Distributional Regression with Tabular Foundation Models: Evaluating Probabilistic Predictions via Proper Scoring Rules
This study investigates how choosing a specific scoring rule—a way to measure prediction quality—affects tabular foundation models like TabPFN and TabICL. Since standard benchmarks often ignore full distribution accuracy, the paper argues that using proper scoring rules for training and evaluation can induce different learning behaviors. This choice of score fundamentally alters the distribution a model learns.