Using ideas from game theory to improve the reliability of language models | MIT News

Picture you and a friend playing a game where the aim is to communicate secret messages using cryptic sentences. Your friend must guess the secret message behind your sentences. Sometimes you give clues directly, other times your friend has to ask yes-or-no questions about the clues you’ve given to guess the message. The challenge is to ensure you both understand each other and agree on the secret message.

Researchers at MIT Computer Science and Artificial Intelligence Laboratory (CSAIL) have developed a similar concept known as a “consensus game” to enhance how AI comprehends and generates text. This game involves two parts of an AI system, one generates sentences (giving clues) and the other understands and evaluates those sentences (guessing the secret message).

The team discovered that by treating this interaction as a game, where both AI parts work together under specific rules to agree on the correct message, they could significantly enhance the AI’s ability to provide accurate and coherent answers. Testing this game-like approach on various tasks such as reading comprehension, math problem-solving, and conversations showed improved performance across the board.

Usually, large language models answer questions by generating answers directly or scoring predefined answers, which can lead to conflicting results. The researchers introduced a new “game-theoretic” method to help language models understand and generate text more effectively.

The team’s algorithm consistently improved the performance of language models across different tasks like reading comprehension, commonsense reasoning, and math problem-solving. The researchers’ work was presented at the International Conference on Learning Representations (ICLR) and received a “best paper award” at the NeurIPS R0-FoMo Workshop in December 2023.