Gael Varoquaux • 7/9/2025

TabICL: Pretraining the best tabular learner

This article details TabICL, a novel tabular foundation model presented at ICML 2025. It explains how the model frames tabular learning as a table completion problem, uses sophisticated data generation to create a rich prior, and employs an improved transformer architecture that accounts for the 2D structure of tables. The result is a powerful, easy-to-use, and efficient tabular learner that works well even without a GPU.

0 comments

#Transformer #Pretraining #Tabular Learning