Ataraxos AI Beats Stratego Champion 15-1 on Small Budget

TL;DR: An AI called Ataraxos beat Stratego champion Pim Niemeijer 15-1 with four draws in a 20-game series reported September 30, 2026 in Nature.

A university team says it has cleared a benchmark that had resisted even large industrial programs. According to the Nature paper, Ataraxos defeated Pim Niemeijer, described as the most decorated Stratego player of all time, by 15 wins, one loss and four draws. According to Ars Technica, the work came from researchers at Carnegie Mellon University, the Massachusetts Institute of Technology, New York University and Stanford University.

The result was presented as the first superhuman result in the game's history to the authors' knowledge. The Nature abstract states that Ataraxos consumed orders of magnitude less compute and data than previous efforts. The authors estimated training cost at less than US$8,000 at 2025 prices, according to a summary in On The Wire that reviewed the Nature paper.

Who built Ataraxos and when it was reported

The Nature article titled "Scalable decision-making for games of imperfect information" was published September 30, 2026. The listed authors are Samuel Sokota and J. Zico Kolter of Carnegie Mellon, Eugene Vinitsky of New York University, Hengyuan Hu of Stanford University, and Zhiyuan Fan and Gabriele Farina of MIT.

According to MIT Schwarzman College of Computing, Sokota was the lead author and Farina was senior author. The MIT report dates the findings to September 30, 2026. Ars Technica reported the result October 1, 2026, in an article by Jacek Krywko.

The institutions framed the system as a general approach to hidden-information decision-making, not only a Stratego program. Nature's abstract says the same techniques produced a superhuman AI for Barrage Stratego and state-of-the-art systems for Hanabi and dou dizhu.

Why Stratego resisted AI

Stratego is a two-player board wargame resembling military chess. Each side arranges 40 pieces in secret, then alternates moves defined by the displacement of a single piece. When a moving piece lands on an occupied square, a battle occurs, both piece types are revealed, and at least one piece is captured based on type and attack-defence roles. Capture of the opponent's Flag wins the game.

The difficulty is the scale of concealed setup. Nature puts the number at more than 10^33 possible piece configurations. Ars Technica describes it as more than a decillion possible setups, with 40 pieces that could be in any order. MIT's account describes 40 pieces per side and identities that remain secret until pieces collide, with the lower-ranking piece eliminated.

That scale separates Stratego from poker. In Texas Hold'em, each player holds only two hidden cards, leaving 1,326 possible hands, a figure cited in both Nature and Ars Technica. Farina told Ars Technica that a machine could weigh all of those possibilities, while Stratego's combinations could not be enumerated the same way.

Length compounds the problem. Farina said a chess game usually lasts 40 moves, while a Stratego game can easily last 2,000 moves, according to Ars Technica. Vinitsky said Stratego involves a massive amount of hidden information that unfolds over a very long time scale.

Bluffing adds another layer. Players sometimes advance a weak piece as if it were a marshal to deter attack. Ars Technica explains the tradeoff described by the team: frequent bluffs lose value because opponents expect them, while never bluffing becomes predictable. Sokota told MIT that the value of a bluff shifts with opponent expectations, unlike chess where the best move remains best regardless of frequency.

Earlier systems did not solve the top-human test. Nature describes industrial research spanning multiple years, involving dozens of researchers and computation costing millions of dollars, without reaching top-human level. That history made Stratego, in the paper's words, potentially the only classical game where well-resourced efforts had failed to produce superhuman performance.

The 20-game match against Niemeijer

The central evaluation was a 20-game online series against Pim Niemeijer played over three weeks. Niemeijer holds four world championships and more than 600 weeks as the world's top-ranked player, according to Ars Technica.

Ataraxos recorded 15 wins, one loss and four draws. Counting draws as half wins gives an effective win rate of 85%, according to On The Wire's calculation from the paper. The Nature paper calls the margin of victory without precedent at the highest level of play.

Match conditions limited adaptation by the machine. Niemeijer was told Ataraxos would not adapt to his play, according to On The Wire. Ars Technica reports that he knew the AI would not adapt, giving him time to search for weaknesses. He managed one win.

Payment terms rewarded human scoring. On The Wire reports Niemeijer was paid US$1,000 to participate plus US$100 per win and US$50 per draw. Ars Technica reports he earned $100 for each win.

The authors included a statistical caution. They wrote that the games were far from independently and identically distributed because Niemeijer could adapt across the series. Only under the assumption that they had been independent does a one-sided binomial test yield a P value below 2.6 x 10^-4, according to On The Wire.

The team attributed the single defeat partly to chance in deployment. Because strong Stratego requires randomizing piece arrangement, Farina said even a perfect strategy will sometimes lose, according to Ars Technica. Vinitsky acknowledged the luck ran both ways, saying the team sometimes got lucky.

Context for past machine milestones helps explain interest. Ars Technica notes Deep Blue defeated Garry Kasparov at chess in 1997, AlphaGo beat Lee Sedol at Go in 2016, and poker programs have beaten professionals for years. Stratego, by contrast, held out.

How Ataraxos learned and reasoned

Ataraxos combined self-play reinforcement learning with test-time search. The paper describes a policy-value network trained by self-play from scratch, a belief network that estimates opponent hidden pieces, and a search step before acting.

Self-play volume was large but far below prior Stratego work. Ars Technica reports 163 million self-play games. On The Wire reports about 160 million self-play games for Ataraxos against about 5.5 billion for DeepNash. MIT quotes Farina saying the system reached strictly higher playing strength than DeepNash while using less than one hundredth of the training examples and less than one thirtieth of the self-play games.

In self-play, moves that led to wins were reinforced and played more often later, while moves that led to losses were played less. Ars Technica describes that as the same simple training idea used before. The difference, it reports, was how much Ataraxos adjusted after each game, because hidden information can send self-play learning around in circles. The team made large changes early in training and small careful changes later.

The larger change was reasoning before each move. Programs such as AlphaGo refine general strategy with search just before acting. Ars Technica reports DeepMind could not make that work in Stratego because the search space was too large. Farina said his group figured out how to do it.

The solution used a second neural network. The belief model guesses opponent piece identities from how those pieces have moved. Instead of iterating through every possible arrangement, Ataraxos samples plausible arrangements, simulates candidate moves in each, and selects based on outcomes. The Nature-linked summary describes sampling possible hidden arrangements, playing out candidate moves, then performing one additional update step to the policy for that decision only.

MIT describes the same two stages as a blueprint strategy plus in-game reasoning. The model first learns a blueprint strategy through self-play, then uses that blueprint as a starting point for setup and for thinking about next moves each round.

Low compute and calm play

Training hardware was modest by frontier-AI standards. The reinforcement-learning run used 16 NVIDIA H100 graphics processing units for one week, with four additional H100s for four days to train the belief network, according to On The Wire. The authors priced that at less than US$8,000 at 2025 prices. Ars Technica summarizes the requirement as 16 GPUs and a few thousand dollars.

The name reflects intended behavior. Ataraxos comes from an ancient Greek word for a state of calm, meaning somebody calm and unbothered, according to Farina in Ars Technica. He suggested the structure and lack of human emotion prevent impulsive reactions even in positions where a human would be losing composure.

Observed play was methodical rather than desperate. While a human far behind might attempt large gambles, Ataraxos worked back slowly and methodically, according to Ars Technica. Vinitsky said the team watched the program bluff its way back from about a two percent victory probability very casually.

Another trait was protecting secrets. When Ataraxos estimated its opponent had no reason to suspect a weak spot, it left that spot alone, even if full-board visibility would make it look vulnerable. Farina said it is very hard for humans who know a secret to decide as if they did not know it, while for machines it is easy. That property helped the system make moves a human would only make while bluffing and follow them up more effectively.

Comparison with DeepNash

DeepNash was DeepMind's Stratego system introduced in 2022 and described in Science. The Nature paper says it won 42 of the 50 counted games on the Gravon site in April 2022 but did not achieve the top ranking, according to On The Wire. The paper argues Gravon opponents were far from top-human level.

Tournament data was also cited. The paper says DeepNash recorded 19 wins and nine losses at the 2023 World Championship and lost to most of the highest-ranked players it faced, including Niemeijer. DeepMind's own 2022 paper said DeepNash achieved a yearly 2022 and all-time top-three rank on Gravon while competing with human expert players, according to On The Wire.

There was no direct Ataraxos-DeepNash match. The authors wrote that DeepMind told them that would not be possible because the DeepNash code was no longer functional. Their cost comparison was an estimate based on published hardware figures and recollection of a DeepNash author, placing an equivalent DeepNash run at roughly US$3 million to US$4.5 million under 2025 pricing.

Wider tests beyond classic Stratego

The researchers applied the same methods to three other games. In Barrage Stratego, a variant with fewer pieces, the AI won four 50-game series against three of the four top-ranked players, which the authors called the first superhuman result for that game to their knowledge, according to On The Wire.

Hanabi and dou dizhu tested cooperation and team play. In Hanabi, the comparison was against earlier AI systems rather than human champions, with the paper reporting a new state of the art. In dou dizhu, it reported wins over the PerfectDou and DouZero bots, according to On The Wire. MIT says the system outperformed top human players in other strategic games with different rules and designs, showing generalization across uses.

A public demonstration added 40 more games. At the 2025 Stratego World Championship held August 1-3, attendees played Ataraxos 40 times, with the paper recording 38 wins, two losses and no draws. On The Wire notes MIT's press release gave the championship record as 39-2 and says it used the paper's figures. Ars Technica reports the AI won 38 of 40 games and says it changed how people play, skewing the metagame.

Peer review, code and funding were disclosed with the paper. Nature published the peer-review file. One referee first called the experiments a weakness and said evaluation scale was relatively limited, while also writing that the 15-1 margin left little doubt about the superhuman claim. After revision adding bot comparisons and three further games, that referee said the responses adequately addressed the weaknesses. Another referee called the work a major empirical milestone.

Code is posted on GitHub under AtaraxosAI in three repositories covering Stratego, Hanabi and dou dizhu, each under the MIT licence, according to On The Wire. Records of the 20 games are on the project site. The paper states the authors declare no competing interests. Funding included U.S. Office of Naval Research awards, National Science Foundation support, a U.S. Department of Transportation grant and Schmidt Sciences AI2050 fellowships.

The authors argue the lesson extends beyond board games. Nature says hidden information characterizes financial markets, military conflict and negotiations. MIT quotes Farina saying real imperfect-information tasks rarely allow enumerating every possibility because there are too many. The paper concludes that practical AI is within reach for many strategic decision problems where fast, accurate simulators can be built.