A 10 million parameter model just beat o3-mini-high on ARC-AGI-2. By almost 4x.

A tiny research model that would fit on a decent laptop just outperformed one of the frontier reasoning models from OpenAI on a hard reasoning benchmark.

That's the kind of thing that makes me drop what I'm doing and actually go read the paper.

The paper is called Generative Recursive Reasoning. It's from KAIST, Mila, and NYU.

Let me set this up properly, because to understand why this paper matters, you need to know where AI reasoning actually is right now in 2026.

Over the last two years, the way frontier models got smarter wasn't really about raw size. It was about something called chain of thought. You give the model more tokens to think with. You let it spend more compute at inference time. You scale up the THINKING, not just the model.