Hierarchical Reasoning Model

Authors: Guan Wang, Jin Li, Yuhao Sun, Xing Chen, Changling Liu, Yue Wu, Meng Lu, Sen Song, Yasin Abbasi Yadkori
Paper: https://arxiv.org/abs/2506.21734
Code: https://github.com/sapientinc/HRM

TL;DR

WHAT was done? The paper introduces the Hierarchical Reasoning Model (HRM), a novel recurrent architecture inspired by the human brain. HRM features two interdependent modules operating at different timescales: a high-level module for slow, abstract planning, and a low-level module for rapid, detailed computation. This design enables significant computational depth while maintaining stability through a "hierarchical convergence" process. Training is made highly efficient by a one-step gradient approximation that bypasses Backpropagation Through Time (BPTT), and the model dynamically allocates resources using an Adaptive Computational Time (ACT) mechanism (see my series of posts on ACT).

WHY it matters? HRM fundamentally challenges the dominant "bigger is better" paradigm of large language models (LLMs). With only 27M parameters and trained from scratch on just ~1000 examples, it achieves near-perfect performance on complex reasoning tasks (like Sudoku-Extreme and Maze-Hard) where massive Chain-of-Thought (CoT) models completely fail. This suggests that sophisticated architecture and computational depth, rather than sheer scale, can be a more efficient and robust path to advanced AI reasoning. Furthermore, the model's learned internal structure spontaneously develops a brain-like dimensionality hierarchy, providing strong empirical validation for a brain-inspired approach.

Details

A New Blueprint for AI Reasoning

The pursuit of artificial general intelligence has long been intertwined with the challenge of complex reasoning. The prevailing approach, embodied by large language models (LLMs), has been to scale up model size and rely on techniques like Chain-of-Thought (CoT) prompting to externalize reasoning into sequential, human-readable steps. While powerful, this paradigm suffers from brittleness, high latency, and an insatiable appetite for data. A recent paper introduces the Hierarchical Reasoning Model (HRM), presenting a compelling alternative that suggests the key to reasoning may lie not in sheer scale, but in a more profound, brain-inspired architectural depth.