Definition
Plain language
A design that runs one part of a model quickly and another part more slowly to deliberate.
As stated in the literature
An architectural pattern, as in HRM-Text, where a fast L-module updates many times per slow H-module update, inspired by frontoparietal-loop hierarchical processing in the brain.
Also called: fast-slow
Why it matters: Splitting compute between fine-grained and coarse-grained timescales mirrors how brains handle hierarchical tasks and can improve long-horizon reasoning without ballooning cost.
For example, one module might update every token while a second module only updates every sixteen tokens to carry longer-term plans.
Heard on the show
“The fast/slow split is what's new here.”Episode 074 — How a Fifteen-Hundred-Dollar Training Run Matched Llama and Gemma on Reasoning