An AI Entered a Game That Required Negotiation

1. Orders Need Allies

In Diplomacy, players move armies and fleets across a map while negotiating with one another between turns. Promises are not binding: each participant submits orders, and their consequences are resolved together. Meta’s researchers set out to build a program that could both choose moves and communicate plans to human players in this setting.

A move planner alone could not carry out the conversations.

2. Combining Two Systems

The team built CICERO by combining a strategic reasoning system with a dialogue model. The planner examined possible moves and other players’ likely actions. The language component used that information to produce messages about proposed cooperation. Researchers sought to keep the messages connected to the program’s intended actions during each turn.

The program needed examples of how people negotiated during games.

3. Learning from Games

Researchers used a dataset of more than 40,000 human games from webDiplomacy.net to help develop the system. Those records included moves and conversations. Training on the language of the game was one part of the work; the program also had to select actions from the current board position while considering what other players might do.

The completed system was tested against people online.

4. Entering the League

Between August and October 2022, CICERO played forty games in an online league using the English-language, short-turn version of Diplomacy. The researchers entered it anonymously. Human participants exchanged messages with the program while deciding their own moves, giving the team a record of both its negotiations and its eventual results.

The evaluation used actual completed games rather than sample dialogue alone.

5. Negotiation Before Orders

During those matches, CICERO could propose coordinated moves and respond to other players’ messages. Its strategic system updated expectations as the board changed. The dialogue model then generated language using those plans. The published research described filtering procedures intended to reduce inconsistent messages, although the system’s performance still had to be assessed from what happened in the games.

The researchers compared its scores with the league’s human results.

6. The Recorded Scores

CICERO obtained more than twice the average human score in the forty-game evaluation. It ranked in the top ten percent of participants who played more than one game. Those figures described the reported league setting, with its own rules and turn times. They were not measurements across every version of the board game.

The team prepared its methods and results for publication.

7. Opening the Research

Meta announced CICERO in November 2022 and released code alongside its research publication. The materials described how the dialogue and strategy components worked together, the data used and the evaluation procedure. Readers could inspect the described system rather than judge it solely from selected messages or an isolated successful match.

The released material kept the result tied to the evaluated competition.

8. Forty Completed Games

Meta reported CICERO’s results from forty online Diplomacy games in its 2022 research. It scored more than twice the average human score and ranked in the top ten percent of participants who played more than one game. Those were measurements from the described competition setting. The team released research material and code; proposed uses beyond the game remained separate from the matches it had actually played.

The reported result came from the forty evaluated games.

Sources

• Meta: CICERO research announcement