Synthesis Without Training: An Inference-Only Pipeline for Tabular, Temporal, and Relational Synthetic Data
Authors: Zilong Zhao, Abdul Raheem, Jiayu Li, Sohei Arisaka, Darius Lim Hong Yi, Milad Abdollahzadeh, Uzair Javaid, Biplab Sikdar
Organizations: Betterdata AI · National University of Singapore · University of Illinois Urbana-Champaign · KAJIMA Technical Research Institute Singapore
Synthetic data generation is dominated by the fit-then-sample paradigm: a generative model is trained on a private dataset and then sampled from. Despite its widespread adoption, this paradigm faces three challenges: (1) a new training run is required for every dataset; (2) different data modalities, such as single tables, time series, and relational databases, require task-specific models and feature engineering; and (3) the resulting model is opaque, making its behavior under data constraints difficult to inspect. We propose GENSCRIPT, an inference-only pipeline that eliminates model training. GENSCRIPT computes a deterministic statistical profile of the source data (column types, ranges, missingness, categories, correlations, etc.) and passes it--rather than raw rows--to a language model to infer field semantics and cross-column integrity constraints. A coding agent then compiles the profile and constraints into an executable, auditable sampler. This unified approach supports single-table, temporal, and relational data without task-specific modeling. Across four single-table benchmarks, GENSCRIPT builds generators in 2 minutes and samples 50k rows within 6 seconds, while remaining within a few points of leading methods in marginal fidelity. Notably, it is the only method that perfectly preserves a 1-to-1 mapping between columns in the Adult dataset. On a smart-building dataset, it produces conditional time series that more closely match the real distribution than two baselines and perfectly preserves primary- and foreign-key relationships in the corresponding relational database.
Figures & tables
Figure 1: The GenScript working pipeline. The implemented UI is provided in Appendix D .
CoverType
Credit
Intrusion
Adult
Method
Shape
Trend
MLE
Shape
Trend
MLE
Shape
Trend
MLE
Shape
Trend
MLE
TabTreeFormer
0.991
0.991
0.919
0.929
0.942
0.971
0.866
0.982
0.565
0.935
0.972
0.822
REaLTabFormer
0.983
0.990
0.937
0.962
0.948
0.947
0.968
0.962
0.999
0.966
0.930
0.925
TabDiff
0.975
0.820
0.894
0.992
0.995
0.999
0.989
0.989
0.995
0.986
0.973
0.924
GenScript
0.984
0.784
0.467
0.930
0.897
0.894
0.874
0.748
0.572
0.936
0.858
0.828
Table 1: Fidelity and utility on the four single-table benchmarks. Best per column in bold.
CoverType
Credit
Intrusion
Adult
Method
Train ↓
Sample ↓
Train ↓
Sample ↓
Train ↓
Sample ↓
Train ↓
Sample ↓
TabTreeFormer
2324.9
474.9
273.6
63.1
217.5
51.8
2117.0
43.6
REaLTabFormer
405.9
159.6
2194.5
533.5
581.0
249.9
175.1
37.8
TabDiff
5893.7
16.5
4934.3
13.1
5567.1
16.1
6760.0
8.6
GenScript
112.2
3.0
93.0
5.4
114.9
2.5
73.3
0.9
Table 2: Generator-construction time (“Train”) and sampling time, in seconds. For GenScript , “Train” covers profiling, constraint induction, and program synthesis.
Appendix figures & tables4 assets
Supplementary material from the paper’s appendix.
Appendix
Table
Domain
Rows
Cols.
Classes
Source
CoverType
forest cover
50 000
55
7
UCI Covertype (581 012 rows)
Credit
card transactions
50 000
31
2
Kaggle creditcardfraud (284 807)
Intrusion
network traffic
50 000
42
20
UCI KDD Cup 1999
Adult
census income
32 561
15
2
UCI Adult (48 842 in full)
acc
building access
100
6
—
industrial partner
aicamera
camera detections
138 341
4
—
industrial partner
Appendix
Table 3: The six source tables. Columns counts include the target where one exists. Following common practice for these benchmarks, CoverType, Credit and Intrusion are subsampled to 50 000 rows stratified on the target; Adult is used at its standard training-split size. Rows generated equals rows in the source table for every method.
Figure 2: The GenScript application user interface, on a completed Adult run. The phase bar tracks the pipeline of Section 2 . The cards report what this run recovered: nine induced constraints, two of them enforced structurally, and 30 of 33 validation checks passed. The generator is exportable, and so auditable and re-runnable independently of the pipeline that wrote it.
Figure 3: t-SNE of aicamera sequences embedded by source.id . GenScript (orange) lies inside the real manifold (blue); TabularARGN (red) and TimeAutoDiff (green) occupy disjoint regions.
Method
Distinct pairs (real: 16)
Violating rows (%) ↓
FD holds
TabTreeFormer
16
61.104
No
REaLTabFormer
19
0.009
No
TabDiff
30
0.190
No
GenScript
16
0.000
Yes
Appendix
Table 4: Consistency of the education ↔ education-num functional dependency on Adult. The real table maps each of its 16 diploma labels to exactly one year count; a row whose year count does not match its label is a violation. Percentages are of the 32 561 generated rows.
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