MIT's New Stratego AI Beats Elite Human Players on a Laptop-Sized Budget
A game-playing system from MIT can outmaneuver top-ranked Stratego players while using far less compute than rival approaches, pointing to better tools for planning under uncertainty.
Researchers at MIT have built an AI system that defeats the best human Stratego players in the world, a milestone for a game long considered a tough test of machine decision-making. The work, reported by MIT News, also shows the system running more efficiently than competing models, meaning it reaches top-level play without the massive compute budgets such approaches usually demand.
Stratego is not chess. Players cannot see each other's pieces, so every move is a bet made on incomplete information, and a single wrong guess can cost the game. That structure makes it a closer stand-in for real-world planning than perfect-information games, which is why researchers have chased it for years. The MIT team's system reportedly handles that uncertainty well enough to topple ranked human opponents, not just other bots.
The efficiency claim may matter as much as the win rate. Many recent game-playing breakthroughs rely on enormous training runs and search budgets that few organizations can afford to reproduce or deploy. A system that reaches elite play with leaner resource use is easier to copy, test and adapt, which lowers the bar for labs and companies that are not sitting on hyperscale infrastructure.
The researchers point to practical uses beyond the board: military maneuvering and business negotiations, both of which involve hidden intentions, limited information and opponents who react to what you do. Those settings are exactly where standard planning tools tend to struggle, because the right move depends on what the other side knows and believes. A model that performs well in Stratego could inform decision-support systems in those areas, though translating a board game into the messy real world is never direct.
Stratego's mix of deception and deduction has made it a stubborn benchmark. Previous systems have made progress, but beating top human players while keeping costs down is a different bar. If the results hold up under scrutiny, the techniques could feed into research on adversarial planning, where an agent must act without knowing the full state of the world.
What to watch next is whether the team publishes enough detail for others to replicate the results and whether the approach scales past a fixed board. The obvious open question is how much of Stratego's success depends on the game's specific rules, and whether the same efficiency carries over to negotiations or military planning, where the rules are not written down and the stakes are not simulated.