“The ARC-AGI V2 benchmark was saturated through a reinforcement learning loop where models generate new tasks, solve them, verify the solutions, and are then fine-tuned on successful reasoning chains.”