Right, with respect to how many RLHF pairs, you can compare it to prior result of ~0%. But what I don't understand, and especially with the hype claims the paper makes about "autonomous super-human reasoning", is why can't they just keep running it and get much higher than 50%? Seems like there's another aspect that is preventing getting higher scores, and makes me wonder if these architectures are really just plateauing.

Don't get me wrong, it's some good work; it's just the language of the paper has some ridiculous hype.