Definition
Plain language
A setting where the model works through a problem privately before giving its answer.
As stated in the literature
A configurable inference regime in which the model emits extended internal reasoning tokens prior to the user-visible response; in this study it appears to aid recovery after capitulation but not to prevent initial yielding.
Also called: thinking enabled, thinking disabled
Why it matters: Extra private deliberation can help a model catch and correct its own mistakes, though it does not necessarily stop it from caving in the first place.
For example, asked a tricky logic puzzle, the model first writes out several lines of private reasoning the user never sees, then replies with just the answer.
Heard on the show
“But the aggregate MATH number is below the base model with thinking enabled.”Episode 040 — Two Frozen Models Learn to Whisper: Coupling Through Hidden States