Hacker News Reader

I'm behind on reading but don't all models use continuous embeddings to represent reasoning?

I believe the "continuous" in Coconut means that the CoT is in the continuous latent space, instead of being on output tokens (see Fig. 1).