Meta recently introduced CICERO, an AI that achieves human-level performance in the online version of Diplomacy, a strategy game in which it communicated and negotiated in natural language with human players without them knowing they were dealing with a bot. Cicero ranked in the top 10% of players who played more than one game and even won an eight-game tournament involving 21 players.

AIs like DeepMind’s AlphaZero, based on reinforcement learning, have demonstrated skills that equal or surpass those of humans for games of chess, shogi or go. Cicero integrates a natural language processing model with planning and reinforcement learning algorithms and is the first AI system to achieve human-level performance in the strategy game Diplomacy.

In this game, seven major powers in 1901: France, England, Germany, Austria-Hungary, Russia, Turkey, Italy, each seek to dominate Europe. Seven players, each controlling a country, compete to own at least half of the supply centers on the game board. To succeed in this challenge, players must cooperate, negotiate, and form alliances. Each round is preceded by two-on-two negotiations, where players form alliances and team up against their opponents. When it’s time to play and move their pieces, it’s up to them whether or not to keep their commitments.

At each stage of the game, Cicero models the likely movements of the other players based on the current state of the board and previous conversations.

For an AI to play this game, it needs to do more than just use natural language; it needs to communicate persuasively. According to an article in Science, Ciceron was trained on a data set of 125,261 games played online at webDiplomacy.net. Of these, 40,408 games contained dialogues, with a total of 12,901,662 messages exchanged between players.

Cicero was registered, but not revealed as a bot, in 40 games running between August 19 and October 13, 2022: it doubled the average score of players. During those 72 hours of play involving the sending of 5,277 messages, the other players did not suspect that he was not human. According to Meta AI, they even preferred to interact with CICERO rather than with other human players.

While he performed well, Cicero still has room for improvement, sending messages that contained errors or contradicted his own plans. His dialogue focused only on the actions of the current turn’s players, he did not model how his dialogue might affect his relationship with other players longer term in the game.

Meta AI states in its blog:

“CICERO marks the beginning of a new era for AI that can collaborate with people in the game using strategic reasoning and natural language processing, and learnings from technology like this could one day lead to intelligent assistants that can collaborate with people.”

Translated from Meta AI présente CICERO, un agent d’IA capable de négocier et coopérer avec les humains