The game of Diplomacy posed a significant challenge for AI researchers due to its complex nature. Unlike games with just two opponents, Diplomacy involves seven players with difficult-to-read motives. Winning requires negotiation and forging cooperative arrangements that can be broken at any time. In 2022, Meta’s AI program Cicero achieved “human-level play” after 40 games, placing in the top 10 percent against human participants.
During the project, Jacob from the Meta team realized Cicero relied on a language model for its dialogues. This led them to consider building a better game to enhance the performance of large language models.
Consensual Interactions
In 2023, Jacob collaborated with Yikang Shen, Gabriele Farina, and adviser Jacob Andreas at MIT to develop the consensus game. This game treats conversations between two people as cooperative games, emphasizing understanding between a speaker and a listener. The consensus game aims to align the language model’s generator and discriminator systems.
The game involves the generator receiving a question and providing candidate responses based on whether it should answer correctly or incorrectly. The discriminator evaluates the responses and rewards both players for reaching agreement. Initial beliefs guide their responses, encouraging them to incorporate knowledge from the internet for accuracy.
