Logit margin

Logit margin refers to the difference between the logit values of two candidate answers in a model's output layer; in multiple-choice questions it is typically computed as the difference between the correct answer and the strongest competing option, providing a continuous measure of the model's preference strength for an answer.