Can a language model reason without writing its steps down? Coconut feeds GPT-2's last hidden state back in as the next ...

Can a language model reason without writing its steps down? Coconut feeds GPT-2's last hidden state back in as the next input instead of a token, and beats written chain of thought on two logic benchmarks with far fewer tokens. Its evidence that the model searches breadth-first in those hidden states comes from one of them, ProsQA, but a later study finds Coconut scores about the same with the hidden-state steps cut to zero.https://benjaminhan.net/posts/20261103-coconut-continuous-latent-reasoning/?utm_source=mastodon&utm_medium=social#AI #LLMs #Reasoning #COLM