“A key innovation in the Hierarchical Reasoning Model (HRM) paper is a training method inspired by Deep Equilibrium (DEQ) models that avoids full backpropagation through time.”