Valid Feature-Level Inference for Tabular Foundation Models via the Conditional Randomization Test

A new research paper by Mohamed Salem introduces a practical method for performing statistically valid feature-level inference on complex machine learning models. The approach marries the Conditional Randomization Test (CRT), a statistical technique, with TabPFN, a specialized foundation model designed for tabular data. This combination directly addresses a critical gap in modern AI: while models like deep neural networks or gradient-boosted trees achieve high predictive accuracy, they are notoriously opaque and rarely provide reliable measures of statistical significance for individual input features.

The resulting procedure generates finite-sample valid p-values that assess whether a specific feature contains meaningful information about a target variable, conditioned on all other features. Crucially, it maintains validity—meaning it controls false positive rates—even in challenging scenarios with nonlinear relationships and high correlation between features. The method is model-agnostic, requiring no retraining of the underlying black-box predictor and making no restrictive parametric assumptions about the data's distribution. This provides data scientists and researchers with a rigorous tool for explainability and feature selection that was previously difficult or impossible to obtain from high-performance, complex models.