocol is a real contribution; the ENT-SW claim is plausible but the arbitrary equilibrium selection and LOGO-selected hyperparameter need fixing before I'd trust the headline. the 3 major comments →

arxiv 2607.27536 v1 pith:ZUBGLX3U submitted 2026-07-30 cs.GT cs.AIcs.LGcs.MA

Strategy, Not Payoffs: A Behavioural Embedding of Normal-Form Games

Joshua Caiata , Sreepriya Pulyassary , Xiang Li , Kate Larson This is my paper
classification cs.GT cs.AIcs.LGcs.MA MSC 91A1091A26
keywords behavioural embeddingtransfer learningnormal-form gamesNash equilibrium entropybest-response switchinglarge language modelsfine-tuningleave-one-game-out
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper asks whether the way a language model's strategic behaviour changes after fine-tuning on one game can be predicted for other games. It proposes a two-feature behavioural embedding, ENT-SW, built from the normalized entropy of the Nash equilibrium and the entropy of the best-response distribution. Across 49 small models and 15 normal-form games, this embedding predicts measured transfer better than a one-hot game-identity baseline, both when a game is fully unseen and when only a game pair is unseen. Published structural embeddings—from payoff geometry, potential-harmonic decompositions, response graphs, and equilibrium-invariant representations—beat identity only in the easier interpolation setting, if at all. The authors conclude that transferability is governed by the decision-making behaviour a game demands, not by the shape of its payoff matrix.

Core claim

The central claim is that two scalar features computed directly from a game's payoff table—HNE, the normalized entropy of the selected Nash equilibrium strategy, and SBR, the normalized entropy of the distribution of pure best responses across opponent actions—capture the behavioural 'shape' that determines how equilibrium-alignment transfer between games unfolds. A weighted ridge regression pooling 49 LLMs and using the pair of game vectors, their absolute difference, and their element-wise product to predict log-relative transfer improvement outperforms the game-identity baseline under both strict leave-one-game-out (LOGO) and leave-one-pair-out (LOPO) cross-validation, and it retains marg

What carries the argument

ENT-SW is a two-dimensional, hand-crafted embedding of a game instance: (i) the normalized entropy of the Nash equilibrium policy (HNE), measuring whether equilibrium requires a pure choice or mixing, and (ii) the normalized entropy of the empirical best-response distribution over opponent actions (SBR), measuring whether optimal play is fixed or switches with the opponent. For each game, per-player vectors are averaged over sampled instances to give one point in the embedding space. A transfer pair is represented by concatenating the two game vectors with their absolute difference and element-wise product. The feature does its work without any learned representation or fit to model behaviou

Load-bearing premise

For games with multiple Nash equilibria, the paper fixes a single canonical mixed equilibrium that defines both the fine-tuning target and the entropy feature; if a different equilibrium is what these games actually demand of a player, the ENT feature—and with it the reported predictive advantage—may not survive.

What would settle it

Re-run the LOGO/LOPO protocol after swapping the selected equilibrium in Stag Hunt, Pure Coordination, and Battle of the Sexes to the payoff-dominant or risk-dominant pure equilibrium; if ENT-SW no longer beats game identity, the result is an artifact of equilibrium selection. A sharper falsifier already sits inside the paper: Harmony has identical ENT-SW coordinates to Prisoner's Dilemma, Deadlock, and Compromise, yet measured transfer between Harmony and those three is strongly negative—so any weighting of these two features that fits the Harmony row without sacrificing LOGO performance woul

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Fine-tuning on Chicken, Entry Game, or Pure Coordination improves the average target game, while Harmony, Trust Game, and Bertrand Duopoly are the weakest sources; the embedding orders games by equilibrium shape and predicts these differences.
  • Prisoner's Dilemma, Deadlock, and Compromise form a mutually compatible cluster—all share a dominant self-interested outcome with ENT=0 and SW=0—so a model tuned on any one transfers strongly to the other two.
  • Published structural embeddings (raw payoffs, potential–harmonic decomposition, response graphs, equilibrium-invariant representations) fail to beat game identity under strict leave-one-game-out, so the predictive signal is not payoff geometry.
  • Because game-pair identity explains 77.1% of transfer variance, a compact embedding offers the practical route to predicting transfer to a game never seen during training, where identity is unavailable.
  • The Harmony counterexample—same ENT-SW coordinates as the dominant-defection trio but opposite behaviour—shows the embedding describes equilibrium shape without naming which action the equilibrium points to; the paper treats this as a limit of the two-feature representation.

Where Pith is reading between the lines

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

  • A direct sensitivity test: recompute ENT and the transfer scores for Stag Hunt, Pure Coordination, and Battle of the Sexes using payoff-dominant or risk-dominant pure equilibria instead of the canonical mixed one; if LOGO performance drops, the reported advantage depends on that selection.
  • The Harmony–Prisoner's Dilemma pair already contradicts the embedding's equivalence claim—both have (0,0) features yet measured transfer between them is among the worst in the matrix—so adding a third feature encoding which action the equilibrium favours could repair this outlier while preserving LOGO gains.
  • The two-feature construction generalizes naturally to n-action and multi-player games, so the authors' untested claim about richer games is directly checkable: entropy over an n-action equilibrium and switching over n opponent actions are well-defined for any finite normal-form game.
  • If the result holds, it suggests a cheap curriculum for fine-tuning: select source games by ENT-SW distance to the target game, without running the expensive 49-model sweep.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper introduces ENT-SW, a two-feature representation of normal-form games: the normalized entropy of a selected Nash equilibrium and the entropy of the best-response distribution across opponent actions, aggregated per game. It uses this embedding as a predictor of log-relative transfer scores (LTS) obtained by fine-tuning 49 small LLMs on one game and evaluating on another. The authors report that ENT-SW outperforms a one-hot game-identity baseline under strict leave-one-game-out (LOGO) and leave-one-pair-out (LOPO) cross-validation, and that published structural embeddings (RSTP, PHD, RG, EI, BRI) fail under LOGO. A variance decomposition attributes 77.1% of transfer variance to game-pair structure and only 2.8% to LLM identity. The central claim is that transfer of strategic capabilities in LLMs is predictable from two hand-computed equilibrium statistics rather than from payoff geometry.

Significance. If the empirical claims are robust, this is a valuable contribution: it provides a simple, interpretable, parameter-free-in-principle embedding that predicts cross-game transfer under a demanding held-out protocol; it uses a large and diverse model roster; and it benchmarks against an identity baseline and published structural representations. The experimental protocol is a step beyond validating embeddings in-sample. The main results are stated with bootstrap CIs and include ablations. However, the correctness of the headline claim depends on two choices: the selection of a 'canonical' mixed equilibrium in multi-equilibrium games and the use of LOGO performance for model selection of the regularization strength. As written, these choices could mask an artifact. The contribution is potentially significant but needs additional robustness analysis before the strong generalization claim can be accepted.

major comments (3)
  1. [§3.3, §5.1, Appendix A (Table 1)] The features ENT and SW and the evaluation target NE-MSE are all defined with respect to a single selected equilibrium per game. For games with multiple Nash equilibria (Stag Hunt, Chicken, Pure Coordination, Battle of the Sexes, Entry Game, Inspection Game; see Appendix A), this selection is not justified. Stag Hunt, for example, also has payoff-dominant and risk-dominant pure Nash equilibria; choosing a pure equilibrium would give ENT=0 instead of the reported mixed value, and the NE-MSE response would change because the target distribution changes. While SW is computed from best responses and is not directly affected by the equilibrium selection, ENT and the response both depend on the choice. Because a change of convention changes both predictor and response, the reported LOGO advantage may be an artifact of this particular convention. No sensitivity analysis over equilibrium selecti
  2. [§5.5, §7 (LOGO/LOPO results)] The regularization strength λ=1000 is selected from a fixed grid based on LOGO performance and then reused across experiments. If the same LOGO folds are used to select λ and to report the improvements in §7, the reported numbers (e.g., 0.360 and 0.370 under LOGO) are not clean out-of-sample estimates: the model-selection procedure has already seen the held-out games. This is especially important for the strict-LOGO claim, since the paper's central argument is that the embedding generalizes to unseen games. I would like to see a nested cross-validation or an independent validation split for λ selection before accepting that the LOGO advantage is real. LOPO is less affected, but the same issue applies to the pooled comparison across both regimes.
  3. [§5.5, identity baseline] The paper calls the one-hot game-identity baseline 'fully flexible,' but in the pooled ridge framework it is penalized with the same λ=1000 that is applied to the continuous embedding features. Because the identity features are binary and the embedding features are z-scored, a single λ does not impose comparable shrinkage across the two feature types. This could make the identity baseline artificially weak and thereby inflate the relative advantage of ENT-SW. The authors should either tune λ separately for each feature representation or report results under comparable regularization scales, so that the comparison reflects information content rather than differential shrinkage.
minor comments (6)
  1. [§3.3] The phrase 'we adopt the mixed equilibrium described in Appendix A' implies a principled criterion; the paper should instead state that this is a selection and explain why this particular equilibrium is chosen across games with multiple equilibria.
  2. [Figure 2] The caption states 'Filled markers exclude 0'; it should define what filled vs. empty markers mean, otherwise the visual encoding is ambiguous.
  3. [§5.5] The family-balanced weight w_m is described verbally; a formula and a table of family sizes would make the weighting scheme reproducible.
  4. [§5.4] The offset c≈0.003 is said to be computed from the median standard error over Monte Carlo trials. The paper should state how sensitive the LTS values and downstream results are to this choice.
  5. [§8] The Harmony discussion is honest and informative, but it shows that ENT-SW cannot distinguish behaviorally opposite games (e.g., Harmony vs. Prisoner's Dilemma). This limitation should be stated more prominently in the abstract or introduction, since it directly qualifies the phrase 'behavioural demands.'
  6. [General] The paper does not provide a data or code availability statement. Given the scale of the experiments (49 LLMs, 480,000 trials per LLM), a release of the generation, fine-tuning, and evaluation code would substantially improve reproducibility.

Circularity Check

0 steps flagged

No circularity: ENT-SW is a parameter-free hand-computed embedding; transfer predictions are out-of-sample and the same-equilibrium coupling is a limitation, not a reduction.

full rationale

ENT-SW is computed solely from each game's payoff table via hand-defined formulas (Eqs. 1–2), with no parameters fitted to the transfer data; the paper states 'neither feature is fit to model behaviour' and the transfer scores LTS are defined from NE-MSE improvements on held-out target instances. The ridge coefficients are fit on training folds and evaluated under strict LOGO/LOPO, so the reported prediction is genuinely out-of-sample and is benchmarked against a one-hot identity baseline; the fact that ENT-SW adds signal even when identity receives full credit is empirical evidence against a definitional collapse. The same selected Nash equilibrium appears in both the ENT feature and the MSE target, creating a conceptual coupling in multi-equilibrium games (Stag Hunt, Chicken, BoS, PCo), but this is not a circular reduction: ENT is only a summary statistic, and the paper's own Harmony outlier (HNE=0, SBR=0, yet behaviourally nearly opposite to PD/Deadlock/Compromise) shows the features do not determine transfer outcomes. The paper's explicit limitations (Section 10) and the lambda selection note ('selected from a fixed grid based on LOGO performance') are robustness/leakage caveats, not circular steps. Self-citations in the references are background and not load-bearing. Therefore no circularity is found.

Axiom & Free-Parameter Ledger

2 free parameters · 4 axioms · 0 invented entities

The embedding itself is parameter-free. The central modeling choices are the ridge penalty λ (selected using LOGO performance), the log offset c (from Monte Carlo standard errors), and the selection of a single mixed equilibrium for multi-equilibrium games, which jointly determine the measured transfer and the ENT feature.

free parameters (2)
  • Ridge regularization λ = 1000
    Selected from a fixed grid based on LOGO performance (§5.5); potential selection on the evaluation protocol.
  • Log-transfer offset c = ≈0.003
    Set using median standard error over Monte Carlo trials (§5.4); stabilizes log-relative transfer scores.
axioms (4)
  • standard math Every finite normal-form game has a Nash equilibrium (Nash 1951).
    Invoked in §3.3 to justify existence of behavioural targets.
  • ad hoc to paper For games with multiple equilibria, the mixed equilibrium specified in Appendix A is the correct behavioural target for both fine-tuning and evaluation.
    A modeling choice; different selections would change ENT values (e.g., Stag Hunt, Pure Coordination, Battle of the Sexes). Not derived from first principles.
  • domain assumption The log-relative transfer score LTS(m,i,j) captures meaningful strategic-capability transfer; MSE to the selected Nash equilibrium is an adequate behavioural distance.
    Evaluation metric choice; alternative metrics (e.g., expected payoff, best-response match) might yield different transfer patterns.
  • domain assumption Per-game payoff constraints (e.g., T>R>P>S in PD) define the game families and sampling distributions.
    Game definitions follow literature (§5.1, Appendix A).

pith-pipeline@v1.3.0-daily-deepseek · 12802 in / 14495 out tokens · 132449 ms · 2026-08-01T06:02:55.105117+00:00 · methodology

0 comments
read the original abstract

Learning a strategic task changes more than what is directly taught: fine-tuning on one game can either enhance or degrade an agent's ability to reason in another. Understanding and predicting this transfer of strategic capabilities, however, remains a key challenge for large language models (LLMs). Normal-form games provide an ideal testbed for analyzing this phenomenon, as they feature explicitly defined payoffs and well-characterized equilibrium behaviours. In this work, we investigate whether game embeddings can explain and predict changes in LLM strategic capabilities following fine-tuning across different games. We propose a lightweight two-feature embedding that captures fundamental behavioural demands: the entropy of the Nash equilibrium and the sensitivity of optimal responses to an opponent's action. We show that while existing published structural embeddings primarily memorize game identities and fail to generalize, our behavioural embedding reliably predicts performance changes on held-out games. These results demonstrate that the transfer of strategic capabilities in LLMs is not dictated by the payoff geometry of a game, but by the underlying structure of the decision-making behaviour it requires.

Figures

Figures reproduced from arXiv: 2607.27536 by Joshua Caiata, Kate Larson, Sreepriya Pulyassary, Xiang Li.

Figure 1
Figure 1. Figure 1: Overview. (a) 15 games embedded by Nash-equilibrium entropy (ENT) and best-response [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Held-out squared-error reduction relative to game identity under (a) LOGO and (b) [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Mean log-relative transfer across LLMs. Rows are source games and columns are targets, [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

35 extracted references · 1 canonical work pages

  1. [1]

    Fowlkes, Stefano Soatto, and Pietro Perona

    Alessandro Achille, Michael Lam, Rahul Tewari, Avinash Ravichandran, Subhransu Maji, Char- less C. Fowlkes, Stefano Soatto, and Pietro Perona. Task2vec: Task embedding for meta- learning. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. arXiv:1902.03545

    Pith/arXiv arXiv 2019

  2. [2]

    How stable are transferability metrics evaluations? InProceedings of the European Conference on Computer Vision (ECCV), pages 303–321, 2022

    Andrea Agostinelli, Michal P´ andy, Jasper Uijlings, Thomas Mensink, and Vittorio Ferrari. How stable are transferability metrics evaluations? InProceedings of the European Conference on Computer Vision (ECCV), pages 303–321, 2022. arXiv:2204.01403

    Pith/arXiv arXiv 2022

  3. [3]

    Playing repeated games with large language models.Nature Human Behaviour, 9(7):1380–1390,

    Elif Akata, Lion Schulz, Julian Coda-Forno, Seong Joon Oh, Matthias Bethge, and Eric Schulz. Playing repeated games with large language models.Nature Human Behaviour, 9(7):1380–1390,

  4. [4]

    Geometric dataset distances via optimal transport

    David Alvarez-Melis and Nicol` o Fusi. Geometric dataset distances via optimal transport. In Advances in Neural Information Processing Systems (NeurIPS), 2020. arXiv:2002.02923

    Pith/arXiv arXiv 2020

  5. [5]

    Camerer.Behavioral Game Theory: Experiments in Strategic Interaction

    Colin F. Camerer.Behavioral Game Theory: Experiments in Strategic Interaction. Princeton University Press, 2003

    2003

  6. [6]

    Ozan Candogan, Ishai Menache, Asuman Ozdaglar, and Pablo A. Parrilo. Flows and decompo- sitions of games: Harmonic and potential games.Mathematics of Operations Research, 36(3): 474–503, 2011. 12

    2011

  7. [7]

    Cardenal, and Max Pellert

    Andrea Cera Palatsi, Samuel Martin-Gutierrez, Ana S. Cardenal, and Max Pellert. Large language models replicate and predict human cooperation across experiments in game theory. arXiv preprint arXiv:2511.04500, 2025. arXiv:2511.04500

    Pith/arXiv arXiv 2025

  8. [8]

    McKee, Joel Z

    Allan Dafoe, Edward Hughes, Yoram Bachrach, Tantum Collins, Kevin R. McKee, Joel Z. Leibo, Kate Larson, and Thore Graepel. Open problems in cooperative AI.arXiv preprint arXiv:2012.08630, 2020. arXiv:2012.08630

    Pith/arXiv arXiv 2012

  9. [9]

    Cooperative AI: Machines must learn to find common ground.Nature, 593(7857):33–36, 2021

    Allan Dafoe, Yoram Bachrach, Gillian Hadfield, Eric Horvitz, Kate Larson, and Thore Graepel. Cooperative AI: Machines must learn to find common ground.Nature, 593(7857):33–36, 2021

    2021

  10. [10]

    Task2box: Box embeddings for modeling asymmetric task relationships

    Rangel Daroya, Aaron Sun, and Subhransu Maji. Task2box: Box embeddings for modeling asymmetric task relationships. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 28827–28837, 2024. arXiv:2403.17173

    Pith/arXiv arXiv 2024

  11. [11]

    Gtbench: Uncovering the strategic reasoning limitations of llms via game-theoretic evaluations

    Jinhao Duan, Renming Zhang, James Diffenderfer, Bhavya Kailkhura, Lichao Sun, Elias Stengel-Eskin, Mohit Bansal, Tianlong Chen, and Kaidi Xu. Gtbench: Uncovering the strategic reasoning limitations of llms via game-theoretic evaluations. InAdvances in Neural Information Processing Systems 37 (NeurIPS 2024), 2024. arXiv:2402.12348

    Pith/arXiv arXiv 2024

  12. [12]

    Can large language models serve as rational players in game theory? a systematic analysis

    Caoyun Fan, Jindou Chen, Yaohui Jin, and Hao He. Can large language models serve as rational players in game theory? a systematic analysis. InProceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 17960–17967, 2024. arXiv:2312.05488

    Pith/arXiv arXiv 2024

  13. [13]

    Kanishk Gandhi, Dorsa Sadigh, and Noah D. Goodman. Strategic reasoning with language models.arXiv preprint arXiv:2305.19165, 2023. arXiv:2305.19165

    Pith/arXiv arXiv 2023

  14. [14]

    Steering language models with game-theoretic solvers

    Ian Gemp, Roma Patel, Yoram Bachrach, Marc Lanctot, Vibhavari Dasagi, Luke Marris, Geor- gios Piliouras, Siqi Liu, and Karl Tuyls. Steering language models with game-theoretic solvers. arXiv preprint arXiv:2402.01704, 2024. arXiv:2402.01704

    Pith/arXiv arXiv 2024

  15. [15]

    Leakage and the reproducibility crisis in machine- learning-based science.Patterns, 4(9):100804, 2023

    Sayash Kapoor and Arvind Narayanan. Leakage and the reproducibility crisis in machine- learning-based science.Patterns, 4(9):100804, 2023. doi: 10.1016/j.patter.2023.100804. arXiv:2207.07048

    arXiv 2023

  16. [16]

    Game of thoughts: Iterative reasoning in game-theoretic domains with large language models

    Benjamin Kempinski, Ian Gemp, Kate Larson, Marc Lanctot, Yoram Bachrach, and Tal Kach- man. Game of thoughts: Iterative reasoning in game-theoretic domains with large language models. InProceedings of the 24th International Conference on Autonomous Agents and Mul- tiagent Systems (AAMAS), pages 1088–1097, Richland, SC, 2025. IF AAMAS

    2025

  17. [17]

    Exploring the effectiveness and consistency of task selection in intermediate-task transfer learning

    Pin-Jie Lin, Miaoran Zhang, Marius Mosbach, and Dietrich Klakow. Exploring the effectiveness and consistency of task selection in intermediate-task transfer learning. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 4: Student Research Workshop), pages 170–185, 2024. arXiv:2407.16245

    Pith/arXiv arXiv 2024

  18. [18]

    Nfgtransformer: Equiv- ariant representation learning for normal-form games

    Siqi Liu, Luke Marris, Georgios Piliouras, Ian Gemp, and Nicolas Heess. Nfgtransformer: Equiv- ariant representation learning for normal-form games. InProceedings of the 12th International Conference on Learning Representations (ICLR), 2024. arXiv:2402.08393

    Pith/arXiv arXiv 2024

  19. [20]

    Equilibrium-invariant embedding, metric space, and fundamental set of 2×2 normal-form games.arXiv preprint arXiv:2304.09978,

    Luke Marris, Ian Gemp, and Georgios Piliouras. Equilibrium-invariant embedding, metric space, and fundamental set of 2×2 normal-form games.arXiv preprint arXiv:2304.09978,

    Pith/arXiv arXiv

  20. [21]

    A general and simple method for obtaining R 2 from generalized linear mixed-effects models.Methods in Ecology and Evolution, 4(2):133–142, 2013

    Shinichi Nakagawa and Holger Schielzeth. A general and simple method for obtaining R 2 from generalized linear mixed-effects models.Methods in Ecology and Evolution, 4(2):133–142, 2013. doi: 10.1111/j.2041-210x.2012.00261.x

    Pith/arXiv arXiv 2013

  21. [22]

    Non-cooperative games.Annals of Mathematics, 54(2):286–295, 1951

    John Nash. Non-cooperative games.Annals of Mathematics, 54(2):286–295, 1951

    1951

  22. [23]

    Nguyen, Tal Hassner, Matthias Seeger, and C´ edric Archambeau

    Cuong V. Nguyen, Tal Hassner, Matthias Seeger, and C´ edric Archambeau. Leep: A new measure to evaluate transferability of learned representations. InProceedings of the 37th Inter- national Conference on Machine Learning (ICML), 2020. arXiv:2002.12462

    Pith/arXiv arXiv 2020

  23. [24]

    Czarnecki, Marc Lanctot, Julien P´ erolat, and R´ emi Munos

    Shayegan Omidshafiei, Christos Papadimitriou, Georgios Piliouras, Karl Tuyls, Mark Rowland, Jean-Baptiste Lespiau, Wojciech M. Czarnecki, Marc Lanctot, Julien P´ erolat, and R´ emi Munos. α-rank: Multi-agent evaluation by evolution.Scientific Reports, 9(1):9937, 2019. doi: 10.1038/ s41598-019-45619-9. arXiv:1903.01373

    Pith/arXiv arXiv 2019

  24. [25]

    Czarnecki, Francisco C

    Shayegan Omidshafiei, Karl Tuyls, Wojciech M. Czarnecki, Francisco C. Santos, Mark Rowland, Jerome Connor, Daniel Hennes, Paul Muller, Julien P´ erolat, Bart De Vylder, Audrunas Gruslys, and R´ emi Munos. Navigating the landscape of multiplayer games.Nature Communications, 11 (1):5603, Nov 2020. ISSN 2041-1723. doi: 10.1038/s41467-020-19244-4. URLhttps://...

    work page doi:10.1038/s41467-020-19244-4 2020

  25. [26]

    O’Brien, Carrie J

    Joon Sung Park, Joseph C. O’Brien, Carrie J. Cai, Meredith Ringel Morris, Percy Liang, and Michael S. Bernstein. Generative agents: Interactive simulacra of human behavior. In Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology (UIST), 2023. arXiv:2304.03442

    Pith/arXiv arXiv 2023

  26. [27]

    Strategic intelligence in large language models: Evi- dence from evolutionary game theory.arXiv preprint arXiv:2507.02618, 2025

    Kenneth Payne and Baptiste Alloui-Cros. Strategic intelligence in large language models: Evi- dence from evolutionary game theory.arXiv preprint arXiv:2507.02618, 2025. arXiv:2507.02618

    Pith/arXiv arXiv 2025

  27. [28]

    Anatol Rapoport and Melvin J. Guyer. A taxonomy of 2×2 games.General Systems: Yearbook of the Society for General Systems Research, 11:203–214, 1966

    1966

  28. [29]

    Routledge, 2005

    David Robinson and David Goforth.The Topology of the2×2Games: A New Periodic Table. Routledge, 2005

    2005

  29. [30]

    Shrout and Joseph L

    Patrick E. Shrout and Joseph L. Fleiss. Intraclass correlations: Uses in assessing rater reliability. Psychological Bulletin, 86(2):420–428, 1979

    1979

  30. [31]

    Game theory meets large language models: A systematic survey

    Haoran Sun, Yusen Wu, Yukun Cheng, and Xu Chu. Game theory meets large language models: A systematic survey. InProceedings of the Thirty-Fourth International Joint Conference on Artificial Intelligence (IJCAI), pages 10669–10677, 2025. arXiv:2502.09053

    Pith/arXiv arXiv 2025

  31. [32]

    Exploring and predicting transferability across nlp tasks

    Tu Vu, Tong Wang, Tsendsuren Munkhdalai, Alessandro Sordoni, Adam Trischler, Andrew Mattarella-Micke, Subhransu Maji, and Mohit Iyyer. Exploring and predicting transferability across nlp tasks. InProceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), 2020. arXiv:2005.00770

    Pith/arXiv arXiv 2020

  32. [33]

    Wellman, Karl Tuyls, and Amy Greenwald

    Michael P. Wellman, Karl Tuyls, and Amy Greenwald. Empirical game-theoretic analysis: A survey.Journal of Artificial Intelligence Research, 82:1017–1076, 2025. 14

    2025

  33. [34]

    Logme: Practical assessment of pre-trained models for transfer learning

    Kaichao You, Yong Liu, Jianmin Wang, and Mingsheng Long. Logme: Practical assessment of pre-trained models for transfer learning. InProceedings of the 38th International Conference on Machine Learning (ICML), 2021. arXiv:2102.11005

    Pith/arXiv arXiv 2021

  34. [35]

    Zamir, Alexander Sax, William Shen, Leonidas Guibas, Jitendra Malik, and Silvio Savarese

    Amir R. Zamir, Alexander Sax, William Shen, Leonidas Guibas, Jitendra Malik, and Silvio Savarese. Taskonomy: Disentangling task transfer learning. InProceedings of the IEEE Con- ference on Computer Vision and Pattern Recognition (CVPR), 2018. arXiv:1804.08328. 15 A Game Suite and Nash Equilibrium Targets Table 1: Game suite and Nash equilibrium targets us...

    Pith/arXiv arXiv 2018

  35. [2024]

    arXiv:2408.05241. 13

    Pith/arXiv arXiv