Prediction--Powered Active Testing
Kianoosh Ashouritaklimi ⋅ Valentin Kilian ⋅ Tom Rainforth ⋅ Francois Caron
Abstract
Active testing estimates model risk from a large unlabelled test pool under a limited labelling budget. We propose Prediction--Powered Active Testing (PPAT), which combines the LURE estimator with a control variate built from cheap black--box predictions on the full pool, yielding an unbiased estimator with reduced variance. On real regression benchmarks, PPAT remains unbiased and consistently achieves lower variance than both LURE and random sampling.
Chat is not available.
Successful Page Load