Definition
Plain language
The fast, reflexive half of an HRM-style recurrent architecture.
As stated in the literature
In HRM and HRM-Text, the low-level module that updates multiple times per H-module step, capturing fast token-level processing in a hierarchical recurrent design.
Why it matters: The fast inner loop is what lets hierarchical reasoners do detailed work without paying the cost of running their full slow module on every step.
For example, while the slow H module updates once to set high-level intent, the L module might iterate four times to handle fine-grained token-level updates underneath it.
Heard on the show
“One forward pass through the model runs the L module three times, then updates H once.”Episode 074 — How a Fifteen-Hundred-Dollar Training Run Matched Llama and Gemma on Reasoning