Testing whether two conditional distributions are equal is a core problem in causal inference, model diagnostics, and transfer learning. A known weakness of existing methods is that their power drops when the conditioning variable is high-dimensional.

A new arXiv preprint (2610.20447) addresses this weakness by turning to generative models. The authors propose a testing procedure built on these models, aiming to keep the test sensitive even when the conditioning set has many dimensions.

The abstract does not disclose the full method or experimental results, so the announcement alone leaves the test's exact construction and performance open. Still, the direction is notable: generative models, already widely used for sampling and density estimation, are being repurposed for a fundamental statistical decision problem.