Statistical Attribute Alignment for Black-Box Generative AI

Abstract: Generative AI offers a powerful way to create synthetic data for simulation, training, and evaluation. But what if a model generates outputs with an attribute distribution that differs from what a downstream task requires? For example, medical research may require image collections with specified distributions of demographic groups or disease categories. This talk develops a statistical framework for attribute alignment through output post-processing, using only query access to the generative model. After observing the attributes of generated outputs, we select a batch whose attributes jointly follow a target distribution, exactly or approximately. We develop sampling algorithms and establish their asymptotic optimality in expected model query cost as the desired batch size grows.