A model is capable of learning the "calibration" during the reinforcement learning phase, in this "old" post from OpenAI: https://openai.com/index/introducing-s

A model is capable of learning the "calibration" during the reinforcement learning phase, in this "old" post from OpenAI: https://openai.com/index/introducing-simpleqa/ https://openai.com/index/introducing-simpleqa/ you can see the positive correlation between stated confidence and accuracy.