Benchmarking and interpreting a tabular foundation model for PISA 2022 mathematical literacy

Recent advances in machine learning have expanded the analytical toolkit available for large-scale assessment data; however, many high-performing approaches remain limited in terms of interpretability and their suitability for assessment-informed decision-making. This study used data from 7,250 students in the Türkiye sample of PISA 2022 mathematical literacy to benchmark and interpret the Tabular Prior-Data Fitted Network (TabPFN) for classifying student achievement. For the purpose of model evaluation, mathematical literacy was operationalised as the mean of ten plausible values and then converted into a binary outcome using the PISA Level 2 proficiency benchmark, consistent with the study’s classification-oriented aim. TabPFN was compared with eXtreme Gradient Boosting (XGBoost), Light Gradient-Boosting Machine (LightGBM), Stacking, Random Forest, and Logistic Regression under consistent preprocessing and stratified cross-validation procedures. Beyond predictive performance, feature-importance patterns were examined using Shapley Additive Explanations (SHAP) for TabPFN and permutation importance for the baseline models. The findings showed that TabPFN achieved competitive performance and produced an interpretable predictor profile. Home resources and mathematics self-efficacy emerged as the most influential predictors, followed by curiosity and classroom disciplinary climate. These results suggest that tabular foundation models can support large-scale educational assessment not only by providing accurate classifications, but also by generating interpretable evidence about achievement-related learner and contextual factors. The findings should be interpreted as classification evidence around a meaningful proficiency threshold rather than as a model of continuous mathematics achievement.