Can Language Model Self Improve?
Artificial Intelligence & Machine Learning
Could Large Language Models use one another to validate all of their output?
Given the way these generative models work chances that they come up with the same responses are likely not very high, in addition they are likely also all going to be trained on very similar corpus, at least the general purpose one, as a result, making it somewhat likely they might be prone to making the same mistakes at times when they do overlap.
Moreover, requiring overlap might also force these to stick with broader answers than specific ones. So far, we’ve seen this quite a bit with GPT in the general purpose cases again. Thus, making the responses more generic and less interesting too.
In conclusion, I am not sure how much of their errors can become fully eliminated in the general purpose case. Lastly, though, may become better to think more broadly about solutions outside of these in specific applications.
Balaji Padmanabhan | USF Muma College of Business
See Dr. Padmanabhan’s other article: What Happened With Google Bard?
