Definition
Plain language
An earlier method that taught AI models to reason in their own internal vectors instead of writing the reasoning out in words.
As stated in the literature
A continuous-latent reasoning approach (Chain of Continuous Thought) where a model reasons in hidden-state vectors rather than tokens; MIRAGE builds on it and replaces its serial latent computation with parallel refinement.
Why it matters: It showed that models can reason in their own internal representations instead of words, opening the door to faster thinking that isn't slowed by writing every step out.
For example, rather than typing out 'first I add these, then I carry the one,' a Coconut-style model holds those steps as internal number-vectors and only writes the final answer.
Heard on the show
“A paper called Coconut, from a couple years back, was the one that first taught models to reason in continuous vectors instead of words.”Episode 115 — Teaching a Phone Agent to Reason Silently, And Keeping It Honest