The Abstraction and Reasoning Corpus (ARC) is a benchmark designed to be a fair intelligence test for both humans and AI.

• It is built on "Core Knowledge" priors that humans share (objectness, basic physics, numbers).
• Crucially, every single task is unique. It is impossible for an AI (or a human!) to practice for the specific puzzles.
• It measures the ability to form abstract concepts from just a few examples.