Tabular foundation models (TFMs) are emerging as a powerful alternative to conventional tabular learning, with the ability to adapt to new datasets in context and without task-specific training. A new arXiv paper, FTFM, introduces a factorized variant of these models.
The paper argues that current TFM architectures face a trade-off, and proposes factorization as a way to address it. The abstract, however, cuts off before detailing the precise nature of that trade-off or the factorization strategy. The full paper presumably elaborates on both.