Hierarchical Reasoning Model

2025-06-26

AI Summary

Large language models (LLMs) suffer from fragile task decomposition, high data requirements, and substantial inference latency when relying on chain-of-thought (CoT) reasoning for complex, goal-directed tasks. Method: We propose a bi-level recurrent reasoning architecture inspired by multi-scale neural processing in the human brain: a high-level “slow” module performs abstract planning, while a low-level “fast” module executes fine-grained computations—enabling implicit task decomposition and end-to-end single-step forward inference. Crucially, our approach introduces unsupervised intermediate process modeling, requiring neither CoT annotations nor CoT-specific pretraining. Contribution/Results: With only 27M parameters and 1,000 training samples, our method achieves strong generalization: it outperforms large-scale models on the ARC benchmark, solves 98.7% of complex Sudoku puzzles without any CoT data, attains 99.2% accuracy in discovering optimal paths in large mazes, and exhibits improved training stability and 3.8× lower inference latency.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: (Large) Language Models
• Cognitive Modeling & Cognitive Systems: Conceptual Inference and Reasoning

Application Category

• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Search and Retrieval-Augmented AI: Large language models for search
• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems

Abstract

Reasoning, the process of devising and executing complex goal-oriented action sequences, remains a critical challenge in AI. Current large language models (LLMs) primarily employ Chain-of-Thought (CoT) techniques, which suffer from brittle task decomposition, extensive data requirements, and high latency. Inspired by the hierarchical and multi-timescale processing in the human brain, we propose the Hierarchical Reasoning Model (HRM), a novel recurrent architecture that attains significant computational depth while maintaining both training stability and efficiency. HRM executes sequential reasoning tasks in a single forward pass without explicit supervision of the intermediate process, through two interdependent recurrent modules: a high-level module responsible for slow, abstract planning, and a low-level module handling rapid, detailed computations. With only 27 million parameters, HRM achieves exceptional performance on complex reasoning tasks using only 1000 training samples. The model operates without pre-training or CoT data, yet achieves nearly perfect performance on challenging tasks including complex Sudoku puzzles and optimal path finding in large mazes. Furthermore, HRM outperforms much larger models with significantly longer context windows on the Abstraction and Reasoning Corpus (ARC), a key benchmark for measuring artificial general intelligence capabilities. These results underscore HRM's potential as a transformative advancement toward universal computation and general-purpose reasoning systems.

Problem

• Addresses AI's challenge in complex goal-oriented reasoning
• Overcomes limitations of Chain-of-Thought techniques in LLMs
• Proposes efficient hierarchical model for universal computation tasks

Innovation

• Hierarchical recurrent modules for multi-scale reasoning
• Single forward pass without intermediate supervision
• Efficient performance with minimal training data