Teaching Machines to Think in Loops – Hierarchical Reasoning Models [HRM]

For years, the race in artificial intelligence has been about scale: bigger models, more data, and ever-larger computing power. But a new approach, called Hierarchical Reasoning Models (HRMs), suggests the future might not be about size at all — but about structure.

The Human-Inspired Twist

Instead of building massive “one-size-fits-all” systems, HRMs take inspiration from how we humans think. Imagine your brain working on a tough puzzle: part of it sets a big-picture plan (“first fill in the corners”), while another part quickly tests and tweaks details (“this number doesn’t fit, try another”). HRMs mimic this dynamic with two modules:

• The strategist (H-Module): slow, deliberate, focused on the bigger plan.
• The worker (L-Module): fast, iterative, focused on details.

Together, they loop — refine, adjust, refine again — until the solution stabilizes.

The following graphic visualizes this architectural approach.
Source: Compare https://www.youtube.com/watch?v=RK7lysjz_G0&list=LL&index=2

Why It Matters

This structure enables HRMs to solve reasoning tasks — like Sudoku, mazes, and abstract reasoning tests — with a fraction of the data and computing power of today’s giants. While large models require billions of examples, HRMs have shown impressive results with just 1,000 training puzzles.

The following graphic compares the performance (i.e. accuracy) on challenging Sudoku problems as as the model depth increases. It compare the HRM with Transformer models. Source: Compare https://www.youtube.com/watch?v=RK7lysjz_G0&list=LL&index=2

Beyond “Bigger is Better”

The lesson? Breakthroughs in AI may come not from feeding models the whole internet, but from teaching them to think more like us: hierarchically, strategically, and iteratively.

In the future, we might look back on 2025 as the year AI stopped growing “upwards” and started growing “smarter.”

Learn more about HRM?

Want to learn more about HRM and get a deep-dive? I’d recommend the following YT-Videos:

Video #1: https://www.youtube.com/watch?v=RK7lysjz_G0&list=LL&index=2

Video #2: https://www.youtube.com/watch?v=QWD55guu0So&t=534s