In recent research conducted by MIT, language models (LMs) in the field of artificial intelligence (AI) are shown to be prone to producing incorrect statements as their size increases. This raises concerns about their reliability, even though they have the capability to handle tasks such as question answering and fact-checking. The challenge lies in extracting the most accurate information from LMs that generate contradictory responses.
To address this issue, MIT researchers have developed a signaling game called the CONSENSUS GAME. This game aims to bridge the gap between generative and discriminative LM decoding processes. The game involves a DISCRIMINATOR agent that conveys whether an answer is correct or wrong to a GENERATOR agent. The challenge is to find a combined policy where both agents agree on the correctness of the generated responses. This requires solving a multi-step game with a string-valued action space.
In their study, the researchers introduced a game-theoretic method known as EQUILIBRIUM-RANKING for decoding LMs. They applied this method to six question-answering benchmarks and found that it outperformed existing generative, discriminative, and mixed decoding techniques. The use of game theory helps formalize and enhance coherence in LMs, resulting in improved accuracy in factual tasks.
If you want to learn more about this research, you can read the full paper here. Credit for this research goes to the MIT researchers involved in the project.
Don’t forget to join our ML SubReddit, Facebook community, Discord Channel, and subscribe to our Email Newsletter to stay updated on the latest AI research news and projects.
If you enjoy our work, you’ll love our newsletter. You can subscribe to it here.
We also have an AI Channel on WhatsApp. You can join it here.