News thumbnail
Science / Thu, 19 Nov 2026 Nature

Quality-controlled synthetic oversampling for imbalanced tabular classification using large language models

Many tabular classification tasks involve imbalanced classes, but synthetic oversampling can produce rows that violate feature constraints or duplicate existing observations. We evaluated QualSynth on eight benchmark datasets with ten seeds against four synthetic oversamplers, including the widely used SMOTE, and class-weighted learning without augmentation. QualSynth achieved the highest F1 score among the synthetic oversamplers on four datasets, while SMOTE had the best average ranks. After correction for multiple comparisons, the only pairwise difference was higher ROC-AUC for QualSynth than for the diffusion model TabDDPM. Class weighting alone had the best overall ranks, and no augmentation strategy improved aggregate ROC-AUC over the unweighted reference.

Many tabular classification tasks involve imbalanced classes, but synthetic oversampling can produce rows that violate feature constraints or duplicate existing observations. We present QualSynth, a training-free framework that prompts a large language model to make sparse edits to real minority-class rows under data-derived feature constraints and a machine-readable output contract. A deterministic validation layer checks each candidate’s structure, values, and duplication before admitting it to the training set. We evaluated QualSynth on eight benchmark datasets with ten seeds against four synthetic oversamplers, including the widely used SMOTE, and class-weighted learning without augmentation. QualSynth achieved the highest F1 score among the synthetic oversamplers on four datasets, while SMOTE had the best average ranks. In equivalence tests, QualSynth and SMOTE differed by less than prespecified margins of 0.03 in F1 score and 0.02 in the area under the receiver operating characteristic curve (ROC-AUC). After correction for multiple comparisons, the only pairwise difference was higher ROC-AUC for QualSynth than for the diffusion model TabDDPM. Class weighting alone had the best overall ranks, and no augmentation strategy improved aggregate ROC-AUC over the unweighted reference. QualSynth is therefore a controlled, inspectable option for minority-class augmentation, with predictive utility comparable to SMOTE’s.

© All Rights Reserved.