It's also interesting because in humans the existence of "Aha!" moments that are not preceded by or are only loosely related to a chain of thought is taken as the proof of the fundamental mystery and irreproducibility of human intelligence. Now the same argument is made to deny that LLMs actually think. Go figure.
replyThat's my personal theory too. The model is stuffing its own context with vaguely related tokens, which helps the attention heads retrieve the right tokens.
replyYup. You basically just need something for probability to push off of
replyI assume theyre searching the local gradient to see if theres a better descent before proceeding.
replyLLMs dont do gradient descent to generate tokens.
replyThey are trained by gradient descent, but inference doesnt involve it.
I don't think there's anything like that going on. They just word vomit into a secondary area, and then there is an internal prompt that says "clean this up and summarize for the user".
replyLess "internal prompt" and more "they are trained to summarize after a </think> token"
replyThe training methods try not to apply any particular rules to the contents of the thinking text. That's called "optimization pressure on CoT" and is thought to reduce safety by inducing the model to lie (or stop clearly printing its intentions) in the thinking text.
reply