They could have rlhfed or finetuned on user thumbs up responses, which could include users who took the test and asked it to explain problems after