Scientists Create A New AI Model

Scientists are quickly making progress on creating AI models that mimic human brain reasoning.

Apparently, there’s a new AI model that can handle advanced reasoning, which sets it apart from well known large language models (LLMs) like ChatGPT.

Scientists say they’re noticing improved performance in important benchmarks.

Scientists at the Singaporean AI firm Sapient have dubbed their latest reasoning AI the hierarchical reasoning model (HRM).

It’s said to be influenced by how the human brain processes information in a hierarchical and multi-timescale manner.This reflects how various regions of the brain combine data over different time spans, from just milliseconds to several minutes.

Scientists say that the new reasoning model outperforms current LLMs and operates more efficiently.This improvement is said to be due to the model requiring fewer parameters and training examples.

They noted that the HRM model has 27 million parameters and utilizes 1,000 training samples.

In AI, parameters are the variables learned during training, like weights and biases.

On the other hand, most top-tier LLMs have billions or even trillions of parameters.

Function

When the HRM was evaluated using the ARC-AGI benchmark, which is recognized as one of the most challenging assessments for determining how close models are to achieving artificial general intelligence, the new model delivered impressive results, as stated in the study.

Many advanced LLMs rely on chain of thought (CoT) reasoning, researchers at Sapient pointed out that this approach has some significant drawbacks, including ‘brittle task decomposition, extensive data needs, and high latency’.

In contrast, HRM utilizes sequential reasoning tasks all at once rather than in a step by step manner.

It consists of two modules: a high-level module for slow and abstract planning and a low-level module for quick and detailed calculations.

This design is inspired by how various parts of the human brain manage planning versus rapid responses.

Additionally, HRM uses a technique called iterative refinement, which means it begins with a rough answer and enhances it through several brief thinking sessions.

After each session, it evaluates whether further refinement is necessary or if the results are satisfactory as the final answer.

According to the researchers, HRM successfully solved Sudoku puzzles that typical LLMs struggle with. The model also performed exceptionally well in navigating mazes, showing that it can tackle structured and logical challenges far better than LLMs.

Although the results are impressive, it’s important to mention that the paper, which has been published in the arXiv database, has not yet undergone peer review.

Nevertheless, the ARC-AGI benchmark team tried to replicate the findings after the model was released as open-source.

The team confirmed the figures. However, they also discovered that the hierarchical structure did not enhance performance as much as claimed.

They concluded that a less documented refinement process during training was likely responsible for the strong results.