The Hidden-Information Game Stratego Finally Falls to AI


Meta description: Ataraxos AI beat Pim Niemeijer 15-1 with four draws after training on 16 GPUs, solving Stratego's long-resisted hidden-information challenge.

Tags: Ataraxos, Stratego AI, imperfect information, game AI, Carnegie Mellon

The Ataraxos AI Stratego result closes a long-standing gap in machine game play. In 1997, Deep Blue defeated Garry Kasparov in chess; in 2016, AlphaGo beat Lee Sedol in Go; and poker bots had already been beating professionals for years. Stratego remained the exception. Researchers from Carnegie Mellon, MIT, New York University, and Stanford University now have: their system, Ataraxos, beat Pim Niemeijer, arguably the best Stratego player of all time, 15 games to one, with four draws.

Background: A hidden-army game that resisted machines

In Stratego, each player commands 40 pieces representing military ranks, from a marshal down to a spy, plus bombs and a flag. The goal is to capture the opponent's flag. The difficulty is not calculation alone. Each player can see where enemy pieces stand, but not what those pieces are. Their identities are revealed only when two pieces collide in battle. The weaker piece is removed, and the winner's identity becomes visible. In Stratego, the hidden identity of pieces means an AI cannot simply search for the strongest move; it must reason about what an opponent may be hiding. That structure makes Stratego an imperfect-information game, similar to poker, where success depends on incomplete knowledge, belief about the opponent, and risk assessment.

Development: Training Ataraxos with a small hardware footprint

The team trained Ataraxos using just 16 GPUs and a few thousand dollars. That small training footprint is notable because even DeepMind, with its exceptional budget, had not produced a machine that reliably beat the best human players. The new result suggests the bottleneck in Stratego may have been more than raw compute: the method used to handle hidden information. For readers following AI milestones, that means a game that had resisted machines now has a concrete machine opponent.

Latest progress: Ataraxos defeats Pim Niemeijer

Ataraxos faced Pim Niemeijer, described as arguably the best Stratego player of all time, and won 15 games to one, with four draws. The match is a 15-1 result in decisive games. It gives the AI a clear victory over the human benchmark, rather than a narrow or inconclusive test. Eugene Vinitsky of New York University, a coauthor of the work, said: “There's something super distinctive about Stratego, which is that it is a massive amount of hidden information that unfolds over a very long time scale.”

What the result changes

The milestone places Stratego alongside chess, Go, and poker in the timeline of games where machines have matched or surpassed elite human play. But the comparison with poker is especially important: both games force players to act without full knowledge of the board or cards. Ataraxos shows that a relatively modest training setup can produce a Stratego system capable of defeating a top human player. The result also leaves a concrete question for the field: whether the same approach can be generalized to other hidden-information problems where identities, intentions, and future consequences remain partially concealed. At present, the milestone stands as a completed match against a named human benchmark, not yet a broad public tournament record.