BIFOLD Colloquium 15/2025

Tabular Foundational Models: An Overview

Abstract. Tabular foundational models (TFMs) have made great progress recently. While they are still limited to small/medium datasets due to the quadratic complexity of self-attention, they now dominate other methods for small/medium datasets. In this talk, I will briefly explain how TFMs work before discussing some of their limitations, in particular their lack of "built-in" symmetries which cause instabilities in high-dimensional classification settings. I will then present an application of TFMs to time-series forecasting, showing the potential impact of these approaches. Finally, I will conclude by presenting TabArena, a recent benchmark developed for tabular models.

Speaker

David Salinas leads a research group at the ELLIS Institute and University of Freiburg, where he is developing high-quality large language models for all official European languages through the OpenEuroLLM and LLMs4EU projects. His work spans the full LLM pipeline—from pre-training and post-training to comprehensive evaluation. His current research interests includes AutoML, tabular models, LLMs, and probabilistic time-series forecasting. He has also contributed to computational topology and geometry processing earlier in his career. Prior to academia, He spent seven years as a Senior Applied Scientist at Amazon and one year at NAVER LABS Europe, where he applied ML research to real-world challenges at scale.