Scientists developed a new AI modeled on the human brain's hierarchical processing, outperforming top LLMs on reasoning tasks.
The "hierarchical reasoning model" (HRM) uses multi-timescale processing, mimicking how different brain regions integrate information.
HRM is far more efficient, using only 27 million parameters vs. billions/trillions in models like GPT-5, and needs fewer training examples.
It excelled on the tough ARC-AGI benchmark, scoring 40.3% vs. OpenAI's 34.5% and Claude's 21.2%, a key test for AGI.
Unlike LLMs that use chain-of-thought, HRM performs sequential reasoning in a single, fast forward pass without intermediate steps.
It has two modules: one for slow, abstract planning and another for rapid, detailed computations, just like the human brain.
The model uses "iterative refinement," thinking in short bursts to repeatedly improve its answer before deciding it's final.
It solved complex Sudoku puzzles and optimal path-finding in mazes, tasks conventional LLMs could not accomplish.
The paper is not yet peer-reviewed. The ARC-AGI organizers replicated the score but found a different driver for the performance gain.
Their findings suggest an under-documented refinement process during training, not the hierarchy, may be the key to its success.