Geometric Filtering of LLM-Generated Samples for Few-Shot Text Classification

2026-08-13

AI Summary

This study addresses the quality heterogeneity of LLM-generated samples by proposing a geometric filtering framework. The method leverages Euclidean distance in embedding space to select geometrically consistent samples and incorporates a soft-weighting mechanism for classifier training, demonstrating that simple distance metrics outperform complex multi-criteria strategies. Experiments across 13 text classification datasets show an average improvement of 2.61 percentage points, significantly surpassing SMOTE. Furthermore, the framework generalizes seamlessly to named entity recognition tasks without modification, achieving a 9.26 percentage point gain and exhibiting strong cross-model robustness. These findings establish geometric filtering as an efficient paradigm for low-resource data augmentation, highlighting the efficacy of geometric consistency over elaborate selection heuristics in enhancing synthetic data utility.

Abstract

Large language models (LLMs) can generate synthetic training data for text classification, but the quality of generated samples is heterogeneous: some fall in correct class regions of the embedding space while others land in peripheral or cross-class zones. We propose a geometric filtering framework that evaluates each LLM-generated sample by its Euclidean distance to real class examples in a sentence embedding space, selecting only geometrically consistent candidates. A soft weighting mechanism transforms filter scores into sample weights for classifier training. Evaluated across 13 datasets, 5 classifiers, 10 augmentation methods, and over 6,700 configurations, our method achieves +2.61 percentage points (pp) over SMOTE ($p<0.0001$, Cohen's $d=0.95$, 88.9% win rate). The approach generalizes to named entity recognition (+9.26pp, 100% win rate) without filter modification, and is robust across 5 LLMs from 4 providers. A key finding is that the simplest distance-based filter consistently outperforms complex multi-criteria alternatives.

Problem

• LLM-generated samples
• Few-Shot Text Classification
• Synthetic Data Quality
• Embedding Space

Innovation

• Geometric Filtering
• Sentence Embedding Space
• Soft Weighting Mechanism
• Synthetic Data Quality
• Few-Shot Text Classification