Deep generative models for tabular data (GANs, diffusion models, and LLM-based generators) exhibit highly non-uniform behavior across datasets; the best-performing synthesizer family depends strongly on distributional stressors such as long-tailed marginals, high-cardinality categorical, Zipfian imbalance, and small-sample regimes. This brittleness makes practical deployment challenging, especially when users must balance competing objectives of fidelity, privacy, and utility. We study {intent-conditioned tabular synthesis selection}: given a dataset and a user intent expressed as a preference over evaluation metrics, the goal is to select a synthesizer that minimizes regret relative to an intent-specific oracle. We propose {stress profiling}, a synthesis-specific meta-feature representation that quantifies dataset difficulty along four interpretable stress dimensions, and integrate it into {SYNTHONY}, a selection framework that matches stress profiles against a calibrated capability registry of synthesizer families. Across a benchmark of 7 datasets, 10 synthesizers, and 3 intents, we demonstrate that stress-based meta-features are highly predictive of synthesizer performance: a kNN selector using these features achieves strong Top-1 selection accuracy, substantially outperforming zero-shot LLM selectors and random baselines. We analyze the gap between meta-feature-based and capability-based selection, identifying the hand-crafted capability registry as the primary bottleneck and motivating learned capability representations as a direction for future work.
Figures & tables
Strategy
Top-1
Top-3
Spear.
NDCG
k NN (k=3) † (oracle-informed)
.556
.778
.578
.918
Random (E[1000])
.111
.304
−.001
.815
Vanilla LLM (GPT-4o-mini)
.000
.111
−.087
.784
Static Heuristic
.333
.778
.422
.889
Synthony (joint opt)
.000
.556
.557
.893
Synthony (SF-only)
.444
.667
.573
.914
Table 1: Test set performance of intent-conditioned model selection strategies (9 dataset-intent pairs). Primary metrics are italicized . k NN † uses oracle neighbor rankings from training datasets and serves as an oracle-informed reference point , not a deployable zero-knowledge method. Bold indicates best among zero-knowledge methods. With 9 test pairs, each Top-1 prediction shifts accuracy by ≈ 11.1 percentage points.
Variant
Top-1
Top-3
Spear.
NDCG
Vanilla LLM (no Synthony )
.000
.111
−.087
.784
− Stress − Focus (bare scoring)
.444
.667
.478
.904
− Focus scaling (stress only)
.444
.667
.602
.910
− Stress profiling (focus only)
.333
.667
.430
.901
Full Synthony (SF-only)
.444
.667
.573
.914
Table 2: Ablation study on the SF-only pipeline (test set, 9 pairs). Components: (A) stress profiling, (B) capability matching, (C) focus scaling. Stress profiling provides the strongest signal; adding focus scaling degrades test performance due to overfitting at this sample size ( n=12 training pairs, 18 parameters).
Predictor
Top-1
Top-3
Spear.
NDCG
Majority vote
.333
.778
.300
.899
Logistic regression
.556
.889
.293
.893
Decision tree
.222
.667
.299
.900
k NN (k=3)
.556
.778
.578
.918
Table 3: Predicting the best synthesizer from dataset meta-features (test set, 9 pairs). Majority vote’s 33.3% Top-1 reflects CART’s frequent wins (8/21 oracle-best); the discriminative signal is in Spearman, where k NN ( 0.578 ) substantially outperforms frequency-based and parametric baselines ( 0.293 – 0.300 ).
Appendix figures & tables11 assets
Supplementary material from the paper’s appendix.
Appendix
Split
Top-1
Top-3
Spearman
NDCG
Training (12 pairs)
.250
.750
.610
.897
Test (9 pairs)
.000
.556
.557
.893
Appendix
Table 4: Joint optimization results on train and test splits.
Skew
Cardinal.
Zipfian
Small data
Correl.
Privacy
Model
Orig
Opt
Orig
Opt
Orig
Opt
Orig
Opt
Orig
Opt
Orig
Opt
AIM
3
2
0
0
1
3
2
4
3
0
4
2
ARF
2
4
4
2
3
4
4
0
4
1
0
1
AutoDiff
1
4
3
1
2
3
2
0
1
3
0
4
BayesianNet.
3
4
4
4
2
2
4
4
3
2
0
4
CART
3
4
4
2
2
2
4
4
4
3
0
4
Appendix
Table 5: Capability scores: hand-crafted (left) vs. optimized (right). Changes of ≥2 are bolded.
Intent
Skew
Cardinal.
Zipfian
Small
Correl.
Privacy
Privacy
5.69
3.78
7.32
9.83
8.90
9.10
Fidelity
2.43
7.36
4.10
3.11
2.78
9.44
Utility
0.68
2.59
3.71
5.36
1.42
2.96
Appendix
Table 6: Optimized scale factors αi,j per intent and capability dimension.
Figure 1: SYNTHONY system architecture. Solid arrows denote the real-time recommendation pipeline; dashed arrows denote the offline calibration loop (Section 3.4 ). Data flows left through profiling, center through scoring, and returns right through the API.
Figure 2: Rule-based scoring pipeline. The match function mj implements a decay curve: 1.0 (exact match), 0.7 (one level below), 0.4 (two levels below), 0.0 (otherwise). Scale factors αi,j are applied when an intent is specified. The hard-problem path is bypassed when scale factors are provided to allow the optimizer full control.
Condition
Match Score
Interpretation
cg,j≥rj
1.0
Capability meets or exceeds requirement
cg,j=rj−1
0.7
Near miss (one level below)
cg,j=rj−2
0.4
Moderate gap (two levels below)
cg,j<rj−2
0.0
Insufficient capability
Appendix
Table 7: Score decay curve: match score as a function of capability gap.
Figure 3: API architecture with SQLite persistence. Double arrows indicate read/write paths. Uploaded files are stored on disk; profiles and recommendations are cached in SQLite for session-based retrieval.
Component status (analyzer, recommender, LLM availability)
Appendix
Table 9: REST API endpoint summary.
Figure 4: MCP server architecture. The server exposes SYNTHONY’s capabilities through three MCP primitives: Tools (executable functions), Resources (read-only data), and Prompts (guided workflows). Communication uses JSON-RPC 2.0 over stdio.
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∣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.
Xiaofeng Lin, Seungbae Kim, Zhuoya Li +3
University of California, Los Angeles, USA · University of South Florida, USA · Cisco, USA +1
Sharing tabular data in high-stakes domains is constrained by privacy regulations. Synthetic data offer a promising alternative, but deep generative models are costly to train and difficult to audit, while LLM-based methods often serialize records as text, obscuring tabular structure and exposing sensitive data. We introduce Tabular Synthesis Strategy Designer (TabSSD), which uses an LLM to design synthesis procedures rather than directly generate records. TabSSD provides the LLM with tree-derived summaries of variable dependence rather than raw records, which produces Python programs for local execution and evaluation. Across twelve datasets, TabSSD strikes a favourable balance among statistical fidelity, predictive utility, and empirical privacy risk, achieving the best average rank across six metrics among ten methods. Moreover, it substantially reduces local computation and token consumption relative to the compared methods. By enabling human-guided refinement and eliminating user-side model tuning, TabSSD lowers the expertise and infrastructure barriers to transparent tabular data synthesis.
Jinmeng Li, Quan Zhang, Hangting Ye +4
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
Synthetic tabular data support use cases like data sharing, model development under access restrictions, and rapid prototyping of analytical workflows. Modern generative models are evaluated by their statistical similarity, correlation structure, privacy, and downstream machine-learning utility. However, such evaluations leave a gap: they rarely test the structure that matters for analytical queries. We present TabQueryBench, a query-centric benchmark for evaluating synthetic tabular data with SQL-shaped analytical queries. TabQueryBench distills 12 public query sources into 44 reusable query templates and grounds them to each dataset through a policy-guided template-to-SQL pipeline. Across 49 datasets and 11 generative models, it produces more than 100 executable queries per dataset while preserving cross-model comparability. Our evaluation reveals a substantial gap between conventional and query-centric fidelity. Even the strongest model, RealTabFormer, reaches only 0.76 +/- 0.23 query-centric fidelity compared with 1.00 for real data. Failures are most pronounced for high-cardinality attributes and local or tail queries, while model quality also trades off against generation cost. These results show that synthetic data can appear statistically realistic yet still fail to preserve the analytical behavior of real data, motivating query-centric fidelity as a necessary complement to existing evaluation metrics. We open source TabQueryBench to facilitate research in this direction.
Jialin Zhang, Fenghao Dong, Yajie Zhou +2
University of Hong Kong · Tongji University · Carnegie Mellon University +1