Research map / Large language models / Language, safety and interpretability

EPIC: Effective Prompting for Imbalanced-Class Data Synthesis in Tabular Data Classification via Large Language Models

NeurIPS 2024 · Large language models › Language, safety and interpretability · approach: LLM pruning and scaling data · 2 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes EPIC, a training-free prompting method that uses class-balanced, grouped CSV-style example sets with unique variable mapping to make GPT-3.5 (and open LLMs) generate synthetic tabular data that improves classifier performance on imbalanced datasets.

Problem

LLM-based synthetic tabular data generation for imbalanced classification is hard: prompt design choices strongly affect minority-class quality and generation efficiency, and existing generators often fail to capture minority classes.

Approach

Prompt LLMs (in-context, no weight updates) with repeated sets of class-balanced, class-grouped samples in CSV format, with categorical values remapped to unique random alphanumeric strings, and trigger completion by appending the header rather than giving explicit instructions.

Key results

• On Travel, EPIC (GPT-3.5) reaches F1 66.65 and sensitivity 78.00 versus 58.12 and 57.00 for the original data, the best among compared methods.
• EPIC is reported as the only method that consistently improves both F1 and balanced accuracy over original data across all six datasets (gradient boosting results and four-classifier averages).
• Unique variable mapping and grouping raise Sick F1 from 86.60 to 91.95 and, on Sick, raise average output samples per inference from 6.93 to 17.68 with 95% success rate.

Measured effects reported in the paper

Surprising findings

• Generating one class at a time gave a low F1 (58.67) on Travel versus 70.37 for balanced multi-class generation, and CSV-style prompting did not always beat sentence-style (e.g., Travel).
• Class-distinction labels in the prompt improved sensitivity but reduced the average output samples from 17.68 to 9.70.

Limitations

• When the training set is too large to fit in the prompt, only subsets are used as examples, risking incomplete distribution coverage.
• Results on imbalanced binary datasets; sensitivity gains come with specificity losses (e.g., Travel specificity -6.54).

Experimental setup

Datasets: Travel Sick HELOC Income Diabetes Thyroid toy dataset (scikit-learn)

Benchmarks: Travel Sick HELOC Income Diabetes Thyroid

Metrics: F1 score balanced accuracy sensitivity specificity generation efficiency (input tokens, output samples, success Pearson and Cramér's V feature correlation

Baselines: SMOTE SMOTENC TVAE CopulaGAN CTGAN CTAB-GAN+ GReaT TabDDPM CuratedLLM LITO

Builds on: GReaT CuratedLLM LITO TabDDPM

Closest papers by meaning

• Synthetic Tabular Data Generation for Imbalanced Classification: The Surprising Effectiveness of an Overlap Class (AAAI 2025 · similarity 0.88)
• Large Language Models Can Automatically Engineer Features for Few-Shot Tabular Learning (ICML 2024 · similarity 0.87)
• CLIMB: Class-imbalanced Learning Benchmark on Tabular Data (NeurIPS 2025 · similarity 0.87)
• TabICL: A Tabular Foundation Model for In-Context Learning on Large Data (ICML 2025 · similarity 0.86)
• Large Scale Transfer Learning for Tabular Data via Language Modeling (NeurIPS 2024 · similarity 0.86)
• Curated LLM: Synergy of LLMs and Data Curation for tabular augmentation in low-data regimes (ICML 2024 · similarity 0.86)