Tiny AI model HRM outpaces ChatGPT by 100x in reasoning
A new artificial intelligence model developed by Singapore-based Sapient Intelligence is challenging the prevailing “bigger is better” philosophy in AI development. This innovative model, named the Hierarchical Reasoning Model (HRM), draws inspiration from the human brain’s architecture to solve complex reasoning problems with remarkable efficiency.
Unlike many contemporary large language models, which often rely on a “shallow” architecture and step-by-step Chain-of-Thought (CoT) prompting, HRM adopts a hierarchical structure. Researchers note that CoT methods can be prone to failure if a single step goes awry. HRM, by contrast, mimics the brain’s approach with two distinct, interconnected modules: a high-level “planner” that engages in slow, strategic thinking (akin to planning a chess move), and a low-level “worker” that performs rapid calculations (like instant facial recognition). This design allows HRM to process complex problems deeply in a single pass, learning to reason from a limited set of examples without requiring extensive pre-training on vast datasets.
Despite its remarkably small size of just 27 million parameters, HRM has demonstrated superior reasoning capabilities in various benchmarks. On the ARC-AGI benchmark, often considered an IQ test for AI, HRM achieved a score of 40.3%, significantly outperforming OpenAI’s o3-mini-high (34.5%) and Claude 3.7 (21.2%). The model’s performance was even more pronounced on specialized tasks: it successfully solved 55% of Sudoku-Extreme puzzles and found the optimal path in 74.5% of 30x30 mazes, while Claude 3.7 and o3-mini-high scored 0% on both. To put HRM’s efficiency into perspective, the original GPT-1 model featured 117 million parameters, over four times HRM’s size. One of HRM’s creators, Guan Wang, highlighted its lean design, noting it can be trained to solve professional-level Sudoku in just two GPU hours.
The implications of HRM’s success are substantial. It suggests that architectural innovation can yield significant advancements in AI, potentially reducing the reliance on massive computational resources. This could lead to more affordable AI deployment, enabling advanced models to run efficiently on a single GPU, and drastically faster training times, measured in hours rather than months. Furthermore, HRM’s design promises improved reasoning capabilities without the need for prohibitively expensive computing infrastructure. The model’s code is also open-source, promoting wider access and further development.
While some skeptics argue that HRM’s current skills might be too specialized, its early performance indicates a promising direction for AI research. This brain-inspired approach is part of a broader trend exploring alternative AI architectures, including Sakana’s continuous thought machines, 1-bit LLMs (bitnets), and diffusion models, which Google is actively experimenting with. These emerging architectures, though currently in their early stages, hint at a future where advanced AI is not exclusively confined to large data centers but can operate efficiently on local machines, democratizing access to powerful artificial intelligence.