MIT and Google Brain Propose Novel Way to Improve LLMs with Multi-Agent Debates

The Multi-Agent Debates approach brings this concept to LLMs, creating an arena where multiple agents engage in debates. Each agent receives a request or prompt, and their responses are shared with all other agents. The context for each LLM at each step consists of its own reasoning as well as the reasoning of its neighbouring agents. Through several iterations of these debates, the models converge to a stable generation, much like humans reaching a common conclusion during a discussion.

To outline the algorithm in simpler terms:

1. Multiple instances of language models generate separate candidate responses for a given query.
2. Each model instance reads and critiques the responses of all other models, incorporating this feedback to update its own response.
3. This process is repeated for multiple rounds until a final answer is reached.

This approach encourages models to generate responses that align with their internal critiques and are reasonable in light of the responses from other agents. The resulting consensus among the models supports multiple chains of reasoning and allows for the exploration of various potential answers before arriving at a final response.

The MIT AI Lab conducted experiments to evaluate the effectiveness of this technique across different domains. The evaluation included tasks such as accurately stating facts about the biography of a famous computer scientist, validating factual knowledge questions, and predicting the next best move in a chess game. The results, as shown in Table 1 below, revealed that the Society of Mind approach outperformed other options in all categories, demonstrating its superiority in enhancing LLMs’ reasoning capabilities.

While this development holds immense promise, questions regarding the stopping criterion for these debates still remain. Determining when to conclude the discussions and reach a final answer is an essential consideration. Researchers and experts are actively exploring metrics such as perplexity or BLEU generation to define this criterion and ensure optimal performance.

The fusion of the Chain of Thoughts approach with the concept of the Society of Mind brings us closer to a future where language models exhibit enhanced reasoning and produce more accurate and nuanced responses.

Read more about AI: