Researchers from MIT, Carnegie Mellon, NYU and Stanford have built an AI system called Ataraxos that dominates the hidden‑information board game Stratego, beating top human players by a large margin. Stratego is a two‑player imperfect‑information game where piece identities remain secret until they clash, yielding more than $10^{66}$ possible setups—far larger than chess—making it a stringent benchmark for strategic AI.\ \ Ataraxos follows a two‑pronged approach. First, it learns a "blueprint strategy" through self‑play reinforcement learning, using specially engineered efficient algorithms that reduce the required training data to less than one‑hundredth of DeepNash’s and the number of self‑play games to under one‑thirtieth, dramatically cutting computational cost. Second, during actual play the system refines its moves on the fly via decision‑time planning. It employs a generative model to assign probabilities to the opponent’s hidden pieces, evaluates a set of candidate actions under these inferred states, and selects the move with the highest expected payoff.\ \ This generative, decision‑time planning component is the missing piece that lifts Ataraxos to superhuman performance. In the world championship it defeated the top‑ranked player with a 15‑1‑4 record and amassed a 39‑2 record against elite human competitors, outperforming DeepMind’s DeepNash which still struggled in this domain.\ \ The team also adapted Ataraxos to other imperfect‑information games—Barrage Stratego (a faster variant), the cooperative card game Hanabi, and the three‑player game Dou dizhu—achieving superhuman results in each, demonstrating the method’s generality. Future work aims to embed interpretability tools so humans can audit the AI’s reasoning, a crucial step for applying such systems to real‑world problems like business negotiations or cybersecurity.\ \ Review: Ataraxos showcases how efficient self‑play learning combined with on‑the‑fly probabilistic planning can overcome the combinatorial explosion of hidden‑information games, offering a promising blueprint for AI‑assisted decision making in complex, real‑world scenarios.