It seems like they self correct when you explain to the LLM something is incorrect. But in reality they take the original question, plus the “incorrect” statem

It seems like they self correct when you explain to the LLM something is incorrect. But in reality they take the original question, plus the “incorrect” statement and get the most likely responses from the model