Definition
Plain language
A method that lets a model think silently inside its own internal state, with special start and stop markers around the silent stretch so it can still be trained with reinforcement learning.
As stated in the literature
A framework for switchable latent reasoning that wraps Coconut-style hidden-state recurrence in two discrete boundary tokens marking entry into and exit from silent latent computation; because only the boundary tokens carry probabilities, the latent steps contribute zero gradient, keeping the scheme compatible with on-policy RL.
Why it matters: It lets a model think in its hidden state while staying trainable with reinforcement learning, combining silent reasoning with on-policy optimization.
For example, a model can drop into a stretch of silent internal reasoning marked by a start token and a stop token, then speak its conclusion.
Heard on the show
“They call the whole framework SWITCH, because the model is learning to switch into and out of thinking silently.”Episode 141 — How Two Tokens Reopened a Reasoning Method the Field Had Given Up On