In a breakthrough that could reshape the future of artificial intelligence, scientists have unveiled a new brain-inspired AI model that is outperforming leading large language models (LLMs) like ChatGPT in reasoning tasks. The system, called the Hierarchical Reasoning Model (HRM), mimics the way the human brain processes information — and it has already surpassed well-known AI models in some of the toughest benchmarks for reasoning.
A Smarter Way of Thinking for AI
Unlike most LLMs that rely heavily on chain-of-thought (CoT) reasoning — breaking down problems into smaller steps expressed in natural language — HRM takes a different approach. Developed by researchers at Sapient, an AI company in Singapore, the model uses a two-module system: one for abstract, high-level planning and another for fast, detailed computations. This structure closely resembles how different regions of the brain work together across multiple timescales.
The researchers explain that this architecture allows HRM to execute reasoning in a single forward pass, without the need for extensive supervision or huge datasets. Remarkably, HRM has only 27 million parameters and was trained on just 1,000 samples. In comparison, advanced models like GPT-5 are estimated to use 3–5 trillion parameters.
Outperforming Leading AI Models
The model was tested on the notoriously difficult ARC-AGI benchmark, designed to measure how close AI is to achieving artificial general intelligence (AGI). HRM delivered groundbreaking results:
- ARC-AGI-1 Test: HRM scored 3%, beating OpenAI’s o3-mini-high (34.5%), Anthropic’s Claude 3.7 (21.2%), and Deepseek R1 (15.8%).
- ARC-AGI-2 Test: HRM achieved 5%, outperforming o3-mini-high (3%), Claude 3.7 (9%), and Deepseek R1 (1.3%).
Beyond benchmarks, HRM showed near-perfect accuracy in solving complex Sudoku puzzles and in optimal pathfinding for mazes — areas where conventional LLMs usually struggle.
Why HRM Could Be a Game-Changer
The Sapient team argues that current LLMs face key limitations, such as “brittle task decomposition, heavy data requirements, and high latency.” HRM addresses these issues through a technique called iterative refinement, where the model refines answers in short bursts of reasoning before finalizing a solution. This method is not only more efficient but also demonstrates human-like adaptability in solving complex problems.
While the research is still awaiting peer review, the AI community is already paying attention. The team has open-sourced the model on GitHub, and independent researchers confirmed the benchmark results. Interestingly, they also noted that much of HRM’s strength came from its refinement process during training — a detail that was not fully highlighted in the original paper.
Outlook
If further validated, HRM could signal a major shift in AI development — moving from massive, resource-heavy models to smaller, more brain-like architectures that prioritize reasoning efficiency over scale. As scientists continue refining this model, HRM could pave the way for AI systems that think more like humans and potentially bring us closer to the dream of true artificial general intelligence.
Read More: Scientists Develop Brain-Inspired AI That Outperforms ChatGPT in Reasoning Tasks