Researchers built an AI advisor for the bluffing card game Coup, and the way you train it to think about winning matters more than the algorithm itself.
The system, called CoupVisor, is a decision-support tool for Coup, a game where players bluff about which hidden role they hold. It handles two jobs: recommending what to do on your turn, and telling you when to challenge an opponent's claim. All of it runs off one shared log of game events, which also powers manual play, replaying recorded games, simulation, and belief tracking. To judge whether a claim is truthful, CoupVisor weighs how plausible the claimed role is against how many cards the claimant still has left - a fix that corrects an earlier flaw where the game's very first claim got flagged as suspicious for no real reason. The researchers then pitted a rule-following advisor against several learned and heuristic players across many simulated games and opponent styles.
The headline result is about reward design, not cleverness. A policy trained to maximize the chance of ultimately winning beat every baseline, while one trained to chase short-term gains lagged behind. That's a small but concrete data point in a debate that shows up everywhere from trading bots to negotiation agents: should you optimize for the next good move, or the eventual outcome.
Coup is a strange but useful test bed precisely because it rewards convincing lies as much as sharp math - which puts it in the same lineage as poker bots and Diplomacy-playing AI, minus the geopolitics.