Singapore startup Sapient Intelligence has recently launched an innovative artificial intelligence architecture called "Hierarchical Reasoning Model" (HRM). The model can match existing large language models (LLMs) in complex reasoning tasks, and in some cases even perform better, while requiring significantly less data and having a smaller model size than traditional models. The design of HRM is inspired by the human brain, aiming to achieve efficient reasoning through different thinking systems.
Current LLMs typically rely on the chain-of-thought (CoT) method to solve complex problems, generating a series of text steps for reasoning. Although this approach has improved the reasoning ability of models to some extent, it also has obvious shortcomings. Researchers point out that the chain-of-thought method relies on manually defined steps, and once an error occurs, it may lead to the failure of the entire reasoning process. Therefore, the research team at Sapient Intelligence has proposed a new approach called "latent reasoning," which allows the model to reason within an abstract space internally rather than relying solely on text generation.
HRM consists of two interacting modules: one is a high-level module responsible for slow and abstract planning, and the other is a low-level module performing fast and detailed computations. This hierarchical design enables HRM to perform deep reasoning without relying on large amounts of input data. Test results show that HRM achieves excellent performance in handling high-difficulty tasks such as abstract reasoning and complex Sudoku, demonstrating its strong capabilities in complex tasks.
In addition to accuracy, HRM also excels in reasoning speed. According to Wang Guan, the founder of Sapient Intelligence, HRM can achieve a 100-fold acceleration in task completion time when performing specific complex reasoning tasks. This means HRM can quickly perform powerful reasoning calculations on edge devices, significantly reducing enterprise time and costs.
Looking ahead, Sapient Intelligence is working to develop HRM into a more general reasoning solution, planning to apply it in multiple fields such as healthcare, climate prediction, and robotics. This development marks that the success of future AI may not lie in simply scaling up model size, but in drawing inspiration from the structure of the human brain to develop smarter and more efficient reasoning architectures.
Key Points:
🚀 HRM surpasses large language models in complex reasoning tasks with a new hierarchical architecture and requires less data.
🔍 HRM's model combines high-level and low-level modules, improving reasoning speed and efficiency.
💼 Future AI may draw more inspiration from the design of the human brain, rather than simply relying on scaling up.
