An article from LiveScience titled “Scientists just developed a new AI modeled on the human brain — it’s outperforming LLMs like ChatGPT at reasoning tasks” caught my attention because my first reaction was “well, that makes sense, I suppose.” But on a second read I realized the author buried the most interesting finding in an underwhelming last paragraph that includes the link to the source post on the ARCPrize blog.
Briefly, a 27-million (with an “M”) parameter model called the “Hierarchical Reasoning Model” achieved impressive results on the ARC-AGI-1 evaluation suite and the team behind it claimed it was because of a novel approach that mimics the way the human brain approaches problem solving.
Because the team open-sourced the model, the ARC Prize team was able to reproduce the result and dig into the architecture. What they found is that the hierarchical structure wasn’t the driver of success. What did matter was the Outer Refinement Loop — HRM doesn’t just output one answer; it takes its first attempt, feeds it back into itself, asks “How can I make this better?”, then repeats. Disable that loop and the model’s performance drops. There’s also Task Augmentation: the model was trained on thousands of variations of a few puzzles. By seeing the same logic applied to different colors or orientations, it got really good at memorizing how to solve those specific problem types — but when presented with tests it hadn’t been trained on, it performed notably worse. Then there’s Adaptive Halting (ACT), a “halting head” that acts as a built-in judge, deciding when the answer is good enough and stopping.
In the end, the architecture mattered less than the iterative attempts at a good solution to a problem the model was well-versed in. Still, even if it struggled with genuinely novel ideas, it proved that a tiny model using a refinement loop can be smarter and cheaper than a giant model taking one pass.