Asking LLMs How Confident They Are

A verbal confidence score is an answer the model generates in natural language. Asking for a percentage does not by itself create a measurement process that knows whether the answer is correct. The score can still correlate with correctness on a particular task, but a team must test that relationship. Calibration asks whether predictions assigned a confidence level are correct at a corresponding rate over a relevant set of cases; it is assessed across examples, not by inspecting one answer.

Research findings are mixed because results depend on models, prompts, and evaluation data. A 2023 study of human-feedback-tuned models found verbalized confidence better calibrated than token probabilities on TriviaQA, SciQ, and TruthfulQA in its experiments. A 2024 evaluation using different language and vision-language models and difficult uncertainty tasks found high calibration error and overconfidence in its tested settings. These findings are not contradictory rules for every modern model: they show why the confidence signal needs evaluation on the intended task and current model version.

Before using a score to escalate or approve decisions, collect labeled examples that resemble production traffic. Compare confidence buckets with observed accuracy, choose a threshold against the costs of wrong answers and unnecessary review, and monitor drift. Expected Calibration Error is one summary of mismatch across bins, but it can conceal uneven performance across subgroups or tasks. Token log probabilities describe likelihoods of generated tokens, not a direct probability that the full answer is correct. Repeated-answer agreement can add evidence about stability, but repeated samples may share the same misconception. Treat verbal confidence as one input to a validated decision process, not a substitute for external evidence or expert review.