They don't. R1 gets the right answer in the thinking part of the response and ignores it in the response more than half the time in my tests.