Скачать или смотреть Hierarchical Reasoning Model: Universal Computation from Latent Reasoning. HRM Model Architecture.

Hierarchical Reasoning Model: Universal Computation from Latent Reasoning.

The podcast provides the technical overview of the Hierarchical Reasoning Model (HRM), a novel recurrent neural architecture inspired by the multi-timescale and hierarchical processing found in the human brain, designed to overcome the limitations of current Chain-of-Thought (CoT) methods in Large Language Models (LLMs). HRM uses two interdependent recurrent modules—a slow, high-level planner and a rapid, low-level executor—to achieve significant computational depth and efficiently solve complex reasoning tasks like Sudoku, maze navigation, and the Abstraction and Reasoning Corpus (ARC) challenges. Crucially, HRM achieves near-perfect performance on these tasks with minimal data (around 1,000 examples) and without pre-training, significantly outperforming much larger CoT-based models. The architecture incorporates techniques like hierarchical convergence to maintain computational activity over many steps and an approximate gradient method to avoid the memory demands of Backpropagation Through Time (BPTT). Furthermore, the document shows that HRM's internal structure develops an emergent dimensionality hierarchy during training, paralleling organizational principles observed in the mammalian cortex.