RESEARCH

Current research

My experiments with language models on ARC grid puzzles and my plans to study how agents learn through play.

← Back to overview

Why did a perfect grid score zero?

I gave an AI a pattern with 46 missing squares. It filled in every one correctly and still failed the puzzle. When the version I’d trained passed, I put the two answers side by side. They showed exactly the same picture.

See what changed
Puzzle with gaps
What both models drew
Black squares mark the gaps in the original puzzle.
ARC-AGI puzzles by François Chollet · Apache 2.0