What are ARC Puzzles?

The Abstraction and Reasoning Corpus (ARC) is a benchmark created by François Chollet to measure general intelligence in AI systems. Each puzzle consists of colored grids where the solver must discover a hidden transformation rule from a few training examples, then apply it to a new test input.

These puzzles test fluid intelligence — the ability to reason about novel problems without relying on memorized patterns — and remain one of the hardest challenges for AI.

What Makes MARC Different?

MARC (Metaphor-Augmented Reasoning for Cognition) extends ARC by adding a metaphorical hint to each puzzle. Instead of discovering the rule purely from visual patterns, solvers get a natural language metaphor like "The floor is lava" or "Robin Hood" that hints at the transformation.

This approach tests the ability to bridge abstract language with visual pattern recognition — much like how humans use figurative language to understand complex ideas. Each puzzle is human-created, encoding a creative metaphor as a grid transformation, ranging from intuitive to deeply abstract.

Creating Good Puzzles

• Start with a metaphor. Begin with an evocative concept, then build a grid transformation that reflects it.
• Keep it solvable. Ensure that a person who understands the metaphor can figure out the rule from a few examples.
• Be consistent. The rule should apply the same way across all training and test pairs.
• Use the colors meaningfully. Assign color significance that aligns with your concept.
• Include 2–4 training pairs and at least 1 test pair.
• Vary your grids so that solvers demonstrate genuine understanding rather than pattern memorization.