4 ms·
One thing about the reasoning is that models are trained to generate a chain of thoughts, but it doesn’t have to be correct, accurate, or reflect the underlying
by dgellow 2mo ago
One thing about the reasoning is that models are trained to generate a chain of thoughts, but it doesn’t have to be correct, accurate, or reflect the underlying logic of the LLM. It’s the same problem we have with the output, it is something plausible, but not that reliable
- anon291 2mo agoI do the same thing in my head. There is no underlying logic to an llm. Logic is an external construct alien to human like forms of reasoning.
- trhway 2mo ago>but it doesn’t have to be correct, accurate, or ... why we don't do GAN here, ie. second model verifying correctness/accuracy/etc. ?