Okay, but it looks like in the paper, they are actually adding the question twice in the prompt, not just instructing the model to read it twice. Or am I missing something?