Prediction-Powered Inference

Prediction-Powered Inference (PPI) is a statistical framework developed by Angelopoulos and Bates that enables valid confidence intervals and calibrated predictions using a mix of large-scale model predictions (on unlabeled data) and small labeled datasets.

The key insight is that a small labeled sample can be used to correct and calibrate predictions from a model applied to a much larger unlabeled dataset, yielding valid statistical inference with minimal labeling cost.

Applications in misinformation detection

NN-PPI (Nearest Neighbor PPI), proposed by Amatya et al. (2026), extends PPI to pointwise claim check-worthiness predictions by applying local residual corrections based on semantically similar labeled examples.

Related topics

• Model calibration and confidence — PPI is fundamentally a calibration technique
• Uncertainty quantification — PPI provides uncertainty quantification via confidence intervals