Towards Reliable Scaling Laws for Tabular Foundation Models
Scaling laws are widely used to predict foundation model performance and guide compute allocation. However, tabular foundation models (TFMs) differ from conventional foundation models in that they are pretrained on virtually unlimited synthetic data from configurable distributions, while ultimately being evaluated on real-world datasets. We investigate whether scaling laws for TabPFNv2-style models reliably extrapolate to larger compute budgets and generalize across…