The Singaporean company Sapient Intelligence introduced a new AI architecture called the Hierarchical Reasoning Model, which achieves results comparable to or even surpassing large language models in complex reasoning tasks. The HRM model uses an approach inspired by the workings of the human brain, combining two recurrent modules for slow planning and fast computations. This structure allows the system to perform deep multi-step reasoning with significantly less data and memory than other modern models require.
HRM demonstrated high efficiency in complex tasks, particularly on ARC-AGI benchmarks, extreme sudoku, and maze tasks. For example, in the “Sudoku-Extreme” and “Maze-Hard” tests, modern CoT models did not solve any tasks, while HRM achieved nearly perfect accuracy, learning from only a thousand examples for each task. On the ARC-AGI benchmark, HRM with 27 million parameters scored 40.3%, surpassing the performance of larger models such as o3-mini-high and Claude 3.7 Sonnet.
According to the company’s founder Guan Wang, HRM provides not only accuracy but also significantly faster task processing. The architecture reduces delays during task execution and lowers costs, as training the model to a professional level in sudoku requires only two hours of GPU work, and for ARC-AGI, up to 200 hours. This makes the model suitable for use in environments with limited computational resources and data scarcity.
HRM is recommended for complex and deterministic tasks where sequential decision-making or long-term planning is required, particularly in robotics or scientific research. The model gradually reduces the number of steps needed to solve a task during training, allowing it to reach an expert level.
Sapient Intelligence is already working on developing HRM towards universal reasoning modules that can be used in medicine, climate forecasting, and robotics. The developers plan to implement self-correction features that will distinguish these models from modern text systems.
