Artificial intelligence is moving fast, and scientists have just taken another big step forward. A team at Sapient, a Singapore-based AI company, has created a new model called the Hierarchical Reasoning Model (HRM). Unlike ChatGPT and other large language models (LLMs), HRM is designed to think more like the human brain. Early results show it can outperform popular LLMs in tough reasoning tests, raising new hopes for human-like artificial general intelligence (AGI).
What is the Hierarchical Reasoning Model (HRM)?
The HRM is a type of AI that doesn’t just process information the way most chatbots do. Instead, it mimics the brain’s method of handling tasks at different speeds. Some parts of the brain take time to plan carefully, while others react quickly. HRM uses the same idea, combining both fast and slow processing.
How HRM Differs from Large Language Models (LLMs)
Most LLMs, including ChatGPT, work by breaking down complex problems into smaller steps, often written out in natural language. HRM, on the other hand, doesn’t need to explain its steps. It processes tasks in a single forward pass, making it faster and less resource-heavy.
The Inspiration from the Human Brain
Scientists built HRM based on how different brain regions process information over time. For example, while one part of the brain might focus on long-term planning, another part reacts instantly to sudden changes. This multi-layered approach is what gives HRM its edge.
Why Fewer Parameters Matter in AI
One of HRM’s most impressive features is its size. While LLMs like GPT-5 are rumored to use trillions of parameters, HRM only uses 27 million. Despite being much smaller, it performs better in reasoning tests. This means it needs less computing power and fewer training examples, making it more efficient.
Training Efficiency of HRM
Training an AI usually requires massive datasets. But HRM was trained on just 1,000 examples, proving that smarter architecture can sometimes beat sheer size. This efficiency could make future AI systems more accessible and eco-friendly.
Performance in ARC-AGI Benchmarks
The ARC-AGI test is known as one of the toughest benchmarks for AI reasoning. HRM achieved a score of 40.3% in ARC-AGI-1, outperforming OpenAI’s o3-mini-high (34.5%), Anthropic’s Claude 3.7 (21.2%), and DeepSeek R1 (15.8%). On the harder ARC-AGI-2 test, HRM still came out ahead with 5%, compared to 3% for o3-mini-high and less than 2% for others.
Chain-of-Thought (CoT) Reasoning in LLMs
Most LLMs use a reasoning method called Chain-of-Thought. They break problems into small steps and explain them in text, like a person thinking out loud. This works well for many tasks but can be slow and unreliable.
Limitations of Chain-of-Thought Reasoning
The Sapient researchers pointed out three big problems with CoT: brittle task breakdowns, huge data requirements, and long delays. In other words, CoT can be clumsy, data-hungry, and slow.
How HRM Uses Sequential Reasoning
Instead of relying on CoT, HRM uses a layered reasoning approach. It doesn’t need to write down each step; it just thinks through them internally, making the process faster and less error-prone.
The Two-Module Structure of HRM
HRM has two main parts: one for slow, abstract planning and another for quick, detailed thinking. This combination is what makes it resemble the human brain more closely than current LLMs.
Iterative Refinement Explained
Another secret to HRM’s success is iterative refinement. The model starts with a rough answer and then improves it step by step in short bursts of thinking. After each burst, it decides whether to refine further or give the final answer.
HRM’s Success with Sudoku and Mazes
Tests show that HRM can solve puzzles that stump LLMs. It excelled at Sudoku and maze-solving, both of which require structured reasoning. This suggests HRM is better suited for logic-heavy problems.
Skepticism and Peer Review Concerns
Although the results are exciting, the study has not yet been peer-reviewed. When others tried to reproduce the results, they found that a hidden refinement process during training may have been a bigger factor than the hierarchical design itself.
Comparison with GPT-5 Parameter Size
While HRM uses just 27 million parameters, GPT-5 may have up to 5 trillion. This contrast highlights a major shift: smaller, smarter models could outperform giant ones if designed well.
AI Industry’s Race Toward Human-Like Reasoning
The development of HRM shows how the AI race is moving beyond size and speed. Researchers now want systems that can reason more like humans, which is a step toward AGI.
Developer Perspectives on AI Coding Skills
Meanwhile, in the software world, surveys reveal that many developers believe AI can already code better than humans. More than half of senior developers said LLMs outperform most programmers.
Survey Results: AI Adoption Among Developers
A recent survey of 800 senior developers found that 78% use AI regularly, with almost half using it daily. Many rely on AI for coding, testing, and reviewing code.
Concerns About Data Privacy and Job Loss
Not everyone is thrilled. Around 24% of developers worry about data privacy, while others fear job loss or unreliable AI-generated code.
Will AI Lower or Raise Barriers for New Developers?
Opinions are split. Some say AI makes it easier for beginners to enter the field, while others believe automation will push newcomers out by replacing entry-level jobs.
AI Skills Becoming Essential in Hiring
Nearly 80% of developers believe AI skills will soon be a must-have for new hires. Employers increasingly value candidates who can work with AI effectively.
The Future of AI: Enabler or Replacement?
Most developers see AI as a tool rather than a replacement. They believe its real value lies in changing how tasks are done, not just speeding them up.
What This Means for Artificial General Intelligence (AGI)
If HRM lives up to its promise, it could push us closer to AGI — AI that can think and reason like a human. This raises both excitement and concerns about how society will handle such powerful technology.
Possible Risks and Ethical Considerations
As with any breakthrough, risks remain. Transparency, fairness, and safety must be priorities. Without proper oversight, human-like AI could bring as many problems as benefits.
Conclusion
The Hierarchical Reasoning Model is a bold step in AI development. By mimicking the brain’s layered approach to thinking, it outperforms even the largest LLMs in reasoning tasks, despite being much smaller. For developers, AI tools are already changing the way coding is done, sparking both excitement and concern. The future of AI may not be about building bigger models but about building smarter, brain-like systems that reason more like us.
FAQs
1. What makes HRM different from ChatGPT?
HRM mimics the human brain with layered reasoning, while ChatGPT relies on breaking problems into text-based steps.
2. How big is HRM compared to GPT-5?
HRM uses only 27 million parameters, while GPT-5 is estimated to use up to 5 trillion.
3. Can HRM really outperform large models?
In reasoning benchmarks like ARC-AGI, HRM scored higher than leading LLMs despite its smaller size.
4. What are the risks of human-like AI models?
Key risks include data misuse, job loss, lack of transparency, and ethical concerns about decision-making.
5. Will AI replace developers?
Most experts believe AI will support rather than replace developers, though entry-level jobs may be affected.
Read more blogs: Alitech Blog
Zeeshan Ali Shah is a professional blog writer at AliTech Solutions, and Realancer renowned for crafting engaging and informative content. He holds a degree from the University of Sindh, where he honed his expertise in technology. With a keen eye for detail and a passion for staying up-to-date on the latest tech trends, Zeeshan’s writing provides valuable insights to his readers. His expertise in the tech industry makes him a sought-after writer, and his work at AliTech Solutions has earned him a reputation as a trusted and knowledgeable voice in the field.
Leave a Reply