Distribution Shift Detection

Distribution shift detection studies whether incoming data, model inputs, or model behavior differ from a reference, and whether those changes threaten reliable predictions. Methods range from statistical tests such as Kolmogorov–Smirnov, maximum mean discrepancy, and change-point scans to representation-based OOD scores, conformal tests, and online betting martingales. Work spans image and language OOD benchmarks, time-series and tabular streams, and deployed systems, with growing attention to locating affected subgroups and triggering adaptation only when needed. A central challenge is balancing sensitivity and detection speed against false alarms, especially under high-dimensional shifts, changing data sources, and shifts that are detectable but harmless—or harmful yet difficult to detect.