Research map / Label-efficient and robust learning / Recognition and long-tail
ReTabSyn: Realistic Tabular Data Synthesis via Reinforcement Learning
Xiaofeng Lin, Seungbae Kim, Zhuoya Li, Zachary DeSoto, Charles Fleming, Guang Cheng
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| fine-tune GReaT generator with DPO on oracle-free perturbation pairs | GReaT (GPT-2) · Adult 1% prevalence | PR-AUC: 0.873 -> 0.906 | Improved |
| add DPO fine-tuning after augmented pre-training (ablation) | ReTabSyn · Wilt | AUROC: 0.815 -> 0.841 | Improved |
| add augmented pre-training before DPO (ablation) | ReTabSyn · Wilt | AUROC: 0.800 -> 0.841 | Improved |
| use oracle-free perturbation instead of classifier-based editing (ablation) | ReTabSyn · Wilt | AUROC: 0.809 -> 0.841 | Improved |
| use oracle-free perturbation instead of classifier-based screening (ablation) | ReTabSyn · Wilt | AUROC: 0.817 -> 0.841 | Improved |
| use only target-perturbation pairs (proportion 1) rather than 50/50 mix (ablation) | ReTabSyn · Wilt | AUROC / Corr.: AUROC 0.841 -> 0.843; Corr. 0.880 -> 0.862 | Mixed |
| use only feature-perturbation pairs (proportion 0) (ablation) | ReTabSyn · Wilt | AUROC / Corr.: AUROC 0.841 -> 0.821; Corr. 0.880 -> 0.888 | Mixed |
| use dynamic augmentation size instead of fixed 1000 rows (ablation) | ReTabSyn · Wilt | AUROC: 0.835 -> 0.841 | Improved |
| increase augmentation to Nx20 (ablation) | ReTabSyn · Wilt | AUROC: 0.841 -> 0.837 | No clear effect |
| add DPO-Positive regularizer (lambda 0.1 vs 0) (ablation) | ReTabSyn · Wilt | AUROC: 0.772 -> 0.841 | Improved |
| vary DPO beta over 0.05, 0.1, 0.2 (ablation) | ReTabSyn · Wilt | AUROC: 0.837 / 0.841 / 0.839 | No clear effect |
| apply ReTabSyn alignment to Qwen3.5-2B backbone | Qwen3.5-2B · Wilt (synthetic only) | AUROC: 0.870 -> 0.900 | Improved |
Closest papers by meaning
- TabSDS: a Lightweight, Fully Non-Parametric, and Model Free Approach for Generating Synthetic Tabular Data (ICML 2025 · similarity 0.89)
- How Realistic Is Your Synthetic Data? Constraining Deep Generative Models for Tabular Data (ICLR 2024 · similarity 0.87)
- CG-TGAN: Conditional Generative Adversarial Networks with Graph Neural Networks for Tabular Data Synthesizing (AAAI 2025 · similarity 0.86)
- Strengthening LLMs for Tabular Prediction with Structural Priors (NeurIPS 2026 · similarity 0.86)
- Tabular Foundation Model for Generative Modelling (NeurIPS 2026 · similarity 0.86)
- TabNAT: A Continuous-Discrete Joint Generative Framework for Tabular Data (ICML 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
