Hierarchal reasoning is a brain-inspired approach to 📝Artificial Intelligence (AI) in which 📝Large Language Model (LLM)s process information across multiple levels of abstraction and timescales. This method contrasts with 📝Chain of Thought Reasoning enabling models to perform complex reasoning internally, without relying on explicit textual logs.
Hierarchal Reasoning architectures, such as the 📝Hierarchical Reasoning Model (HRM), employ two modules: a high-level system responsible for abstract planning and a low-level system dedicated to rapid, granular computation. Together, these modules iteratively refine states until converging on both local and global solutions. Unlike standard Transformer-based models that have fixed layer depths, HRMs achieve arbitrarily deep reasoning through nested recurrent structures. Research highlights that HRMs provide greater efficiency, requiring fewer parameters and less data while outperforming larger models in tasks such as Sudoku, mazes, and the Abstraction and Reasoning Corpus. They are further inspired by human cognitive architecture, where higher cortical regions handle planning and lower regions execute details. By offering implicit task decomposition and adaptive depth, HRMs represent a significant step toward more general-purpose artificial intelligence.