Zero-Flow Two-Sample Tests: Distribution Comparison Built on Flow Models

arXiv 2607.21542·medium signal

Yakun Wang, Leyang Wang and Song Liu propose a new two-sample test for deciding whether two sample sets come from the same distribution, built on a zero-flow construction. Two-sample testing is the underlying primitive for production drift detection — deciding whether today's input distribution has moved from the training distribution. A test with better power on high-dimensional data is directly usable in ML monitoring, where standard nonparametric tests degrade badly as dimension grows.