Tabular foundation models are trained on large collections of real-world datasets and increasingly served through black-box APIs. That makes it hard for data owners to know whether their private or proprietary information was used in training. A new arXiv paper proposes a dataset inference method to address this gap.

Dataset inference aims to determine whether a specific dataset was part of a model's training data. The paper frames this as a tool for data provenance and privacy auditing, allowing auditors to check black-box models without access to internal weights or training logs. This is particularly relevant for tabular data, which often contains sensitive records.

The approach is described in a single preprint, arXiv:2610.24278, and has not yet been peer-reviewed. The authors position it as a way to bring more transparency to a class of models that is growing in commercial use but remains opaque to outside observers.