ARC-AGI 2 is a dataset for general tasks, specifically compositional reasoning tasks that require the simultaneous application of rules or multiple interacting rules. It includes a semi-private set for testing remotely-hosted commercial models and a fully-private set for testing self-contained models during the ARC Prize competition. The dataset contains 1000 tasks for training and 120 tasks for evaluation, combining tasks from ARC-AGI-1 and new tasks.
No results indexed yet — be the first to submit a score.
Submit a checkpoint and a reproduction script. We will run it, publish the score, and — if it takes the top — annotate the step on the progress chart with your name.