Skip to content
BENIBIX Learn

Abstraction and Reasoning Corpus (ARC-AGI)

AI📏 Evaluation & BenchmarksBenchmarks and Leaderboards

Also known as ARC-AGI, ARC-AGI-2, ARC Prize

Grid puzzles. Each one asks an AI to learn a new rule from just a few pairs.

Think of it like

It is like a puzzle book where each page has a new kind of puzzle. You must spot the trick from a few tries.

Example

One task shows three pairs of grids where every blue square turns red. The solver must apply that rule to a fourth grid, with every cell right.

Why it matters

Most tests reward what a model already knows. This one tries to check how well it learns something new.

Definition

ARC-AGI is a benchmark of puzzles made of small colored grids. Each task shows a few example pairs, an input grid and the output it turns into. The solver must find the hidden rule and build the output for a new grid. Every task is new, so memory does not help, only spotting the rule. François Chollet made it in 2019. Harder versions followed, ARC-AGI-2 in 2025 and ARC-AGI-3, where the puzzles are games.
Two examples show the rule. The solver must fill the empty grid with every cell right.

Common mistake

Thinking a high score means a model is as smart as a person. It tests one skill, learning a rule from a few pairs.

Key takeaways

  1. 1
    Few examples. Find the rule from a few pairs.
  2. 2
    Exact answer. Every cell must match.
  3. 3
    Easy for people. Hard for AI, by design.