LLM Augmentation Underperforms Embedding-Space SMOTE for Imbalanced Data
August 14, 2026
A benchmark of 11 augmentation methods shows that LLM-based generative augmentation is often statistically equivalent or inferior to classical embedding-space SMOTE-style retrieval for imbalanced text classification. The study evaluated performance across seven datasets with imbalance ratios up to 500.
HOW THIS AFFECTS YOU
●
builderYou may achieve better results for imbalanced datasets by using embedding-space retrieval rather than generative LLM augmentation.
●
researcherThis benchmark provides a baseline for comparing generative versus classical data augmentation.