This artificial intelligence surpasses current models by being 100 times faster

The major contemporary language models (LLM) approach the complex problems mainly through the technique of the “chain of thought” (Chain-of-Thought. Therefore, COT). Therefore, This method consists in breaking down a problem into intermediate steps in the form of text. Furthermore, forcing the model to verbalize its process of reflection via words, while there would be a much faster method.

Although this approach has improved LLM reasoning capacities, it has certain limits. Moreover, In their publication, the researchers of Sapient Intelligence qualify the COT method of “crutch” rather than satisfactory solution. However, They claim that this technique is based on fragile decompositions defined by man. Nevertheless, where only one error or a bad ordering of the steps can make the entire reasoning process fail.

This dependence on the generation of explicit language links the reasoning of the model at the level of “tokens”, which often requires large quantities of training data and produces long and slow responses, even if we have the impression that our current AIs are already very fast, we can do up to 100 times better according to these researchers.

A hierarchical approach inspired by the human brain – This artificial intelligence surpasses current

To go beyond the limits of current AI. Furthermore, researchers have explored a new concept of reasoning that aligns more about human thought. In addition, As the study notes. In addition, “The brain maintains long and consistent reasoning chains with remarkable effectiveness in a latent space, without constant translation towards language”.

The implementation of such deep reasoning in AI is complex. For example, Recurrent. Therefore, alternative architectures can suffer from “early convergence” this artificial intelligence surpasses current (early convergence), where the model adopts a solution too quickly, sometimes to please the user, but where certain responses are not convincing. Who has never issued a chatgpt query with a response from the latter completely next to the plate. just to answer the question?

The Sapient team turned to neuroscience to design their hierarchical reasoning model (HRM). The model has two coupled recurring modules:

• A high level module (H) for slow and abstract planning.
• A low level (L) module for quick and detailed calculations.

This structure allows a process that the team calls “hierarchical convergence”. The rapid module deals with part of the problem until reaching a stable local solution. Then. the slow H module incorporates this result, updates its global strategy and assigns a new sub-probler refined to the L module. This mechanism resets the L module. preventing it from blocking itself and allowing the global system to execute a long sequence of reasoning stages with a light architecture which does not suffer from the problems of early convergence.

Regarding interpretability, Guan Wang, founder and CEO of Sapient Intelligence, explains that the internal HRM processes can be decoded and visualized. He adds that COT transparency can be misleading. citing studies showing that models can sometimes provide correct answers with erroneous reasoning, and vice versa.

This artificial intelligence surpasses current

HRM performance in action

To assess their model. the researchers tested the HRM on test benches requiring significant search and back-back capacities, such as the Arc-Agi corpus (Abstraction and Reasoning Corpus), sudoku grids of extreme difficulty and complex labyrinths.

The results indicate that the HRM can solve problems that are out of reach for advanced LLMs. even the next version of GPT 5. On the tests “sudoku-extreme” and “Maze-Hard”, the cutting-edge models based on the COT obtained a score of 0% success. On the other hand, the this artificial intelligence surpasses current HRM has reached an almost perfect precision after training on only 1000 examples for each task.

On the Arc-Agi test bench, which assesses abstract reasoning, the HRM of 27 million parameters obtained a score of 40.3%. This performance exceeds that of much larger COT models like O3-Mini-High (34.5%) and Claude 3.7 Sonnet (21.2%).

An upcoming change for language models?

According to this artificial intelligence surpasses current Guan Wang. if the LLM remain relevant for linguistic or creative tasks, an HRM type architecture offers higher performance for “Complex or deterministic tasks”especially “Sequential problems requiring complex decision -making or long -term planning”. Domains such as robotics. on -board AI or scientific exploration, where the data is rare and the latency is critical, are cases of relevant use, more than conversational AI.

The efficiency of architecture results in concrete advantages for companies. The parallel treatment of the HRM could allow a “100 times reduction in the accomplishment time of tasks” Compared. to current AI. This means a lower inference latency and the ability to operate powerful reasoning systems on devices on the periphery.

Cost savings are also notable. Model training for tasks such as sudoku resolution at a professional level requires about two hours of GPU. and for the Arc-Agi test, between 50 and 200 hours this artificial intelligence surpasses current of GPU. It is a fraction of the resources necessary for the major models that we know today.

SAPIENT Intelligence is already working to develop HRM to a more generalist reasoning module. with preliminary results in the fields of health, climate forecasting and robotics. These future models will also incorporate self-correction features, in order to strongly limit errors.

Want to save even more? Discover our selected promo codes for you.