🤯 New AI "Reasons" Like a Human

Singapore startup Sapient has built an AI inspired by the human brain. Instead of the usual chain-of-thought reasoning of language models, researchers have developed a "hierarchical reasoning model" (HRM) that operates like how humans think.

The human brain doesn't just follow a list of steps. It uses two kinds of thinking at once: one that looks at the big picture and plans, and another that dives into the details and works things out quickly. In this principle, HRM also works.

The HRM uses only 1,000 training samples for each task and consists of 27 million parameters, thousands of times less than the flagship models from OpenAI, Google, and other companies. At the same time, this small AI model outperforms leading LLMs in logical tasks and strategic thinking. For example, it quickly solves complex Sudoku puzzles. It finds its way out of most confusing mazes, which even the "smartest" models cannot do.

In the ARC-AGI benchmark, which is considered one of the most difficult exams for testing the reasoning capabilities of AI, the model from Sapient scored 40.3%. For comparison: o3-mini-high achieved 34.5%, Claude Sonnet 3.7—21.2%, and DeepSeek-R1—15.8%.

"These results underscore HRM's potential as a transformative advancement toward universal computation and general-purpose reasoning systems," the model's creators insist.

➡️ The model is available on GitHub.

#news #benchmarks @hiaimediaen