New AI system Ataraxos beats top human players at Stratego

October 1, 2026, 06:01Source: MIT Schwarzman College of Computing

A research team from MIT, Carnegie Mellon University, New York University and Stanford University has built an AI system called Ataraxos that defeated top-ranked human players of the board wargame Stratego by a large margin. According to MIT News, no AI system had achieved this before, and Ataraxos did it at a fraction of the training cost of earlier efforts.

Key points

Ataraxos beat the world's strongest Stratego player 15-1-4 and went 39-2 against top humans at the world championship.
According to the researchers, it beats DeepMind's DeepNash using less than one hundredth of the training examples and less than one thirtieth of the self-play games.
It combines self-play reinforcement learning with decision-time planning based on a generative model of hidden pieces.
The system also reached superhuman performance in Barrage Stratego, Hanabi and Dou dizhu.
The team plans to add interpretability so decisions can be audited before any real-world use.

A game built on hidden information

Stratego is a two-player game of imperfect information. Players arrange 40 pieces on their side of a board and move them across it to capture the opponent's flag. The identity of every piece stays secret until two pieces collide, and then the lower-ranking one is eliminated.

MIT says the possible piece configurations number more than 10 to the 66th power, an exponentially greater number than in chess. That makes the game very hard for an AI system, and it is often used as a benchmark for the strategic thinking of powerful models.

Hidden information changes how decisions work. Sam Sokota, lead author of the paper and a graduate student at Carnegie Mellon, put it this way: "The more you bluff, the more your opponent expects it, and the less each bluff is worth." He contrasted this with chess, "where the best move is still the best move no matter how often you've played it."

According to the article, earlier efforts such as Google's DeepMind relied on sophisticated operations that were computationally demanding and costly. Even with millions of dollars in training costs, those models were not strong enough to beat top human Stratego players.

How Ataraxos works

The name comes from a Greek word describing one who is unbothered or free from anxiety. The system combines two main ingredients.

Self-play reinforcement learning: the model plays against itself many times to learn a strong "blueprint strategy". The team designed especially efficient algorithms so that Ataraxos learns faster than prior methods and does not get stuck trying to predict every possible move.
Decision-time planning: during a game, Ataraxos starts from the blueprint strategy, then refines its choices on the fly. A generative model uses probabilities to estimate the likely identities of the opponent's hidden pieces, and the system evaluates future choices before selecting a move.

Gabriele Farina, an assistant professor in EECS at MIT and senior author of the paper, said the approach lets the system "really zoom in on the specific board and opponent we are facing." MIT describes the use of this generative model for decision-time planning as the missing piece that enabled superhuman performance.

Results and efficiency

Ataraxos beat the strongest Stratego player in the world by a record margin of 15-1-4, according to MIT. It also achieved a 39-2 record against top human players at the Stratego world championship.

On efficiency, Farina said: "Our system reaches strictly higher playing strength than DeepNash (DeepMind's system) while using less than one hundredth of the training examples and less than one thirtieth of the self-play games, indicating a massive improvement in efficiency."

The researchers also adapted the system to other imperfect information games: Barrage Stratego, a faster-paced variant with fewer pieces; Hanabi, a cooperative card game with many players; and Dou dizhu, in which two players cooperate against a third. Ataraxos reached superhuman performance in each, which the team presents as evidence that the method generalizes.

Farina also described how the bot plays: "A human might start freaking out if their most valuable piece is exposed, but the bot can be surprisingly composed. It doesn't overcorrect and give away its secrets."

What it could mean beyond games

The researchers say the system could be adapted to real-world problems with hidden information, such as business negotiations, cybersecurity and military maneuvers. Farina noted that in such tasks "you often don't have the luxury of enumerating through all the possibilities."

Adoption is not imminent. The team wants to build interpretability measures into Ataraxos so it can explain its decisions in a way a human can understand. "Humans must have the final say in whether a recommendation is followed, so before adoption can happen, we need a way to audit the model's decisions," Farina said.

The paper, whose authors also include Eugene Vinitsky of NYU, Zico Kolter of Carnegie Mellon, Hengyuan Hu of Stanford and Zhiyuan Fan of MIT, appears in Nature. The research is funded in part by the Office of Naval Research, the New York University Department of Civil and Urban Engineering, the C2SMART Center, the National Science Foundation, and a Schmidt Sciences AI2050 Early Career Fellowship.