Definition
Plain language
Letting a model keep thinking inside its own head, feeding its internal thoughts straight back in as the next input, instead of forcing each thought into words first.
As stated in the literature
A latent-reasoning approach (originating with Coconut) that loops a model's top-layer thought vector back as the next input without decoding it to a token; reasoning capacity and compute decouple, but the absence of token-level probabilities made it incompatible with on-policy RL until switchable boundary tokens were introduced.
Why it matters: It lets a model reason without spending effort turning every intermediate thought into text, separating how much it thinks from how much it has to write.
For example, instead of writing out 'let me carry the one' as words, the model passes its raw internal thought straight back to itself to continue calculating.
Heard on the show
“They call it hidden-state recurrence.”Episode 141 — How Two Tokens Reopened a Reasoning Method the Field Had Given Up On