# Nash equilibrium

> Source: https://aiwiki.ai/wiki/nash_equilibrium
> Updated: 2026-07-24
> Fact-checked: 2026-07-24
> Categories: Algorithms, Computer Science, Mathematics, Reinforcement Learning
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "Nash equilibrium." aiwiki.ai, 24 Jul 2026. https://aiwiki.ai/wiki/nash_equilibrium
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

A Nash equilibrium is a combination of strategies, one for each player in a game, such that no player can raise their own payoff by changing strategy alone while everyone else keeps theirs unchanged. The concept comes from John Nash, who called it an "equilibrium point" and set it out in two papers published in 1950 and 1951 [1][2]. In Nash's own words, an equilibrium point is a strategy profile "such that each player's mixed strategy maximizes his payoff if the strategies of the others are held fixed", so that "each player's strategy is optimal against those of the others" [1].

Nash also proved that such a point always exists. Theorem 1 of the 1951 paper reads, in full, "Every finite game has an equilibrium point", provided players are allowed to randomize over their available actions [1]. That result is what made the idea usable as a general prediction about strategic behavior across all of [game theory](https://aiwiki.ai/wiki/game_theory), rather than a property of the two-player zero-sum games John von Neumann had analyzed in the 1920s [4]. Nash shared the 1994 Sveriges Riksbank Prize in Economic Sciences in Memory of Alfred Nobel with John Harsanyi and Reinhard Selten, "for their pioneering analysis of equilibria in the theory of non-cooperative games" [3]. The name "Nash equilibrium" came later and honors the existence result; Nash's own papers speak only of equilibrium points [1][4].

For [artificial intelligence](https://aiwiki.ai/wiki/artificial_intelligence) the concept cuts two ways. It is the target that superhuman poker programs, adversarial training procedures, and a growing family of language-model alignment methods are built to approximate [8][13][21]. It is also an obstacle: computing a Nash equilibrium of a general two-player game is PPAD-complete [4], which is one reason practical systems settle for approximate equilibria of drastically simplified games, and why systems built for games with more than two players often set the solution concept aside [12].

## Definition

A finite game in Nash's formulation has n players, each with a finite set of pure strategies and a payoff function mapping every combination of pure strategies to a real number. A mixed strategy is a [probability distribution](https://aiwiki.ai/wiki/probability_distribution) over one player's pure strategies, and payoffs extend to mixed profiles by taking expectations [1]. Writing s for a profile of mixed strategies, s is an equilibrium point when, for every player i, no alternative strategy r_i gives i a higher expected payoff against the others' fixed choices [1].

Two consequences follow directly from the definition and are used by algorithms that search for equilibria. First, a player's mixed strategy is optimal exactly when every pure strategy it plays with positive probability is itself a best reply, which turns the equilibrium condition into a finite set of comparisons rather than an optimization over a continuum [1]. Second, "no equilibrium point can involve a dominated strategy", so strictly dominated actions can be deleted before the search begins [1].

An equilibrium is called pure when every player's strategy places all its probability on a single action, and mixed otherwise. Nash's existence proof needs mixing: many ordinary games have no pure equilibrium at all. He also proved a symmetry counterpart, Theorem 2, that "any finite game has a symmetric equilibrium point", so a game whose rules treat players interchangeably always admits an equilibrium in which they behave interchangeably [1].

## Existence and its proof

Nash gave two proofs. The 1950 note in the Proceedings of the National Academy of Sciences derived existence from Kakutani's generalized fixed point theorem, an approach suggested to him by David Gale [1][2]. The 1951 Annals of Mathematics paper replaced it with what Nash called "a considerable improvement over that earlier version": a continuous map on the space of strategy profiles whose fixed points are exactly the equilibrium points, so that Brouwer's fixed point theorem suffices [1]. The map shifts each player's mixture toward whichever pure strategies currently beat their average payoff, then renormalizes, so a profile survives it unchanged only when no such improving strategy exists [1].

Finiteness matters. The theorem is stated for games with finitely many pure strategies per player, and the guarantee is for mixed strategies. Equilibria are also known to exist in many infinite games, but not automatically, and locating one can be difficult even when existence is assured [12].

The two-player zero-sum case is a special instance rather than a separate theory. Nash observed that "the set of equilibrium points of a two-person zero-sum game is simply the set of all pairs of opposing 'good strategies'", the optimal strategies of von Neumann and Morgenstern's earlier theory [1]. This has a practical payoff that carries over to modern game-playing systems: in a balanced two-player zero-sum game, a player who follows an equilibrium strategy cannot lose in expectation whatever the opponent does, which makes the strategy unbeatable without requiring any model of the opponent [12].

## Standard examples

| Game | Equilibria | What it illustrates |
| --- | --- | --- |
| Prisoner's dilemma | One, both players defect | An equilibrium can leave every player worse off than an available alternative; each does better when both cooperate, but cooperation is not an equilibrium [4] |
| Rock-paper-scissors | Mixed only, both randomize uniformly | No pure equilibrium exists; the best any opponent can achieve against the equilibrium mixture is a tie in expectation [4][12] |
| Matching pennies | Mixed only, each player heads with probability 1/2 | A zero-sum game with no pure equilibrium; alternating best replies cycle through the four profiles rather than converging [5] |
| Chicken | Three, two pure and one mixed | Multiple equilibria create a selection problem: knowing the game does not tell a player which equilibrium the others are aiming at [4] |
| Nash's three-man poker game | One | Nash's own worked application, solved by eliminating dominated behavior parameters and then solving the resulting algebraic system; at the unique equilibrium the first player opens on a high card with probability 0.308, and the three seats have game values of about -0.147, -0.096 and 0.243 [1] |

The poker example is not incidental. Nash devoted a section near the end of the 1951 paper to it and wrote that "poker is the most obvious target" for the theory, with a caveat that anticipated the computational turn the subject would take: "The complexity of the mathematical work needed for a complete investigation increases rather rapidly, however, with increasing complexity of the game; so that analysis of a game much more complex than the example given here might only be feasible using approximate computational methods" [1].

## Computing equilibria

Two-player zero-sum games are easy. Their equilibria can be written as a [linear program](https://aiwiki.ai/wiki/linear_programming) and solved in polynomial time [4]. Everything beyond that is harder.

For general two-player games given as a payoff matrix, the classical method is the Lemke-Howson algorithm, a pivoting procedure related to the simplex method. It always terminates, but Rahul Savani and Bernhard von Stengel constructed bimatrix games on which its path is exponentially long [5].

The complexity question was settled in the mid-2000s. Because Nash's theorem guarantees that a solution always exists, the problem cannot be NP-complete in the usual way: NP-hardness proofs for search problems lean on the possibility that no solution exists [4]. Instead the problem was placed in PPAD, a class Christos Papadimitriou defined in 1994 for problems whose solutions are guaranteed by a parity argument on a directed graph, the same argument that underlies Brouwer fixed points [5]. Constantinos Daskalakis, Paul Goldberg and Papadimitriou proved that finding a Nash equilibrium is PPAD-complete, first for games with four or more players; the three-player case followed within months, and Xi Chen and Xiaotie Deng then extended it to two-player games [4][5].

| Problem | Status |
| --- | --- |
| Equilibrium of a two-player zero-sum game | Polynomial time via linear programming [4] |
| Equilibrium of a general two-player (bimatrix) game | PPAD-complete [4][5] |
| Equilibrium maximizing total payoff, or using a specified strategy with positive probability | NP-complete [4] |
| Epsilon-approximate equilibrium, two players | No polynomial-time algorithm under the exponential time hypothesis for PPAD; a quasi-polynomial lower bound matching the best known algorithm [7] |
| Correlated equilibrium | Computable in polynomial time for games given by explicit payoff matrices, since correlated equilibria form a convex set [5] |

Approximation does not rescue the problem, which matters because every large-scale AI system computes an approximation rather than an exact equilibrium. Aviad Rubinstein's 2016 result shows that for some constant epsilon, computing an epsilon-approximate equilibrium of a two-player game requires quasi-polynomial time under an exponential time hypothesis for PPAD, which rules out a polynomial-time algorithm and matches the best known algorithm up to lower-order terms [7]. Weaker solution concepts are the practical escape route: simple no-regret learning dynamics converge to the coarse correlated equilibria of any finite game, and to Nash equilibrium in the two-player zero-sum case [18]. Much of what looks like equilibrium computation in [machine learning](https://aiwiki.ai/wiki/machine_learning) is regret minimization aimed at that weaker target.

## Price of anarchy

Nash equilibria are stable, not efficient. Elias Koutsoupias and Papadimitriou proposed measuring the gap in 1999, defining what they called the coordination ratio: "the ratio between the worst possible Nash equilibrium and the social optimum as a measure of the effectiveness of the system" [5][6]. Papadimitriou introduced the name price of anarchy for it in 2001 [5].

The canonical case is network routing. In Pigou's two-link example the price of anarchy is exactly 4/3, and 4/3 remains an upper bound for every network whose edge costs are affine functions of load, however large or complex the network [5]. With arbitrary cost functions the ratio can be unbounded even on the simplest network [5]. Braess's paradox, discovered by Braess in 1968, is the sharpest illustration: adding an intuitively helpful edge to a congested network can make every driver worse off at equilibrium, and with affine costs the damage is bounded by the same 4/3 factor [5]. A companion measure, the price of stability, compares the best equilibrium rather than the worst to the optimum [5].

## Nash equilibrium in AI systems

### Self-play and game-playing programs

The strongest link between the concept and modern AI runs through two-player zero-sum games, where an equilibrium strategy cannot be beaten. Counterfactual regret minimization (CFR), introduced by Martin Zinkevich, Michael Johanson, Michael Bowling and Carmelo Piccione in 2007, minimizes regret independently at each decision point in an extensive-form game and provably drives the average strategy to a Nash equilibrium in two-player zero-sum games [8]. It scaled to poker abstractions with up to 10^12 states on first publication [8].

| System | Year | Result | Relationship to equilibrium |
| --- | --- | --- | --- |
| Bowling, Burch, Johanson and Tammelin | 2015 | Heads-up limit Texas hold'em "essentially weakly solved" using CFR+ | Direct equilibrium approximation of a game with about 10^13 decision points [9][11] |
| Libratus (Brown and Sandholm) | 2017 | Beat four top-10 heads-up no-limit specialists over 120,000 hands in 20 days, by 147 mbb/hand at 99.98% statistical significance | Blueprint equilibrium approximation, nested subgame solving, and a self-improvement module; about 25 million core hours [10][11] |
| Pluribus (Brown and Sandholm) | 2019 | Beat elite professionals over 10,000 hands of six-player no-limit hold'em, 48 mbb/game with standard error 25 and p = 0.028 | Explicitly abandons Nash equilibrium as the goal; blueprint cost about $144 of compute [12] |
| DeepNash (Perolat et al.) | 2022 | Human expert play at Stratego, with a 2022 and all-time top-3 rank on the Gravon platform | Regularised Nash Dynamics, a model-free method that converges to an approximate equilibrium rather than cycling around it, and uses no search [16] |
| [CICERO](https://aiwiki.ai/wiki/cicero) (Meta FAIR) | 2022 | More than twice the average human score across 40 online Diplomacy games, top 10% of players with multiple games | Equilibrium-finding regularized toward human play (piKL), since equilibrium alone gives no guarantee in a seven-player game [17][18] |

Libratus marked the first time an AI beat top humans at heads-up no-limit hold'em, a game with roughly 10^161 decision points that its abstraction cut to about 10^12 [11]. Its successor [Pluribus](https://aiwiki.ai/wiki/pluribus) is the clearest published statement of the concept's limits. The 2019 experiment ran five human professionals against one copy of Pluribus, the five drawn each day from a pool of 13 players over 12 days, plus a second format in which Chris Ferguson and Darren Elias each played 5,000 hands against five copies of the bot; Pluribus won 32 mbb/game across those 10,000 hands, at p = 0.014 [12]. Brown and Sandholm write that in six-player poker "our goal should not be a specific game-theoretic solution concept but rather to create an AI that empirically consistently defeats human opponents", because outside two-player zero-sum games a profile assembled from independently computed equilibria need not be an equilibrium at all [12]. Their illustration is the Lemonade Stand Game, where players placing themselves around a ring want to be far from everyone else: the equilibria are exactly the evenly spaced configurations, of which there are infinitely many, and independent choices almost never line up [12].

The compute figures cut against the usual expectations for a superhuman game AI. Pluribus trained its blueprint strategy in 8 days on a 64-core server, 12,400 CPU core hours in total and under 512 GB of memory, which the authors priced at about $144 at cloud computing spot instance rates, and it played on two Intel Haswell E5-2695 v3 CPUs using less than 128 GB of memory [12]. The same paper notes for comparison that [AlphaGo](https://aiwiki.ai/wiki/alphago) used 1,920 CPUs and 280 GPUs for real-time search in its 2016 matches against Lee Sedol, [Deep Blue](https://aiwiki.ai/wiki/deep_blue) used 480 custom chips in 1997, and Libratus used 100 CPUs in 2017 [12].

Perfect-information programs sit slightly to the side of this story. [AlphaZero](https://aiwiki.ai/wiki/alphazero) and its predecessors combine [Monte Carlo tree search](https://aiwiki.ai/wiki/monte_carlo_tree_search) with learned evaluation in perfect-information games with alternating turns, where a deterministic policy always does at least as well as the best randomized one [20]. Equilibrium language returns in messier multi-agent settings: [AlphaStar](https://aiwiki.ai/wiki/alphastar) trained a league of competing agents and, per DeepMind's description, the final agent "consists of the components of the Nash distribution of the league", a mixture over discovered strategies rather than a single policy [26]. Michael Littman's 1994 paper on Markov games as a framework for multi-agent [reinforcement learning](https://aiwiki.ai/wiki/reinforcement_learning) introduced minimax-Q for environments shared by two agents with diametrically opposed goals, and noted that unlike a Markov decision process such a game need not have a deterministic optimal policy: the optimal stationary policy is sometimes probabilistic, rock-paper-scissors being the classic case [20]. Policy-Space Response Oracles, published by Marc Lanctot and colleagues in 2017, generalized fictitious play, iterated best response and the double oracle method into a single [multi-agent](https://aiwiki.ai/wiki/multi_agent_system) framework built around approximate best responses to mixtures of opponent policies [19].

### Generative adversarial networks

The [generative adversarial network](https://aiwiki.ai/wiki/generative_adversarial_network) is a game by construction. Ian Goodfellow and coauthors framed it in 2014 as "a minimax two-player game" between a generator and a discriminator, and proved that in the space of arbitrary functions there is a unique solution, with the generator recovering the data distribution and the discriminator outputting 1/2 everywhere [13].

Real training does not take place in the space of arbitrary functions, and the mismatch is where the concept bites. Tim Salimans, Goodfellow and coauthors put it plainly in 2016: "training GANs requires finding a Nash equilibrium of a non-convex game with continuous, high-dimensional parameters", while "GANs are typically trained using gradient descent techniques that are designed to find a low value of a cost function, rather than to find the Nash equilibrium of a game", and when used that way "these algorithms may fail to converge" [14]. Their example is two lines long: if one player minimizes xy over x and the other minimizes -xy over y, simultaneous gradient descent "enters a stable orbit, rather than converging to x = y = 0, the desired equilibrium point" [14]. That orbiting is the gap between descending a [loss function](https://aiwiki.ai/wiki/loss_function) and reaching an equilibrium of a game, and it is why the paper presents techniques such as feature matching and minibatch discrimination as "heuristically motivated to encourage convergence" rather than as guarantees [14]. Much of the practical instability associated with the GAN [minimax loss](https://aiwiki.ai/wiki/minimax_loss), including oscillation and failure to [converge](https://aiwiki.ai/wiki/convergence), sits in that gap [14]. Sanjeev Arora and coauthors later showed that an approximate pure equilibrium does exist for a special class of generators with natural training objectives when generator capacity and training set size are moderate, and proposed a mixture-based training protocol, MIX+GAN, on that basis [15].

### Language model alignment

Since 2023 the concept has been imported into post-training for [large language models](https://aiwiki.ai/wiki/large_language_model). Standard [RLHF](https://aiwiki.ai/wiki/rlhf) fits a point-wise reward model, typically Bradley-Terry, and then maximizes it, which cannot express intransitive or cyclic preference relations [22][23]. Nash Learning from Human Feedback, introduced by Remi Munos and colleagues in December 2023, instead learns a preference model directly and seeks "a policy that consistently generates responses preferred over those generated by any competing policy, thus defining the Nash equilibrium of this preference model", using a mirror-descent algorithm they call Nash-MD [21].

Two follow-ups made the approach competitive on benchmarks. Self-Play Preference Optimization treats alignment as a constant-sum two-player game and approximates its Nash equilibrium through iterative self-play; applied to Mistral-7B-Instruct-v0.2 with a 0.4B preference model and 60,000 prompts, it reached a 28.53% length-controlled win rate against GPT-4-Turbo on AlpacaEval 2.0, rising to 38.77% from Llama-3-8B-Instruct [22]. Direct Nash Optimization, published by Corby Rosset and coauthors in April 2024, reported a 33% win rate against GPT-4-Turbo on the same benchmark from a 7B Orca-2.5 model, up from 7% for the baseline [23]. Work since then has concentrated on convergence behavior under realistic policy parametrization: a May 2025 paper analyzes a simple self-play policy gradient update, finds a stability limitation in its dynamics, and embeds the update in a proximal point framework, proving high-probability last-iterate convergence for the resulting algorithm [28].

A separate 2025-2026 line asks whether language models actually play equilibria. A replication of 121 dyadic games across four classical game types found that Llama reproduced human cooperation patterns closely while Qwen aligned with Nash equilibrium predictions, so "does the model play the equilibrium" turns out to be model-specific rather than a property of LLMs in general [24]. In simulated Cournot markets, a January 2026 study reported that language-model agents sustained tacit collusion and drove prices as much as 200% above Nash equilibrium levels, and that forcing a few dominant agents to play best responses restored competitive pricing [25].

### Mechanism design

Mechanism design inverts the problem: rather than predicting equilibrium behavior in a given game, it designs the game so that equilibrium behavior produces a desired outcome [5]. The textbook instance is Vickrey's second-price auction, in which the item goes to the highest bidder but the winner pays the second-highest bid. That rule makes reporting one's true valuation a dominant strategy regardless of what anyone else does, and a dominant strategy solution is automatically a Nash equilibrium [5]. When dominant strategies are unavailable, designers fall back on Bayesian Nash equilibrium as the behavioral assumption [5]. Leonid Hurwicz, Eric Maskin and Roger Myerson received the 2007 economics Nobel "for having laid the foundations of mechanism design theory" [27]. The same logic runs through ad auctions, resource allocation among software agents, and protocol design in which the price of anarchy is used directly as the design objective [5].

## Limitations

The critique that comes from computer science is computational, and it targets the concept as a behavioral prediction rather than as mathematics. Daskalakis, Goldberg and Papadimitriou argue that "an equilibrium concept should be efficiently computable if it is to be taken seriously as a prediction of what a group of agents will do", because if computing a particular kind of equilibrium "is an intractable problem, of the kind that take lifetimes of the universe to solve on the world's fastest computers, it is ludicrous to expect that it can be arrived at in real life" [4]. Their PPAD-completeness result is evidence that Nash equilibrium fails that test in general.

Multiplicity is the second problem. A game can have many equilibria with very different payoffs, and nothing in the definition says which one players should coordinate on. This is the gap the price of anarchy measures at the aggregate level and the equilibrium selection problem addresses at the individual level [5][12].

Third, the guarantees that make equilibrium attractive are narrow. Outside balanced two-player zero-sum games, playing an equilibrium strategy does not guarantee a good outcome, and the algorithms that converge to equilibria in that setting, CFR among them, carry no such guarantee elsewhere [12][18]. Pluribus and CICERO both responded by treating equilibrium computation as a useful subroutine rather than as the specification of correct play [12][17].

## See also

- [Game theory](https://aiwiki.ai/wiki/game_theory)
- [Generative adversarial network](https://aiwiki.ai/wiki/generative_adversarial_network)
- [Multi-agent system](https://aiwiki.ai/wiki/multi_agent_system)
- [Pluribus](https://aiwiki.ai/wiki/pluribus)
- [Shapley value](https://aiwiki.ai/wiki/shapley_value)
- [Reinforcement learning](https://aiwiki.ai/wiki/reinforcement_learning)

## References

1. Nash, John. "Non-Cooperative Games." Annals of Mathematics, Second Series, Volume 54, Issue 2 (September 1951), pages 286-295. Received October 11, 1950. https://www.cs.upc.edu/~ia/nash51.pdf
2. Nash, John F., Jr. "Equilibrium Points in N-Person Games." Proceedings of the National Academy of Sciences, Volume 36, Issue 1 (January 1950), pages 48-49. https://pmc.ncbi.nlm.nih.gov/articles/PMC1063129/
3. The Royal Swedish Academy of Sciences. "The Prize in Economic Sciences 1994" (press release, 11 October 1994). https://www.nobelprize.org/prizes/economic-sciences/1994/press-release/
4. Daskalakis, Constantinos; Goldberg, Paul W.; Papadimitriou, Christos H. "The Complexity of Computing a Nash Equilibrium." Communications of the ACM, Volume 52, Issue 2 (February 2009), pages 89-97. https://people.csail.mit.edu/costis/simplified.pdf
5. Nisan, Noam; Roughgarden, Tim; Tardos, Eva; Vazirani, Vijay V. (editors). Algorithmic Game Theory. Cambridge University Press, 2007. https://www.cs.cmu.edu/~sandholm/cs15-892F13/algorithmic-game-theory.pdf
6. Koutsoupias, Elias; Papadimitriou, Christos. "Worst-case Equilibria." STACS 99, Lecture Notes in Computer Science volume 1563, pages 404-413, 1999. http://cgi.di.uoa.gr/~elias/publications/paper-kp99.pdf
7. Rubinstein, Aviad. "Settling the complexity of computing approximate two-player Nash equilibria." arXiv:1606.04550, 14 June 2016. https://arxiv.org/abs/1606.04550
8. Zinkevich, Martin; Johanson, Michael; Bowling, Michael; Piccione, Carmelo. "Regret Minimization in Games with Incomplete Information." Advances in Neural Information Processing Systems 20 (NIPS 2007). https://papers.nips.cc/paper_files/paper/2007/hash/08d98638c6fcd194a4b1e6992063e944-Abstract.html
9. Bowling, Michael; Burch, Neil; Johanson, Michael; Tammelin, Oskari. "Heads-up limit hold'em poker is solved." Science, Volume 347, Issue 6218 (9 January 2015), pages 145-149. https://pubmed.ncbi.nlm.nih.gov/25574016/
10. Brown, Noam; Sandholm, Tuomas. "Superhuman AI for heads-up no-limit poker: Libratus beats top professionals." Science, Volume 359, Issue 6374 (26 January 2018), pages 418-424. https://pubmed.ncbi.nlm.nih.gov/29249696/
11. Brown, Noam; Sandholm, Tuomas. "Libratus: The Superhuman AI for No-Limit Poker" (demonstration paper), IJCAI-17. https://noambrown.com/papers/17-IJCAI-Libratus.pdf
12. Brown, Noam; Sandholm, Tuomas. "Superhuman AI for multiplayer poker." Science, Volume 365, Issue 6456 (30 August 2019), pages 885-890. https://noambrown.com/papers/19-Science-Superhuman.pdf
13. Goodfellow, Ian J.; Pouget-Abadie, Jean; Mirza, Mehdi; Xu, Bing; Warde-Farley, David; Ozair, Sherjil; Courville, Aaron; Bengio, Yoshua. "Generative Adversarial Networks." arXiv:1406.2661, 10 June 2014. https://arxiv.org/abs/1406.2661
14. Salimans, Tim; Goodfellow, Ian; Zaremba, Wojciech; Cheung, Vicki; Radford, Alec; Chen, Xi. "Improved Techniques for Training GANs." arXiv:1606.03498, 10 June 2016. https://arxiv.org/abs/1606.03498
15. Arora, Sanjeev; Ge, Rong; Liang, Yingyu; Ma, Tengyu; Zhang, Yi. "Generalization and Equilibrium in Generative Adversarial Nets (GANs)." arXiv:1703.00573, ICML 2017. https://arxiv.org/abs/1703.00573
16. Perolat, Julien; et al. "Mastering the Game of Stratego with Model-Free Multiagent Reinforcement Learning." arXiv:2206.15378, 30 June 2022. https://arxiv.org/abs/2206.15378
17. Meta Fundamental AI Research Diplomacy Team (FAIR). "Human-level play in the game of Diplomacy by combining language models with strategic reasoning." Science, Volume 378, Issue 6624 (9 December 2022), pages 1067-1074. https://pubmed.ncbi.nlm.nih.gov/36413172/
18. Meta Fundamental AI Research Diplomacy Team (FAIR). "Human-level play in the game of Diplomacy" (technical report with supplementary materials). https://noambrown.com/papers/22-Science-Diplomacy-TR.pdf
19. Lanctot, Marc; Zambaldi, Vinicius; Gruslys, Audrunas; Lazaridou, Angeliki; Tuyls, Karl; Perolat, Julien; Silver, David; Graepel, Thore. "A Unified Game-Theoretic Approach to Multiagent Reinforcement Learning." arXiv:1711.00832, NIPS 2017. https://arxiv.org/abs/1711.00832
20. Littman, Michael L. "Markov games as a framework for multi-agent reinforcement learning." Proceedings of the Eleventh International Conference on Machine Learning, 1994. https://www2.cs.duke.edu/courses/spring07/cps296.3/littman94markov.pdf
21. Munos, Remi; et al. "Nash Learning from Human Feedback." arXiv:2312.00886, 1 December 2023. https://arxiv.org/abs/2312.00886
22. Wu, Yue; Sun, Zhiqing; Yuan, Huizhuo; Ji, Kaixuan; Yang, Yiming; Gu, Quanquan. "Self-Play Preference Optimization for Language Model Alignment." arXiv:2405.00675, 1 May 2024. https://arxiv.org/abs/2405.00675
23. Rosset, Corby; Cheng, Ching-An; Mitra, Arindam; Santacroce, Michael; Awadallah, Ahmed; Xie, Tengyang. "Direct Nash Optimization: Teaching Language Models to Self-Improve with General Preferences." arXiv:2404.03715, 4 April 2024. https://arxiv.org/abs/2404.03715
24. Cera Palatsi, Andrea; Martin-Gutierrez, Samuel; Cardenal, Ana S.; Pellert, Max. "Large language models replicate and predict human cooperation across experiments in game theory." arXiv:2511.04500, 6 November 2025. https://arxiv.org/abs/2511.04500
25. Deshpande, Sanyukta; Jacobson, Sheldon H. "Strategic AI in Cournot Markets." arXiv:2601.17263, 24 January 2026. https://arxiv.org/abs/2601.17263
26. DeepMind. "AlphaStar: Mastering the real-time strategy game StarCraft II." https://deepmind.google/discover/blog/alphastar-mastering-the-real-time-strategy-game-starcraft-ii/
27. The Royal Swedish Academy of Sciences. "The Prize in Economic Sciences 2007." https://www.nobelprize.org/prizes/economic-sciences/2007/summary/
28. Tiapkin, Daniil; Calandriello, Daniele; Belomestny, Denis; Moulines, Eric; Naumov, Alexey; Rasul, Kashif; Valko, Michal; Menard, Pierre. "Proximal Point Nash Learning from Human Feedback." arXiv:2505.19731, 26 May 2025. https://arxiv.org/abs/2505.19731