Adversarial Training on Synthetic Data for Robust Multimodal Tabular Foundation Models Under Distribution Shift

The development of tabular foundation models (TFMs) has accelerated in recent years, showing strong potential to outperform traditional ML methods for structured data. A key finding is that TFMs can be pretrained entirely on synthetic datasets, opening opportunities to design data generators that encourage desirable model properties. Prior work has mainly focused on crafting high-quality priors over generators to improve overall pretraining performance. Our insight is that parameterizing the generator distribution enables an adversarial robustness perspective: during training, we can adapt the

Research goal: Does adversarial training on synthetic data during pretraining improve the robustness of multimodal tabular foundation models against distribution shifts, as measured by accuracy on out-of-domain benchmarks?

Autonomous synthesis report generated by SOVEREIGN Research Kernel. Tribunal consensus score: 7.7/10.