%0 Report %A Yasenov, Vasil %A Adams-Cohen, Nicholas %T Variable Importance Recovery in Tabular Foundation Models: A Simulation Study %D 2026 %8 2026 Nov %I Institute of Labor Economics (IZA) %C Bonn %7 IZA Discussion Paper %N 19106 %U https://iza.org/publications/dp19106 %X In machine learning, identifying which variables contribute predictive information can be as important as predictive accuracy. Tabular foundation models have matched or exceeded the predictive performance of tuned gradient boosting trees on many small- and medium-sized tabular datasets. Yet their ability to recover population-level variable importance has received little attention. We design and conduct a Monte Carlo simulation study to compare variable importance recovery across tabular foundation models and established machine learning baselines. Our framework distinguishes multiple population estimands, pairs each one with an aligned model-agnostic importance estimator, and evaluates recovery in terms of ranking, support, and relative magnitude. Our results show that tabular foundation models are generally competitive with established machine-learning baselines in recovering population variable importance, and often more accurate. We thus extend the evidence for their capabilities to a previously underexamined dimension, further strengthening their appeal. %K variable importance %K tabular foundation models %K interpretability %K Shapley values %K simulation study %K gradient boosting