ARC-AGI-2, developed by the ARC Prize Foundation, is the second iteration of the Abstraction and Reasoning Corpus. Like its predecessor, it presents systems with grid-based input-output demonstrations and asks them to infer the underlying transformation rule and apply it to a novel test input, allowing two attempts per test case (pass@2). The benchmark comprises 1,360 tasks in total – 1,000 training tasks and 360 evaluation tasks split evenly across public, semi-private, and private sets of 120 each.

ARC-AGI-2 is designed to be substantially harder for AI systems while remaining accessible to humans. Tasks that were susceptible to brute-force program search in the original ARC-AGI have been removed, and new tasks target weaknesses in symbolic interpretation, compositional reasoning, and contextual rule application – areas where current systems struggle to assign meaning beyond surface-level visual patterns or to apply multiple interacting rules simultaneously. Controlled human testing with over 400 participants confirmed that every evaluation task can be solved by at least two people in two attempts or fewer, with an average human score of 60%. By contrast, pure LLMs score 0% and frontier reasoning systems achieve only single-digit percentages.