Evaluation
Model confidence estimation
ActiveUpdated: 22 September 2026Published
Key
innovation
Quantifying trust in a model's answer, not just the answer itself.
Category
Evaluation
Abstraction level
Pattern
Operation level
InferenceEvaluation (runtime)
Use cases
Hallucination detectionSelective prediction / abstentionRouting to verification or a human
How it works
Confidence is estimated from token probabilities, entropy, cross-sample agreement, or the model's verbalized confidence.
Problem solved
Models often answer confidently even when wrong (poor calibration, hallucinations).