Using multi-agent discussion to improve reasoning and factual accuracy of Long Language Models (LLMs)

In recent times, big language models have managed to grab everyone’s attention with their advanced capabilities. LLMs with some outstanding language comprehension and production capabilities, such as OpenAI’s GPT-3.5, the latest multimodal GPT 4, etc., are being used significantly by industries. Generating meaningful answers to questions, summarizing text prompts, translating languages, and transforming text to text are some of the use cases.

LLMs can efficiently produce coherent text, understand and respond to prompts, and even learn from a small number of instances, called low-try learning. With few shot learning, LLMs use supervised information to classify new data with just a few training samples. Since LLMs have scope for improvement, in a recent research paper, a team of researchers from MIT and Google Brain proposed a complementary approach based on “multi-agent discussion” to improve the quality of linguistic responses generated by LLMs. .

The team has introduced a mechanism where multiple instances of the LLM participate to propose and argue their unique answers and reasoning processes over multiple rounds, rather than relying solely on one model instance. The goal is to arrive at a final answer that has been carefully reviewed and improved through a collaborative effort. This complementary method for improving language responses uses the ‘partnership of minds’ approach, which is inspired by the idea that the collective intelligence of multiple minds working together can lead to better performance and more accurate results.

This approach involves a number of models or agents, all of whom are asked the same question at the outset. By allowing these models to repeatedly evaluate and revise their actions in light of responses from other agents, the goal is to improve the performance of these models. The ‘multi-agent debate’ used in this method has been used to improve the deductive reasoning and factual accuracy of language models in order to use discussion between various instances of the language model to achieve better response output.

The team has seen significant improvements in mathematical and strategic reasoning using the “society of minds” approach, thus showing how the collective intelligence of multiple LLM instances leads to better performance. The suggested method also addresses the formation of false conclusions and hallucinations, a known weakness of modern models. The team has found that their method reduces the likelihood of such errors and increases the factual value of the generated content.

The adaptability of this approach is one of its benefits, as it can be used with existing black-box LLMs without requiring significant changes. All the investigated tasks follow the same process, with the same indications, which guarantees consistency and simplicity of use. Upon evaluation, the team has found that increasing the number of agents in a multi-agent discussion or increasing the number of discussion rounds improves the performance of the models. It has also been found that multi-agent discussion can allow two different instances of language models, such as ChatGPT and Bard, to cooperatively solve a task that they cannot solve individually.

In conclusion, the ‘partnership of minds’ strategy has the potential to greatly improve LLM performance, creating new opportunities for advances in language creation and understanding. By using this method, LLMs can provide more accurate and reliable answers, have greater reasoning skills, and make fewer errors that are often found in language models.

review the Paper, Code, and Project. Don’t forget to join our 22k+ ML SubReddit, discord channel, and electronic newsletter, where we share the latest AI research news, exciting AI projects, and more. If you have any questions about the article above or if we missed anything, feel free to email us at [email protected]

🚀 Check out 100 AI tools at AI Tools Club