ARC-AGI 1 and 2 are spatial reasoning benchmarks. ARC-AGI 3 is advanced spatial reasoning with agentic flavor. They're adversarial benchmarks - they intentiona

ARC-AGI 1 and 2 are spatial reasoning benchmarks. ARC-AGI 3 is advanced spatial reasoning with agentic flavor. They're adversarial benchmarks - they intentionally hit the weak point of existing LLMs. Not "AGI complete" by any means. But not useless either.