Brain-inspired, multi-level reasoning & planning AI model
A revolutionary 27M-parameter AI model that performs complex sequential reasoning in a single forward pass. Featuring dual recurrent modules for high level planning and sharp detail, it outperforms larger models on puzzles and maze challenges.
HRM is open-source with full implementation details, trained checkpoints, and evaluation code available, enabling researchers to reproduce results, modify the architecture, and build upon the hierarchical reasoning framework.
HRM implements a recurrent architecture with high-level abstract planning and low-level detailed computation modules, mirroring human brain processing, allowing researchers to study and deploy brain-inspired AI reasoning systems.
HRM outperforms larger models on the Abstraction and Reasoning Corpus (ARC), a key AGI benchmark, demonstrating potential for universal computation and general-purpose reasoning without Chain-of-Thought limitations.
HRM maintains training stability and efficiency through its recurrent design, executing sequential reasoning in a single forward pass without explicit intermediate supervision, reducing latency and computational overhead compared to CoT approaches.
HRM achieves near-perfect performance on challenging reasoning tasks (Sudoku, maze solving, ARC benchmarks) using only 1,000 training samples and 27 million parameters, enabling efficient reasoning without extensive data requirements or pre-training.