HRM Hierarchical Reasoning Model: A Breakthrough in Complex Reasoning Tasks with 27 Million Parameters

The Hierarchical Reasoning Model (HRM) open-sourced by the Sapient team is inspired by the brain's hierarchical processing mechanism and adopts a dual-module recurrent architecture. With only 27 million parameters and 1000 training samples, and without pre-training or chain-of-thought (CoT) data, it achieves near-perfect performance on complex reasoning tasks like Sudoku, mazes, and ARC-AGI, providing new ideas for efficient intelligent systems.

In the field of artificial intelligence, complex reasoning ability is an important measure of intelligence level. Mainstream large language models rely on chain-of-thought (CoT) technology to complete complex tasks, but they face issues like fragile task decomposition, high data demand, and long reasoning latency, calling for more efficient architectural solutions.

HRM's core consists of two recurrent modules:

• High-level module: Slow abstract planning, runs at low frequency, responsible for understanding the overall structure of the problem and formulating strategies (similar to the function of the brain's prefrontal cortex);
• Low-level module: Fast detailed computation, runs at high frequency, executes high-level strategies and handles specific details. The two modules interact bidirectionally to form a complete reasoning cycle, enabling complex sequence reasoning to be completed in a single forward pass without intermediate supervision.

HRM has significant advantages:

1. Extremely small parameter scale: Only 27 million parameters, low training cost, fast reasoning speed, and runnable on consumer-grade hardware;
2. Minimal data requirement: Only 1000 training samples needed, suitable for data-scarce fields;
3. No dependency on pre-training or CoT: Trained from scratch, no pre-training phase or CoT annotated data required;
4. Single inference completion: Solves complex tasks in a single forward pass, reducing reasoning latency.

HRM performs outstandingly in multiple benchmark tests:

• Complex Sudoku: Near-perfect accuracy on extremely difficult 9x9 Sudoku;
• Large mazes: Finds optimal paths in 30x30 difficult mazes;
• ARC-AGI benchmark: Outperforms models with larger parameter sizes, demonstrating general reasoning potential (ARC is a key benchmark for measuring general intelligence).

HRM is deployment-friendly:

• Training efficiency: Training an extreme Sudoku model takes about 10 hours on a single RTX4070 laptop GPU, and training ARC tasks takes about 24 hours on an 8-card cluster;
• Hardware requirements: Supports CUDA extensions and FlashAttention, consumer-grade GPUs can meet training and inference needs;
• Pre-trained models: HuggingFace has released pre-trained checkpoints for tasks like Sudoku, mazes, and ARC, which can be used directly.

The release of HRM is of great significance:

• Proves that small models can compete with large models in complex reasoning through architectural design, promoting the development of efficient AI;
• The hierarchical mechanism provides a case for interdisciplinary research between neuroscience and AI;
• Moves towards general reasoning systems, suitable for scenarios like edge computing and mobile devices; The team has open-sourced the code, datasets, and pre-trained models, providing references for researchers.