Prediction-Powered Inference

Prediction-Powered Inference combines abundant model predictions or unlabeled data with a smaller set of trusted labels to estimate quantities while correcting for prediction errors. Common approaches include calibration, residual-based control variates, cross-fitting, active label sampling, and combining multiple predictors. These methods support estimates and tests for means, risks, causal effects, rankings, and other quantities, including in AI evaluation and dependent time-series settings. The central tension is that noisy predictions can improve precision, but gains depend on their relationship to true outcomes and methods must preserve valid uncertainty estimates under finite samples, model error, or domain shift.