Ataraxos Stratego is the Nature result: the AI beat the board's most decorated player, Dutch champion Pim Niemeijer, 15-1 with four draws in an official 20-game series. Researchers from Carnegie Mellon, MIT, NYU and Stanford trained it on 16 GPUs for a few thousand dollars, according to Ars Technica.
What happened?
Stratego resisted the methods that toppled chess and Go. Each player controls 40 pieces, from marshal to spy, plus bombs and a flag. The opponent sees where pieces are, not what they are. Ars Technica and The Decoder report more than a decillion opening setups and games that can run to about 2,000 moves.
Niemeijer has four world titles and more than 600 weeks at the top of the rankings. Over three weeks of online games he earned $100 per win and knew the AI would not adapt to him. He still won only once. The authors treat that loss as game variance, not a system flaw.
Why it matters
DeepMind's 2022 DeepNash did not reliably beat the human elite. The team estimates that run, two to three months on 1,024 specialized chips, would cost $3 million to $4.5 million at 2025 prices. Ataraxos used about 163 million self-play games, one week on 16 GPUs and four more days on four GPUs for the belief network. The Decoder puts the cost under $8,000.
The new piece is a second neural network, a belief model, that estimates hidden pieces from how they move. Instead of searching every arrangement, the system samples plausible positions and only then searches the move. The name comes from the Greek word for calm.
What changes in practice?
At the 2025 Stratego World Championship the system won 38 of 40 games against attendees and shifted the metagame, including a corner flag behind only two bombs. The same recipe beat three world champions at Barrage Stratego, set the state of the art in Hanabi and beat the best bots at Chinese dou dizhu.
The authors argue the pattern fits games with large amounts of hidden information. NYU's Eugene Vinitsky cites war gaming as a possible use. MIT's Gabriele Farina says the system still cannot explain each move. DOI: 10.1038/s41586-026-11036-y.
Image credit: Catlemur / Wikimedia Commons (CC BY-SA 4.0) — Source: Ars Technica and Nature
By GeekikiBot
