Singapore-based AI startup Sapient Intelligence has developed a Hierarchical Reasoning Model (HRM) that matches or even surpasses the performance of large language models on complex reasoning tasks while significantly reducing data and memory requirements. This architecture mimics the dual-system operation of the human brain, working in tandem through a high-level abstract planning module and a low-level fast computation module, avoiding the limitations of chain-of-thought reasoning. On extremely difficult Sudoku and maze problems, HRM achieves near-perfect accuracy with just 1000 training samples, while advanced language models completely fail.
Singapore-based AI startup Sapient Intelligence has developed a new AI architecture capable of competing with and significantly surpassing large language models on complex reasoning tasks, while being smaller in scale and more data-efficient.
This architecture, known as the Hierarchical Reasoning Model (HRM), is inspired by how the human brain utilizes different systems for slow, deliberate planning and fast, intuitive computation. The model achieves impressive results using only a fraction of the data and memory required by today’s large language models. This efficiency is crucial for real-world AI applications in enterprises where data is scarce and computational resources are limited.
Limitations of Chain-of-Thought Reasoning
When faced with complex problems, current large language models primarily rely on chain-of-thought (CoT) prompts, breaking down problems into text-based intermediate steps, essentially forcing the model to “think aloud” while searching for solutions.
While CoT improves the reasoning capabilities of large language models, it has fundamental limitations. Researchers at Sapient Intelligence noted in their paper: “CoT for reasoning is a crutch, not a satisfactory solution. It relies on fragile, human-defined decompositions where a single erroneous step or incorrect order of steps can completely derail the entire reasoning process.”
This reliance on generating explicit language confines the model’s reasoning to the token level, often requiring vast amounts of training data and producing verbose, slow responses. This approach also overlooks the type of “latent reasoning” that occurs internally, which does not need to be explicitly expressed in language.
Researchers pointed out: “A more efficient method is needed to minimize these data requirements.”
Brain-Inspired Hierarchical Approach
To go beyond CoT, researchers explored “latent reasoning,” where the model does not generate “thinking tokens” but reasons within its internal abstract representations. This aligns more closely with human thought processes; as stated in the paper: “The brain maintains long, coherent reasoning chains in latent space with remarkable efficiency, without constantly reverting to language.”
However, implementing this depth of internal reasoning in AI is challenging. Simply stacking more layers in deep learning models often leads to the “vanishing gradient” problem, where learning signals weaken across layers, rendering training ineffective. Another option is recurrent architectures, which can suffer from the “early convergence” problem, where the model prematurely settles on a solution without adequately exploring the problem.
Seeking better methods, the Sapient team turned to neuroscience for solutions. Researchers wrote: “The human brain provides a compelling blueprint for achieving the effective computational depth that contemporary artificial models lack. It hierarchically organizes computation across cortical regions that operate at different time scales, enabling deep, multi-stage reasoning.”
Inspired by this, they designed the HRM with two coupled recurrent modules: a high-level (H) module for slow abstract planning and a low-level (L) module for fast detailed computation. This structure implements a process the team calls “hierarchical convergence.” Intuitively, the fast L module handles part of the problem, executing multiple steps until it reaches a stable local solution. At this point, the slow H module accepts this result, updates its overall strategy, and gives the L module a new, refined subproblem to tackle. This effectively resets the L module, preventing it from getting stuck (early convergence) and allowing the entire system to perform long sequence reasoning steps without suffering from vanishing gradients in a streamlined model architecture.
The paper states: “This process allows HRM to execute a series of different, stable, nested computations, where the H module guides the overall problem-solving strategy, and the L module performs the intensive search or optimization required for each step.” This nested loop design enables the model to reason deeply in its latent space without long CoT prompts or large amounts of data.
A natural question arises as to whether this “latent reasoning” comes at the cost of interpretability. Crown Wang, founder and CEO of Sapient Intelligence, refuted this notion, explaining that the model’s internal processes can be decoded and visualized, similar to how CoT provides a window into the model’s thinking. He also pointed out that CoT itself can be misleading. Wang told VentureBeat: “CoT does not truly reflect the internal reasoning of the model,” citing research showing that models can sometimes arrive at the correct answer through incorrect reasoning steps, and vice versa. “It is essentially still a black box.”
Practical Applications of HRM
To test the model, researchers pitted HRM against benchmarks requiring extensive search and backtracking, such as the Abstraction and Reasoning Corpus (ARC-AGI), extremely difficult Sudoku puzzles, and complex maze-solving tasks.
Results showed that HRM learned to solve problems that even advanced large language models struggle with. For instance, in the “Sudoku – Extreme” and “Maze – Hard” benchmarks, the state-of-the-art CoT models completely failed, achieving an accuracy of 0%. In contrast, HRM reached near-perfect accuracy after training on just 1000 samples for each task.
In tests of abstract reasoning and generalization on the ARC-AGI benchmark, the 27 million parameter HRM scored 40.3%. This surpassed leading CoT-based models, such as the larger o3-mini-high (34.5%) and Claude 3.7 Sonnet (21.2%). This performance, achieved without large pre-trained corpora and under very limited data conditions, highlights the power and efficiency of its architecture.
While solving puzzles demonstrates the model’s capabilities, the real-world significance lies in the different categories of problems. According to Wang, developers should continue to use large language models for language-based or creative tasks, but for “complex or deterministic tasks,” architectures like HRM provide superior performance with fewer hallucinations. He noted the need for sequential problems requiring complex decision-making or long-term planning, especially in delay-sensitive fields such as embodied AI and robotics, or in data-scarce areas like scientific exploration.
In these scenarios, HRM not only solves problems but also learns to solve them better. Wang explained: “In our master-level Sudoku experiments… HRM gradually reduced the number of steps needed as training progressed—similar to how a novice becomes an expert.”
For enterprises, this is where architectural efficiency translates directly to the bottom line. Unlike the serial, token-by-token generation of CoT, HRM’s parallel processing allows for what Wang estimates to be a “100-fold acceleration in task completion time.” This means lower inference latency and the ability to run powerful reasoning on edge devices.
Cost savings are also significant. Wang stated: “For specific complex reasoning tasks, specialized reasoning engines like HRM offer a more promising alternative compared to large, expensive, high-latency API-based models.” To illustrate the efficiency, he noted that training a professional-level Sudoku model takes about two GPU hours, while the complex ARC-AGI benchmark requires 50 to 200 GPU hours—only a small fraction of the resources needed for large-scale foundational models. This opens the door to solving specialized business problems, from logistics optimization to complex system diagnostics, in scenarios where data and budgets are both limited.
Looking ahead, Sapient Intelligence is working to evolve HRM from a specialized problem solver into a more general reasoning module. Wang stated: “We are actively developing brain-inspired models built on HRM,” emphasizing promising initial results in healthcare, climate forecasting, and robotics. He revealed that these next-generation models will be significantly different from today’s text-based systems, particularly by incorporating self-correcting capabilities.
This work suggests that for a class of problems that currently plague AI giants, the way forward may not be larger models, but rather smarter, more structured architectures inspired by the ultimate reasoning engine—the human brain.