ARC-AGI

Why we included this project

ARC-AGI is a set of 800 small grid puzzles that reward genuine reasoning rather than pattern matching. Each task shows a few demonstration input/output pairs, and the solver has to infer the underlying rule well enough to produce the exact output grid for unseen test inputs, with three tries allowed. The data ships as plain JSON in clean training and evaluation splits, so wiring it into any solver is simple, and a browser interface lets humans attempt tasks by hand. Because scoring demands an exact grid on first exposure, it is a serious test for program synthesis and neuro-symbolic systems. Teams building reasoning models will find it useful both as a scoring target and as a steady supply of small, tricky examples for studying generalization and inductive bias.