Language is the most important communication and expression tool for human beings, and it is also an important input and output method for artificial intelligence (AI) systems. With the development of deep learning, large pre-trained language models (such as GPT-3, T5, etc.) have been able to achieve amazing results on various natural language processing (NLP) tasks, and even answer some common sense or professional questions. However, these language models still have some limitations, such as lack of logical reasoning ability, lack of factual accuracy, lack of diversity and interpretability, etc.
To address these issues, researchers at the Massachusetts Institute of Technology (MIT) and the IBM Watson AI Lab (MIT-IBM Watson AI Lab) have proposed a new approach that uses language to understand machines. They introduced a strategy that utilizes discussions and debates between multiple AI systems to arrive at an optimal answer. This approach allows these large language models to improve adherence to factual data and improve the quality of decision-making.
Specifically, their approach consists of the following steps:
First, given a question, an initial answer is generated using a pre-trained encoder-decoder model (e.g., T5).
Then, another pre-trained encoder-decoder model (e.g., BART) is used to generate a critique of the initial answer, i.e., pointing out the shortcomings or errors of the initial answer.
Then, use the first model to respond to the criticism, i.e., explain or revise the initial answer.
Repeat the process until a consensus is reached or a preset number of rounds is exceeded.
Finally, an evaluation model (e.g., rouge) is used to score the final answers, selecting the answer with the highest score as the output.
The advantage of their approach is that it can take advantage of the interaction and competition between different AI systems to improve the quality and credibility of answers. Through discussion and debate, AI systems can check and correct each other, complement and refine each other, motivate and innovate each other, and thus arrive at a better answer. In addition, their approach can also use language to understand machine thought processes, improving the interpretability and intelligibility of answers by generating and presenting conversations.
Their method has been experimented on a number of publicly available datasets, and the results show that their method can significantly improve the number of answers and diversity, and their method can generate more factual and logical answers, as well as more creative and interesting answers than a single language model.
In conclusion, using language to understand machines is a multi-agent collaboration method, which can use discussions and debates between different AI systems to improve the quality and credibility of answers, and can also use language to understand the thinking process of machines to improve the interpretability and intelligibility of answers. This method provides a new way of thinking and direction for the application and development of large language models.