LATABLE: TOWARDS LARGE TABULAR MODELS
Keywords: Tabular data, generative models
TL;DR: LaTable is a novel tabular diffusion model trained across tabular datasets and displays early signs of scaling laws.
Abstract: Tabular data is one of the most ubiquitous data modalities, yet the literature on tabular generative foundation models is lagging behind its text and vision counterparts. Large Tabular Models (LTMs) could revolutionize the way tabular data is used: not as any single dataset analyzed in a vacuum, but contextualized using their metadata and with respect to related datasets. Creating an LTM is difficult, due to the heterogeneous feature spaces of different tabular datasets, metadata, and prior knowledge. In this work, we propose LaTable: a novel tabular diffusion model that addresses these challenges. We show LaTable can be trained across tabular datasets. Through extensive experiments, we find that LaTable displays early signs of scaling laws previously encountered in foundation model regimes. Moreover, LaTable outperform baselines in out-of-distribution few-shot data generation.
Primary Area: foundation or frontier models, including LLMs
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 9883
Loading