ReTabSyn: Realistic Tabular Data Synthesis via Reinforcement Learning

Xiaofeng Lin; Seungbae Kim; Zhuoya Li; Zachary DeSoto; Charles Fleming; Guang Cheng

stat.MLMar 11, 2026

ReTabSyn: Realistic Tabular Data Synthesis via Reinforcement Learning

Authors: Xiaofeng Lin, Seungbae Kim, Zhuoya Li, Zachary DeSoto, Charles Fleming, Guang Cheng

Organizations: University of California, Los Angeles, USA · University of South Florida, USA · Cisco, USA · University of Mississippi, USA

Abstract

Deep generative models can help with data scarcity and privacy by producing synthetic training data, but they struggle in low-data, imbalanced tabular settings to fully learn the complex data distribution. We argue that striving for the full joint distribution could be overkill; for greater data efficiency, models should prioritize learning the conditional distribution $P(y\mid \bm{X})$ , as suggested by recent theoretical analysis. Therefore, we overcome this limitation with ReTabSyn, a Reinforced Tabular Synthesis pipeline that provides direct feedback on feature correlation preservation during synthesizer training. This objective encourages the generator to prioritize the most useful predictive signals when training data is limited, thereby strengthening downstream model utility. We empirically fine-tune a language model-based generator using this approach, and across benchmarks with small sample sizes, class imbalance, and distribution shift, ReTabSyn consistently outperforms state-of-the-art baselines. Moreover, our approach can be readily extended to control various aspects of synthetic tabular data, such as applying expert-specified constraints on generated observations.

Figures & tables

Figure 1: Illustration of the challenging scenario for tabular generator: in scenarios with limited or imbalanced training data, tabular generators may produce synthetic datasets containing unrealistic entries, potentially degrading the performance of downstream machine learning tasks.

Figure 2: Overall workflow of ReTabSyn. Starting from scarce, imbalanced real data, we first train a base tabular generator $\pi_{\mathrm{ref}}$ via supervised learning. Guided by the utility decomposition (Thm. 3.1), we construct oracle-free preference-labeled tuples by label/target perturbation : for each row, we keep the conditioning context (prompt) fixed and create a chosen tuple with the original label and a rejected tuple with a perturbed label, forming prompt–chosen–rejected training pairs. Finally, we fine-tune the generator with DPO , using $\pi_{\mathrm{ref}}$ as the reference policy, to enlarge the likelihood margin between chosen and rejected tuples, improving feature–target alignment and downstream ML utility.

Panel of Figure 3 — Figure 3: AUROC scores of synthetic data-trained models on in-distribution test sets, across varying training set sizes. The left panel shows performance on pure synthetic data, and the right panel on real data augmented with synthetic data.

Dataset	Rows	Num Cols	Cate Cols
Adult	48,812	6	8
Bean	13,611	16	1
Churn-Modelling	10,000	6	4
HTRU2	17,898	8	1
Indian Liver Patient	579	9	2
Obesity	2,111	8	9

Table 1: Summary statistics of benchmark datasets

Model	Adult	Churn	HTRU2	Magic	Shoppers
Real	$0.924_{\pm.009}$	$0.883_{\pm.014}$	$0.897_{\pm.008}$	$0.916_{\pm.012}$	$0.701_{\pm.013}$
TVAE	$0.861_{\pm.015}$	$0.821_{\pm.013}$	$0.851_{\pm.012}$	$0.861_{\pm.011}$	$0.590_{\pm.012}$
GReaT	$0.873_{\pm.015}$	$0.834_{\pm.010}$	$0.857_{\pm.014}$	$0.873_{\pm.009}$	$0.604_{\pm.011}$
SMOTE	$0.891_{\pm.010}$	$0.852_{\pm.016}$	$0.869_{\pm.013}$	$0.889_{\pm.015}$	$0.649_{\pm.009}$
TabSyn	$0.882_{\pm.017}$	$0.843_{\pm.011}$	$0.862_{\pm.010}$	$0.882_{\pm.012}$	$0.632_{\pm.015}$
SynRL	$0.862_{\pm.014}$	$0.825_{\pm.012}$	$0.852_{\pm.011}$	$0.863_{\pm.011}$	$0.635_{\pm.011}$

Table 2: PR-AUC at 1% prevalence (mean ±SE over 10 seeds). Best synthetic generator per dataset in bold (Real excluded).

Model	Adult	Churn	Liver	Titanic
Real	$0.868_{\pm.011}$	$0.720_{\pm.012}$	$0.603_{\pm.013}$	$0.862_{\pm.012}$
TVAE	$0.806_{\pm.014}$	$0.686_{\pm.014}$	$0.565_{\pm.015}$	$0.823_{\pm.013}$
GReaT	$0.819_{\pm.013}$	$0.701_{\pm.014}$	$0.576_{\pm.015}$	$0.836_{\pm.012}$
SMOTE	$0.817_{\pm.012}$	$0.704_{\pm.013}$	$0.596_{\pm.013}$	$0.851_{\pm.011}$
TabSyn	$0.808_{\pm.015}$	$0.704_{\pm.016}$	$0.594_{\pm.016}$	$0.849_{\pm.014}$
SynRL	$0.809_{\pm.013}$	$0.688_{\pm.013}$	$0.565_{\pm.014}$	$0.825_{\pm.012}$

Table 3: AUC under distribution shift (mean ±SE over 10 seeds). Best synthetic generator per dataset in bold (Real excluded).

Figure 4: Correlation similarity matrix visualized using a color scale. Shades closer to red indicate high similarity, meaning the correlation between real and synthetic data is well preserved. In contrast, shades of blue signify low similarity, suggesting a weaker alignment between real and synthetic feature correlation.

Model	Precision	Recall	Shape	Corr.
TVAE	$0.280_{\pm.016}$	$0.180_{\pm.010}$	$0.760_{\pm.006}$	$0.760_{\pm.005}$
SMOTE	$\mathbf{0.530}_{\pm.011}$	$0.320_{\pm.008}$	$\mathbf{0.830}_{\pm.005}$	$0.810_{\pm.003}$
TabSyn	$0.410_{\pm.019}$	$0.290_{\pm.010}$	$0.810_{\pm.005}$	$0.800_{\pm.004}$
GReaT	$0.440_{\pm.017}$	$0.210_{\pm.010}$	$0.800_{\pm.005}$	$0.740_{\pm.003}$
SynRL	$0.300_{\pm.015}$	$0.200_{\pm.009}$	$0.780_{\pm.005}$	$0.780_{\pm.005}$
PTA	$0.430_{\pm.017}$	$0.220_{\pm.010}$	$0.790_{\pm.005}$	$0.750_{\pm.006}$

Table 4: Fidelity metrics across models (mean ±SE over 10 seeds; averaged over all small-data splits). Best per column in bold . Shape is the average marginal-distribution similarity (1–KS/TVD) and Corr. is the mean pairwise correlation similarity.

Model	AUC	TPR@FPR=0.01	Authenticity
TVAE	$0.627_{\pm.011}$	$0.061_{\pm.006}$	$0.743_{\pm.008}$
SMOTE	$0.831_{\pm.009}$	$0.438_{\pm.004}$	$0.635_{\pm.006}$
TabSyn	$0.600_{\pm.011}$	$0.052_{\pm.005}$	$0.671_{\pm.011}$
GReaT	$0.587_{\pm.010}$	$0.054_{\pm.005}$	$0.650_{\pm.009}$
SynRL	$0.625_{\pm.008}$	$0.060_{\pm.004}$	$0.740_{\pm.008}$
PTA	$0.589_{\pm.012}$	$0.055_{\pm.005}$	$0.652_{\pm.009}$

Table 5: Privacy attack metrics over 10 seeds. Leakage AUC and TPR@FPR=0.01 are the empirical worst-case (maximum over 13 MIAs), lower-is-better; Authenticity is higher-is-better. Values reported as mean ±SE where available.

Training stages
	AUROC	Corr.	Prec.	Recall
Default ReTabSyn	0.841	0.880	0.763	0.471
Only DPO (no aug.)	0.800	0.810	0.713	0.429
Only Aug. pre-train	0.815	0.851	0.749	0.420
Reward supervision (§ 3.3 )
Oracle editing	0.809	0.867	0.731	0.427

Table 6: Ablation on the Wilt dataset. In each block, a single parameter or settings is varied while others remain at their default values. Metrics include AUROC on synthetic data trained classifiers, correlation similarity, precision, and recall.

Figure 5: Effect of varying the proportion $\rho$ of DPO steps relative to total fine-tuning steps. As $\rho$ increases, fidelity remains stable while utility (ROC AUC) consistently improves.

Appendix figures & tables16 assets

Supplementary material from the paper’s appendix.

Appendix

Dataset Name	Split Column	Train / Test Composition	N Train	N Test
Adult	native-country	United States / non-US	43,152	5,128
Indian liver patient	Gender	Male / Female	441	142
Obesity	Gender	Male / Female	1068	1043
Churn Modeling	Country	France & Germany / Spain	7,519	2,477
Titanic	Sex	Male / Female	577	314

Appendix

Table 7: Dataset splits based on demographic variables to induce distributional drift.

Attack	Threat model	Signal type
DOMIAS ( van Breugel et al., 2023 )	Shadow-box	Density ratio
DPI ( Ward et al., 2024 )	Shadow-box	Local density
Classifier ( Houssiau et al., 2022 )	Shadow-box	Density ratio
Density Estimator ( Houssiau et al., 2022 )	Black-box	Density estimation
DCR ( Chen et al., 2020 )	Black-box	Distance-based
DCR-Diff ( Chen et al., 2020 )	Shadow-box	Distance difference

Appendix

Table 8: Membership-inference attacks used in this study.

Model	Train	Fine‑tune	Sampling
ReTabSyn	1hrs	3mins	7mins
GReaT	1hrs	-	7mins
TabSyn	55mins	-	2mins
SMOTE	-	-	3mins

Appendix

Table 9: Run‑time decomposition of generator operations (Train, Fine‑tune, Sampling) on full Wilt dataset

Model	Hyperparameter	Values
XGBoost	n_estimators	{50, 100, 200}
	max_depth	{3, 4, 5}
	learning_rate	{0.01, 0.1, 0.2}
CatBoost	iterations	{100, 200, 500}
	depth	{4, 6, 8}
	learning_rate	{0.01, 0.05, 0.1}

Appendix

Table 10: Hyperparameter grids for XGBoost and CatBoost

Model	I1	I2	I3
SMOTE	3.28%	2.13%	2.73%
TabSyn	1.02%	1.22%	2.00%
GReaT	1.00%	1.20%	2.00%
ReTabSyn	0.00%	0.00%	0.00%

Appendix

Table 11: Constraint-violation rate (lower is better) on synthetic Adult data.

Model	Abalone	Insurance	Shoppers
Real	0.291	0.714	0.201
SMOTE	0.250	0.690	0.190
TabSyn	0.170	0.200	0.010
GReaT	0.123	-1.417	-0.058
ReTabSyn	0.191	0.226	0.093

Appendix

Table 12: Testing R 2 of downstream models trained on different generators

Dataset	§ 4.3 (imbalance)	§ 4.4 (shift)
Adult	eligible	eligible
Bean	multi-class	no demo/geo column
Churn	eligible	eligible
HTRU2	eligible	no demo/geo column
ILP	low support @1%	eligible
Magic	eligible	no demo/geo column

Appendix

Table 13: Per-dataset eligibility for the imbalanced (§ 4.3 ) and distribution-shift (§ 4.4 ) settings.

Size	Real	GReaT	SMOTE	TabSyn	ReTabSyn	TVAE	SynRL	PTA
32	$0.761_{\pm.018}$	$0.679_{\pm.024}$	$0.722_{\pm.019}$	$0.675_{\pm.029}$	$\mathbf{0.778}_{\pm.021}$	$0.670_{\pm.026}$	$0.686_{\pm.024}$	$0.694_{\pm.022}$
64	$0.798_{\pm.015}$	$0.717_{\pm.019}$	$0.774_{\pm.016}$	$0.736_{\pm.022}$	$\mathbf{0.792}_{\pm.017}$	$0.692_{\pm.020}$	$0.698_{\pm.019}$	$0.720_{\pm.018}$
128	$0.815_{\pm.013}$	$0.744_{\pm.015}$	$0.792_{\pm.012}$	$0.760_{\pm.018}$	$\mathbf{0.813}_{\pm.014}$	$0.710_{\pm.017}$	$0.713_{\pm.016}$	$0.759_{\pm.015}$
256	$0.840_{\pm.011}$	$0.775_{\pm.012}$	$0.810_{\pm.009}$	$0.782_{\pm.015}$	$\mathbf{0.834}_{\pm.010}$	$0.732_{\pm.014}$	$0.738_{\pm.013}$	$0.801_{\pm.012}$
512	$0.878_{\pm.009}$	$0.821_{\pm.010}$	$0.845_{\pm.008}$	$0.814_{\pm.013}$	$\mathbf{0.853}_{\pm.009}$	$0.729_{\pm.011}$	$0.751_{\pm.010}$	$0.826_{\pm.010}$

Appendix

Table 14: Small-data synthetic-only AUROC (mean ±SE over 10 seeds, pooled across datasets). Best synthetic generator per size in bold (Real excluded).

	Wilt			Churn
Method	AUROC	Prec.	Recall	AUROC	Prec.	Recall
Real data	$0.884_{\pm.015}$	$0.631_{\pm.011}$	$0.820_{\pm.015}$	$0.841_{\pm.009}$	$0.686_{\pm.007}$	$0.781_{\pm.007}$
SMOTE	$0.845_{\pm.018}$	$0.605_{\pm.012}$	$0.790_{\pm.018}$	$0.847_{\pm.008}$	$0.689_{\pm.006}$	$0.792_{\pm.006}$
Qwen3.5 Pretrain	$0.870_{\pm.016}$	$0.598_{\pm.014}$	$0.800_{\pm.020}$	$0.830_{\pm.012}$	$0.670_{\pm.010}$	$0.768_{\pm.012}$
Qwen3.5 + ReTabSyn	$\mathbf{0.900}_{\pm.014}$	$\mathbf{0.638}_{\pm.013}$	$\mathbf{0.835}_{\pm.016}$	$\mathbf{0.858}_{\pm.010}$	$\mathbf{0.700}_{\pm.009}$	$\mathbf{0.798}_{\pm.010}$
GPT-5.4 (prompted)	$0.714_{\pm.018}$	$0.541_{\pm.004}$	$0.682_{\pm.015}$	$0.742_{\pm.013}$	$0.709_{\pm.015}$	$0.697_{\pm.013}$

Appendix

Table 15: Cross-backbone validation: synthetic-only downstream utility (mean ±std over 10 seeds), same protocol as § 4.2 . Best generator per dataset in bold (Real excluded).

Metric	SMOTE	GReaT	TabSyn	ReTabSyn	TVAE	SynRL	PTA
Mean AUROC (alternative)	0.678	0.655	0.663	0.685	0.646	0.652	0.673
Mean AUROC (default)	0.744	0.731	0.730	0.756	0.711	0.718	0.740
Wins (alternative)	2/12	1/12	1/12	8/12	0/12	0/12	0/12

Appendix

Table 16: Alternative-target evaluation summary (macro-OVR AUROC, 10 seeds, 5 classifiers). Best in bold .

Target Column	Real	SMOTE	GReaT	TabSyn	ReTabSyn	TVAE	SynRL	PTA
Exited (default)	$0.710_{\pm.011}$	$0.658_{\pm.010}$	$0.667_{\pm.012}$	$0.669_{\pm.017}$	$\mathbf{0.694}_{\pm.005}$	$0.650_{\pm.011}$	$0.654_{\pm.010}$	$0.666_{\pm.009}$
Geography	$0.582_{\pm.003}$	$0.565_{\pm.007}$	$0.573_{\pm.005}$	$0.583_{\pm.005}$	$\mathbf{0.589}_{\pm.003}$	$0.562_{\pm.006}$	$0.566_{\pm.005}$	$0.578_{\pm.005}$
Gender	$0.506_{\pm.005}$	$0.505_{\pm.006}$	$0.501_{\pm.007}$	$0.506_{\pm.004}$	$\mathbf{0.510}_{\pm.006}$	$0.500_{\pm.005}$	$0.502_{\pm.005}$	$0.505_{\pm.004}$
IsActiveMember	$0.543_{\pm.009}$	$0.525_{\pm.004}$	$0.525_{\pm.005}$	$0.529_{\pm.011}$	$\mathbf{0.540}_{\pm.004}$	$0.520_{\pm.006}$	$0.522_{\pm.005}$	$0.531_{\pm.005}$
HasCrCard	$0.502_{\pm.003}$	$0.505_{\pm.002}$	$\mathbf{0.506}_{\pm.003}$	$0.505_{\pm.002}$	$0.503_{\pm.007}$	$0.502_{\pm.003}$	$0.503_{\pm.003}$	$0.504_{\pm.002}$

Appendix

Table 17: Churn alternative targets (DPO target: Exited).

Target Column	Real	SMOTE	GReaT	TabSyn	ReTabSyn	TVAE	SynRL	PTA
NObeyesdad (default)	$0.909_{\pm.006}$	$0.865_{\pm.005}$	$0.820_{\pm.005}$	$0.812_{\pm.010}$	$\mathbf{0.879}_{\pm.010}$	$0.803_{\pm.011}$	$0.810_{\pm.010}$	$0.838_{\pm.009}$
Gender	$0.906_{\pm.007}$	$\mathbf{0.875}_{\pm.016}$	$0.843_{\pm.011}$	$0.831_{\pm.011}$	$0.847_{\pm.013}$	$0.829_{\pm.012}$	$0.835_{\pm.011}$	$0.844_{\pm.010}$
FAVC	$0.743_{\pm.029}$	$0.696_{\pm.018}$	$0.684_{\pm.020}$	$0.672_{\pm.023}$	$\mathbf{0.711}_{\pm.009}$	$0.666_{\pm.017}$	$0.673_{\pm.016}$	$0.695_{\pm.014}$
CAEC	$0.781_{\pm.011}$	$0.726_{\pm.014}$	$0.674_{\pm.012}$	$\mathbf{0.739}_{\pm.015}$	$0.734_{\pm.037}$	$0.682_{\pm.016}$	$0.689_{\pm.015}$	$0.717_{\pm.014}$
family_history	$0.834_{\pm.011}$	$0.811_{\pm.011}$	$0.774_{\pm.006}$	$0.788_{\pm.015}$	$\mathbf{0.813}_{\pm.008}$	$0.760_{\pm.010}$	$0.769_{\pm.009}$	$0.795_{\pm.008}$

Appendix

Table 18: Obesity alternative targets (DPO target: NObeyesdad).

Target Column	Real	SMOTE	GReaT	TabSyn	ReTabSyn	TVAE	SynRL	PTA
Survived (default)	$0.799_{\pm.011}$	$\mathbf{0.789}_{\pm.003}$	$0.752_{\pm.012}$	$0.783_{\pm.023}$	$0.785_{\pm.005}$	$0.744_{\pm.012}$	$0.752_{\pm.011}$	$0.780_{\pm.009}$
Sex	$0.784_{\pm.012}$	$0.765_{\pm.009}$	$0.731_{\pm.013}$	$0.771_{\pm.023}$	$\mathbf{0.783}_{\pm.012}$	$0.723_{\pm.013}$	$0.730_{\pm.012}$	$0.767_{\pm.010}$
Pclass	$0.922_{\pm.008}$	$0.875_{\pm.012}$	$0.864_{\pm.015}$	$0.824_{\pm.010}$	$\mathbf{0.895}_{\pm.010}$	$0.845_{\pm.012}$	$0.852_{\pm.011}$	$0.879_{\pm.010}$
Embarked	$0.762_{\pm.008}$	$0.713_{\pm.017}$	$0.660_{\pm.008}$	$0.652_{\pm.021}$	$\mathbf{0.730}_{\pm.022}$	$0.646_{\pm.018}$	$0.654_{\pm.017}$	$0.700_{\pm.015}$

Appendix

Table 19: Titanic alternative targets (DPO target: Survived).

Target Column	Real	SMOTE	GReaT	TabSyn	ReTabSyn	TVAE	SynRL	PTA
Dataset (default)	$0.701_{\pm.006}$	$0.664_{\pm.016}$	$\mathbf{0.686}_{\pm.022}$	$0.656_{\pm.032}$	$0.667_{\pm.021}$	$0.648_{\pm.019}$	$0.655_{\pm.018}$	$0.675_{\pm.016}$
Gender	$0.617_{\pm.015}$	$\mathbf{0.580}_{\pm.021}$	$0.527_{\pm.017}$	$0.555_{\pm.011}$	$0.566_{\pm.022}$	$0.520_{\pm.018}$	$0.529_{\pm.016}$	$0.562_{\pm.015}$

Appendix

Table 20: Indian Liver Patient alternative targets (DPO target: Dataset).

Setting	Real	SMOTE	GReaT	ReTabSyn	TVAE	SynRL	PTA
Adult, synth	0.853	$0.836_{\pm.001}$	$0.835_{\pm.002}$	$\mathbf{0.841}_{\pm.011}$	$0.823_{\pm.004}$	$0.829_{\pm.003}$	$0.824_{\pm.003}$
Adult, augmented	0.853	$0.849_{\pm.001}$	$0.841_{\pm.001}$	$\mathbf{0.851}_{\pm.001}$	$0.839_{\pm.003}$	$0.844_{\pm.003}$	$0.848_{\pm.002}$
Magic, synth	0.884	$0.878_{\pm.000}$	$0.854_{\pm.002}$	$\mathbf{0.880}_{\pm.003}$	$0.866_{\pm.004}$	$0.851_{\pm.003}$	$0.846_{\pm.003}$
Magic, augmented	0.884	$\mathbf{0.885}_{\pm.001}$	$0.879_{\pm.001}$	$\mathbf{0.885}_{\pm.002}$	$0.873_{\pm.003}$	$0.877_{\pm.003}$	$0.882_{\pm.002}$

Appendix

Table 21: Full-data balanced results: AUROC (mean ±std , 10 seeds). Best synthetic generator per row in bold (Real excluded).

$\beta$	0.05	0.1	0.2
AUROC	0.837	0.841	0.839
$\lambda$	0	0.1	0.2
AUROC	0.772	0.841	0.822

Appendix

Table 22: Sensitivity of downstream AUROC to

\beta

and

\lambda

on Wilt (one parameter varied at a time; default in bold ).

Synthetic Tabular Data Generation

Synthetic tabular data generation creates artificial records that preserve useful feature relationships while limiting exposure of sensitive data.

See all 98 papers

Limited Reference, Reliable Generation: A Two-Component Framework for Tabular Data Generation in Low-Data Regimes
Sep 12, 2025Mingxuan Jiang, Keyang Chen, Yongxin Wang +10Synthetic Data Generation Synthetic Tabular Data Generation
School of Computer Science, Fudan University, Shanghai, 200438, China · Northwestern Polytechnical University, Xi’an, 710129, China · Institute of Financial Technology, Fudan University, Shanghai, 200438, China
Creation begins with understanding: LLMs as strategy designers for privacy-preserving tabular data synthesis
Aug 30, 2026Jinmeng Li, Quan Zhang, Hangting Ye +4Synthetic Tabular Data Generation Privacy-Preserving ML
School of Artificial Intelligence, Jilin University, Changchun, China · Broad College of Business, Michigan State University, East Lansing, USA · Commonwealth Scientific and Industrial Research Organisation (CSIRO), Australia +2
SYNTHONY: A Stress-Aware, Intent-Conditioned Agent for Deep Tabular Generative Models Selection
Mar 31, 2026Hochan Son, Xiaofeng Lin, Jason Ni +1Synthetic Data Selection Tabular Data Synthesis
Department of Statistics, University of California, Los Angeles · Department of Mathematics, University of California, Los Angeles

Sep 12, 2025J/K move · Enter open · S save

Limited Reference, Reliable Generation: A Two-Component Framework for Tabular Data Generation in Low-Data Regimes

Mingxuan Jiang, Keyang Chen, Yongxin Wang, Yongsheng Zhao+9

School of Computer Science, Fudan University, Shanghai, 200438, China · Northwestern Polytechnical University, Xi’an, 710129, China · Institute of Financial Technology, Fudan University, Shanghai, 200438, China

Synthetic Data Generation Synthetic Tabular Data Generation

Open paper AskPDF arXiv

Synthetic tabular data generation is increasingly essential in machine learning, supporting downstream applications when real-world, high-quality tabular data is insufficient. Existing tabular generation approaches, such as generative adversarial networks (GANs) and fine-tuned Large Language Models (LLMs), typically require sufficient reference data, limiting their effectiveness in domain-specific datasets with scarce records. While prompt-based LLMs offer flexibility without parameter tuning, they often generate distributionally drifted data with localized redundancy, leading to degradation in downstream task performance. To overcome these issues, we propose ReFine, a framework that (i) extracts symbolic if-then rules from interpretable models and embeds them into prompts to explicitly guide the generation process toward the domain-specific distribution, and (ii) applies dual-granularity filtering that mitigates over-sampling patterns while preserving rare but informative samples to reduce localized redundancy. Extensive experiments on diverse benchmarks demonstrate that ReFine provides robust downstream utility, achieving a top-tier average rank across datasets and data regimes, with an average relative improvement of 7.48% in extreme low-data regimes.

Abstract

Synthetic Tabular Data Generation

Explore similar work

Limited Reference, Reliable Generation: A Two-Component Framework for Tabular Data Generation in Low-Data Regimes

Creation begins with understanding: LLMs as strategy designers for privacy-preserving tabular data synthesis

SYNTHONY: A Stress-Aware, Intent-Conditioned Agent for Deep Tabular Generative Models Selection