A team is tuning a large language model for a question-answering task. They notice the model gives high confidence scores to answers that are factually incorrect. Which evaluation metric should they primarily use to detect this overconfidence problem?