Meta’s Cicero reaches human-level play in Diplomacy
Meta AI announced Cicero on 22 November 2022, an agent that paired a language model with strategic planning and reinforcement learning to play the negotiation game Diplomacy. Across 40 anonymous games on webDiplomacy.net it more than doubled the average human score.
Why it mattered Cicero placed in the top 10 percent of players who played more than one game without its opponents realizing it was a machine, extending game-playing systems into open-ended negotiation with people.
Diplomacy is a seven-player game of alliances set in Europe before the First World War. Players have no dice and no hidden units; they have conversation. Each turn is preceded by private negotiation, and the game is decided by who trusts whom and who is lying.
Cicero combined a language model, trained to hold those negotiations, with a planning component that worked out what moves to make and what to ask for. The two had to agree: the system needed to say things consistent with what it intended to do, and to infer intentions from what others said.
Meta reported that between August and October 2022 Cicero played 40 anonymous games on webDiplomacy.net, more than doubling the average score of its human opponents and placing in the top 10 percent of players who played more than one game. Its opponents did not know they were playing a program.