Definition
Plain language
Getting more reasoning out of a model by having it think harder in its internal state rather than by writing more words.
As stated in the literature
A class of test-time scaling techniques that increase computation per token through depth recurrence, iterative refinement, or hidden-state persistence rather than through longer generated chains of thought.
Why it matters: It promises gains from spending more compute per token without bloating context length or producing visible scratchpads.
For example, instead of generating a longer chain of thought, the model loops its hidden state through extra recurrent steps before producing the next token.