“This process allows the HRM to perform a sequence of distinct, stable, nested computations, where the H-module directs the overall problem-solving strategy and the L-module executes the intensive search or refinement required for each step.” This nested-loop design allows the model to reason deeply without needing huge amounts of data.