Hierarchical Reasoning Model Puts Traditional LLMs to the Test

One such brain-inspired AI architecture, the Hierarchical Reasoning Model (HRM), created by researchers at Sapient in Singapore, is proving to be a formidable competitor to traditional large language models (LLMs) deployed by leaders such as OpenAI and Anthropic. However, HRM has performed better with more advanced reasoning tests even though they are smaller and trained with less data.

Dual-Module Brain-Inspired Design.

HRM replicates the two-layered thought process of the brain with two modules:

• High level module - deals with slow abstract planning.
• Low level module - does fast detailed calculations.

HRM uses iterative refinement as opposed to chain-of-thought (CoT) reasoning found in LLMs, which solve problems in a series of steps, using short bursts of reasoning to continuously refine solutions.

Weakness in Chain-of-Thought Reasoning.

The traditional CoT reasoning is capable but relies on large problems and is characterized by inflexible problem decompositions. It is also latent due to processing steps. The way HRM escapes these traps is by dynamically refining answers rather than use a fixed chain.

Good AGI Benchmark Performance.

In the ARC-AGI benchmark, which is a gold standard in testing general intelligence, HRM scored:

• 40.3 on ARC-AGI-1, beating OpenAI on o3-mini-high (34.5) and Claude 3.7 (21.2) and DeepSeek R1 (15.8) on the same metric.
• 5% on ARC-AGI-2, better than OpenAI (3%), DeepSeek (1.3%), and Claude (0.9%).

Complex Task Performance.

HRM was able to solve Sudoku puzzles and maze navigation, which involves abstract thinking and high precision. These findings demonstrate its capacity to transcend what regular LLMs are capable of doing.

Hidden Factor in Its Success

In independent reviews, it was found that the positive performance of HRM was not necessarily because of its hierarchical architecture, but also because of an under-documented process of refining training, implying that training practices are as important as design.

Why HRM Stands Out

HRM is not based on a linear task decomposition as ChatGPT or Claude. Its iterative refining structure coupled with the two-module system makes it more versatile and practical especially in areas where planning and detailed execution are required to cooperate.