calibration

Definition

Whether a model's stated confidence matches its actual accuracy: among answers given at seventy percent confidence, roughly seventy percent should be right. Calibration is independent of accuracy — a model can be accurate and badly overconfident, or weak but honest about it — and it is what makes a probability usable for decisions like abstaining or escalating to a human. Preference-based post-training is often reported to degrade it.

Explainers using this term

• Bayes' Theorem in AI — Priors, Posteriors, and UncertaintyOn Calibration of Modern Neural Networks