Researchers have developed a new method for evaluating the performance of generative models in preserving inter-column dependencies in tabular data. The method, called gradient-boosted C2ST, is more effective than previous methods at detecting when these dependencies are not being preserved. In experiments with two types of generators, the researchers found that destroying all inter-column dependency still allowed the generators to produce realistic-looking data, but this came at a cost: minority-class accuracy dropped by 0.38-0.61. The study suggests that current generative models may be relying too heavily on mean-field objectives and not enough on direct dependency supervision. --- Why it matters: This matters because it highlights a limitation in current generative models for tabular data, which can have significant consequences for applications such as data augmentation and transfer learning. Source: https://arxiv.org/abs/2607.21636