Definition
Plain language
How many times a model effectively re-runs the same blocks to refine its answer.
As stated in the literature
The number of iterations a recurrent or looped architecture applies its core block during a forward pass, providing additional computation per parameter; central to HRM-Text and attractor-model designs.
Why it matters: It lets a model spend more compute on harder inputs without growing in size, decoupling reasoning depth from parameter count.
For example, a small model might loop its core block four extra times on a hard puzzle, effectively giving itself more thinking depth without adding new parameters.
Heard on the show
“And I want to land the payoff piece, which is the question of whether the recurrent depth actually does anything.”Episode 074 — How a Fifteen-Hundred-Dollar Training Run Matched Llama and Gemma on Reasoning