Article contents:
- Introduction: A new way of thinking for AI
- What is the Hierarchical Reasoning Model (HRM)?
- Stunning superiority in intelligence tests
- Beyond “Chain of Thought”: How does HRM think?
- Smaller and more efficient
- Success in hard tasks: From Sudoku to mazes
- A note of caution: Study not yet peer-reviewed
- Conclusion: A step toward more human-like AI
Introduction: A new way of thinking for AI
Scientists have developed a new type of artificial intelligence model that can reason differently than most large language models (LLMs) like ChatGPT, leading to vastly better performance on key logical reasoning benchmarks. This new model, called the Hierarchical Reasoning Model (HRM), is inspired by hierarchical and multi-scale temporal processing in the human brain, and may represent a qualitative leap in machines’ ability to solve complex problems.
What is the Hierarchical Reasoning Model (HRM)?
The Hierarchical Reasoning Model is inspired by the way different brain regions integrate information over varying time intervals (from fractions of a second to minutes). The model consists of two units: a high-level unit responsible for slow, abstract planning, and a low-level unit that handles fast, detailed calculations. This dual structure mimics how the human brain deals with problems, where long-term strategic planning occurs in parallel with fast tactical execution.
Stunning superiority in intelligence tests
When researchers tested HRM on the ARC-AGI benchmark—a very difficult test designed to measure how close models are to achieving Artificial General Intelligence (AGI)—the system achieved amazing results. HRM scored 40.3%, compared to 34.5% for OpenAI’s model, 21.2% for Anthropic’s Claude, and 15.8% for DeepSeek’s model. This significant superiority indicates that the HRM architecture may be much better suited for abstract reasoning and problem-solving tasks than traditional large language model architectures.
Beyond “Chain of Thought”: How does HRM think?
Most advanced large language models use “Chain-of-Thought” reasoning, where a complex problem is broken down into simpler intermediate steps expressed in natural language. But scientists argue that this approach has drawbacks, namely that it requires massive training data and suffers from high latency. Instead, HRM executes sequential reasoning tasks in a “single forward pass,” without any explicit supervision of intermediate steps. It works through “iterative refinement”—a computing technique that improves solution accuracy by repeatedly refining an initial estimate—over several short bursts of “thinking.”
Smaller and more efficient
One of the most impressive aspects of HRM is its efficiency. The model contains only 27 million parameters and uses 1,000 training samples. By comparison, most advanced large language models contain billions or even trillions of parameters. For example, some estimates suggest that the recently released GPT-5 contains between 3 and 5 trillion parameters. This efficiency makes it possible to run powerful models on less expensive hardware that consume less power.
Success in hard tasks: From Sudoku to mazes
HRM achieved near-perfect performance on difficult tasks such as complex Sudoku puzzles—tasks that traditional large language models have failed to accomplish—in addition to excelling at finding the optimal path in mazes. This demonstrates that its logical reasoning capability is not merely theoretical, but can be applied to practical problems requiring multi-step planning and solving.
A note of caution: Study not yet peer-reviewed
Although the results are promising, the study has not yet been peer-reviewed. However, the organizers of the ARC-AGI benchmark tried to recreate the results themselves after the scientists open-sourced their model. Although they were able to reproduce the numbers, they found that the hierarchical structure itself had little impact on performance, and instead, an undocumented “refinement” process during training was what led to the massive performance gains. This indicates that further research is needed to understand exactly what makes the model successful.
Conclusion: A step toward more human-like AI
The HRM model represents an exciting new approach in AI design. By drawing inspiration from the architecture of the human brain, scientists have created a model that excels in logical reasoning tasks that have long been a weak point for large language models. Even if the precise details are still under investigation, this work opens the door to a new generation of more efficient and problem-solving-capable AI systems, bringing us one step closer to achieving truly human-like artificial intelligence.
FAQs
Q: What is the main difference between HRM and models like ChatGPT?
A: ChatGPT relies on a linguistic “Chain-of-Thought” approach, whereas HRM uses a brain-inspired hierarchical architecture to process abstract planning and detailed calculations separately, making it better at logical reasoning.
Q: What is the ARC-AGI benchmark?
A: It is a set of abstract intelligence tests designed to measure AI’s ability to solve problems it has never seen before, and it is considered one of the hardest benchmarks for evaluating progress toward artificial general intelligence.
Q: Why is the small model size important?
A: Smaller models require less computing power and less training data, making them cheaper to develop and run, and more accessible to a wide range of researchers and companies.
