Tabular Foundation Models Versus Established Tabular Learners: A Regime-Stratified Critical Review

Tabular foundation models (TFMs) are pretrained once across many tables and applied to a new table by conditioning on its labelled rows, with no training on that table. They challenge gradient-boosted decision trees, the long-standing default for tabular prediction, but the published comparisons are hard to interpret, because studies differ in data regime, split design, baseline tuning, and metric. This review assesses the evidence regime by regime. From 76 peer-reviewed works selected under a documented protocol, it defines TFMs by reusable cross-table pretraining, organises them by pretraining data, context mechanism, scale, modality, and task interface, and compares them with tree ensembles, deep tabular networks, and tabular-to-image methods. Seven recurring claims are graded on peer review, evaluator independence, split match, baseline strength, and replication, with all scores published. One claim is strongly supported: on small-to-medium independent and identically distributed tables, TFMs match or exceed tuned tree ensembles without per-dataset tuning. Tuned tree-based and deep models remain as strong or stronger on temporal, feature-engineered data; above 10,000 rows the evidence is developer-reported except for one independent clinical benchmark; and no eligible study tests a frozen TFM under a grouped split or beyond its feature envelope. Tabular-to-image methods win only on small or very high-dimensional tables. Among the comparison-bearing studies, about one in ten use a deployment-matched split, none of them an independent evaluation of a TFM, and more than half of those evaluating a pretrained model report no contamination check. A reporting checklist, an evidence ladder, and seven research directions turn these findings into practice.