Artificial intelligence has long been measured by the size of its models and the scale of its training data. Bigger usually meant better. But a new AI system developed by researchers at Sapient, a Singapore-based AI firm, is challenging that assumption. Their Hierarchical Reasoning Model (HRM), modeled loosely on how the human brain processes information, has achieved results that rival, and in some cases surpass, massive language models like ChatGPT, despite being far smaller in size.
The breakthrough, unveiled in a recent preprint study, could mark a turning point in how scientists think about AI’s future. Instead of chasing ever-larger models, the focus may shift toward smarter, more efficient systems that reason in ways closer to humans.
The New AI Architecture
HRM is inspired by the human brain’s ability to manage information on multiple levels at once. Traditional large language models (LLMs) like ChatGPT rely on statistical pattern recognition across vast datasets.
The system uses two interconnected modules: a high-level abstract planner that works like a strategist, and a low-level computation engine that handles details. Together, they mimic how the brain balances broad planning with immediate decision-making.
This allows HRM to solve problems in a single forward pass, rather than depending on long step-by-step reasoning prompts.
Training Efficiency
One of the most striking aspects of HRM is its efficiency. The model has just 27 million parameters, a fraction of the billions used in ChatGPT, and it was trained on only 1,000 examples. Despite its small scale, HRM still managed to outperform much larger AI systems on some of the most demanding reasoning benchmarks.
This efficiency suggests that AI doesn’t need massive data or endless computational power to excel. Instead, architecture and training methods may be the key to unlocking smarter, more versatile systems.
The New AI Shows Promising Results
The true test of any AI system lies in its performance. HRM was evaluated on the ARC-AGI benchmark, widely regarded as one of the toughest reasoning challenges for artificial intelligence.
On the even tougher ARC-AGI-2, HRM achieved 5%, edging out other systems that scored closer to 1%.
What is more, HRM excelled in specialized reasoning tasks. It solved complex Sudoku puzzles with near-perfect accuracy and found optimal paths through mazes, tasks that often trip up traditional language models.
Reasoning has always been a weak spot for AI. While LLMs like ChatGPT generate text that feels natural, they struggle when asked to solve puzzles, plan strategies, or handle problems that require multi-step logic. By outperforming them in these areas, HRM suggests that brain-inspired models may be the future.
Despite the promising results, experts urge caution. The HRM study is still a preprint, meaning it has not yet undergone peer review. Independent reviewers have also noted that the model’s performance may stem as much from its training refinement process as from its architecture. In other words, it is not yet clear whether the brain-inspired design is the main factor behind its success.
