Cicero plays Diplomacy
From 19 August to 13 October 2022 Meta FAIR's agent Cicero played 40 games of Diplomacy anonymously on webDiplomacy.net against 82 people, negotiating with them in plain language. Its mean score was 25.8% against its opponents' 12.4%. The paper was published in Science on 22 November 2022.
Why it matters
Diplomacy is won by agreement: alliances and betrayals are made in private conversation. Cicero joined a language model to strategic planning and showed that an agent can negotiate with people at the level of a strong player, not only calculate moves.
A blitz league with five-minute negotiation turns: over 72 hours of play Cicero sent 5,277 messages, ranked in the top 10% of those who played more than one game, second of the 19 who played five or more, and first in an eight-game tournament of 21 participants. The language model is a 2.7-billion-parameter transformer. Opponents did not know they were playing a machine; one player suspected a bot after a game, but Cicero was not exposed. Meta released the code.