Definition
Plain language
An early tiny-reasoner model architecture for puzzle-solving via iterative latent refinement.
As stated in the literature
Hierarchical Reasoning Module, a small recurrent architecture for structured reasoning tasks, prior art for attractor-style and looped reasoners; also used as the basis of HRM-Text's two-module fast/slow language model design.
Also called: HRM-Text
Why it matters: It's a precursor to current interest in small, looped reasoning models that try to compete with large transformers via depth rather than width.
For example, HRM was an early demonstration that a small recurrent model could solve ARC-style puzzles by iteratively refining a hidden representation.
Heard on the show
“There's a line of work on tiny recursive reasoners — HRM, TRM — that hit really impressive scores on tasks like Sudoku and maze-solving with just a few million parameters.”Episode 041 — When the Iteration Teaches the Model to Skip the Iteration