# Hierarchical Reasoning Model

**Authors:** Guan Wang, Jin Li, Yuhao Sun, Xing Chen, Changling Liu, Yue Wu, Meng Lu, Sen Song, Yasin Abbasi Yadkori  
**arXiv:** [2506.21734](https://arxiv.org/abs/2506.21734)  
**Date:** 2025-06-26  
**Summary page:** https://www.summarizepaper.com/en/paper/2506.21734v1/

## Key points

- The Hierarchical Reasoning Model (HRM) is a revolutionary recurrent architecture developed by Sapient Intelligence in Singapore to address reasoning challenges in artificial intelligence.
- HRM utilizes a hierarchical and multi-timescale processing approach inspired by the human brain, consisting of two interdependent recurrent modules: a high-level module for abstract planning and a low-level module for detailed computations.
- With only 27 million parameters, HRM outperforms larger models on complex reasoning tasks with just 1000 training samples, particularly excelling at tasks like solving complex Sudoku puzzles and navigating large mazes.
- HRM achieves exceptional performance on benchmarks like the Abstraction and Reasoning Corpus (ARC) without requiring pre-training or Chain-of-Thought (CoT) data.
- These results demonstrate HRM's potential as a transformative advancement towards universal computation and general-purpose reasoning systems in AI.

## Summary

The Hierarchical Reasoning Model (HRM) is a revolutionary recurrent architecture developed by the team at Sapient Intelligence in Singapore to tackle the challenge of reasoning in artificial intelligence. Unlike current large language models that rely on Chain-of-Thought (CoT) techniques, HRM utilizes a hierarchical and multi-timescale processing approach inspired by the human brain. This model consists of two interdependent recurrent modules: a high-level module for abstract planning and a low-level module for detailed computations. With only 27 million parameters, HRM achieves exceptional performance on complex reasoning tasks with just 1000 training samples, surpassing larger models on benchmarks like the Abstraction and Reasoning Corpus (ARC). The model does not require pre-training or CoT data and excels at tasks such as solving complex Sudoku puzzles and navigating large mazes. These results highlight HRM's potential as a transformative advancement towards universal computation and general-purpose reasoning systems in AI.