New recurrent AI model solves complex reasoning with tiny dataset and no pre-training

A new approach to machine reasoning sidesteps the computational overhead and data hunger of current large language models by drawing inspiration from how brains organize thought across different timescales.

The Hierarchical Reasoning Model (HRM), introduced by researchers including Yuhao Sun, compresses reasoning into a single forward pass using just 27 million parameters—orders of magnitude smaller than modern LLMs. The model learns to tackle intricate problem-solving tasks from only 1,000 training samples, without requiring pre-training or Chain-of-Thought supervision.

How it works#

HRM's architecture mirrors the brain's hierarchical structure through two interlocked recurrent modules. A slower, abstract-reasoning layer handles high-level planning and goal decomposition, while a faster, detail-oriented layer executes concrete computational steps. This separation mirrors how human cognition balances deliberate strategic thinking with rapid tactical responses.

Unlike dominant approaches that chain explicit reasoning steps (Chain-of-Thought), HRM performs sequential reasoning implicitly during a single forward pass. The model receives no supervision on intermediate reasoning steps—it learns to decompose and solve problems end-to-end.

Demonstrated performance#

The model achieves near-perfect accuracy on demanding tasks:

• Complex Sudoku puzzles requiring logical constraint satisfaction
• Maze pathfinding in large environments demanding optimal route discovery
• Abstraction and Reasoning Corpus (ARC) tasks, where HRM surpassed much larger models with extended context windows

ARC represents a key challenge for measuring progress toward artificial general intelligence, as it demands abstract pattern recognition and compositional reasoning that current systems struggle with.

Why this matters#

Current LLM reasoning methods carry three core limitations: fragile task decomposition that fails on novel problems, massive training data requirements, and high computational latency from generating intermediate reasoning chains. By eliminating these bottlenecks, HRM points toward more efficient, robust, and scalable approaches to machine reasoning that may better capture the fundamental computational principles underlying human thought.