Geometric Filtering Boosts LLM-Generated Data for Classification

Large Language Models (LLMs) are increasingly used to generate synthetic training data for tasks like text classification, especially in few-shot scenarios where real data is scarce. However, the quality of this generated data can vary widely; some samples accurately represent their intended class, while others are less relevant or even misaligned. Researchers have introduced a geometric filtering framework designed to address this variability. This method assesses each LLM-generated sample by calculating its Euclidean distance to actual examples within the sentence embedding space. Only those synthetic samples that are geometrically consistent with their target class are selected for training. A soft weighting mechanism then transforms these filter scores into sample weights for the classifier. Extensive evaluations across 13 datasets, 5 classifiers, and 10 augmentation methods demonstrated the framework's effectiveness, achieving a significant performance improvement over traditional methods like SMOTE. The approach also proved robust across various LLMs and generalized successfully to other tasks like named entity recognition, highlighting that even simple distance-based filtering can yield superior results compared to more complex multi-criteria alternatives.