Abstract

One of the key features of the human brain is its deep reasoning ability derived from its structural dynamics. In contrast, current artificial intelligence models are constrained by fixed depth, with a predetermined number of layers that do not adapt based on the complexity of the task at hand. Additionally, the monolithic nature of their latent space restricts the representation of diverse, task-specific features, further impairing their reasoning flexibility. To tackle the current limitations of reasoning in large language models, we present the Hierarchical Reasoning Model (HRM), a novel architecture that sets new benchmarks in artificial general intelligence (AGI) by leveraging neurocognitive principles of modularity, recurrence, and synchronised rhythms.

HRM consists of two modules that operate hierarchically and at different frequencies: the Foundational Module (FM) receives input and performs rapid, iterative local processing through multiple recurrent cycles, while the Metamodule (MM) operates at a slower temporal scale, integrating abstract representations and providing top-down strategic modulation to the FM. Empirical evaluations demonstrate that HRM , with 0.027B parameters, significantly outperforms state-of-the-art models, including current LLMs, on complex reasoning tasks such as the Abstraction and Reasoning Corpus (40.3% in ARC -1 and 5% in ARC -2) and Sudoku (55%), using around 1000 training samples. These results highlight HRM ’s ability to generalise from minimal data and execute multi-step reasoning effectively. Model analysis demonstrates that the FM refines local features through iterative self-recurrence cycles in a way similar to sensory cortices in the brain, while the MM acts like the prefrontal cortex, providing strategic, higher-order guidance and ensuring convergence through slower, more deliberate updates. Furthermore, HRM exhibits key emergent properties, including self-correction, functional segregation, and hierarchical dimensionality, which enable dynamic exploration, efficient global reasoning, and adaptive computation. In summary, HRM integrates key brain-inspired principles to propose a compelling framework for AGI , offering significant advancements in both generalization and reasoning.