ARC-AGI-3 is actually about action efficiency not solve rate, current LLMs just make a bunch of moves that inefficient, thereby lowering their score