Source-linked AI summary
Synthesis Without Training: An Inference-Only Pipeline for Tabular, Temporal, and Relational Synthetic Data
Zilong Zhao, Abdul Raheem, Jiayu Li, Sohei Arisaka, Darius Lim Hong Yi, Milad Abdollahzadeh, Uzair Javaid, Biplab Sikdar
TL;DR
Per-dataset training, modality-specific pipelines, and opaque constraints complicate synthetic-data generation. GENSCRIPT replaces model fitting with profile-driven constraint inference and code generation, achieving fast generation with near-leading marginal fidelity while preserving explicit dependencies.
Problem
Fit-then-sample generation requires new training for each dataset, separate approaches across modalities, and leaves rule violations difficult to inspect or repair.
Source: §1 · pp. 1–2Method
GENSCRIPT deterministically profiles data, has an LLM infer semantics and integrity constraints, and compiles them into an executable sampler without training a generator.
Source: §1 · p. 2, §2.2 · p. 3Results
Across four single-table benchmarks, GENSCRIPT is fastest on construction and sampling while close to leaders in marginal fidelity; it also matches building-telemetry data more closely than two temporal models and emits no Adult rows violating a functional dependency.
Source: §4 · p. 4Takeaways & Limitations
Practitioners receive a readable program they can inspect, edit, and rerun, with constraints represented explicitly rather than buried in a fitted generator.
Source: §1 · p. 2, §4 · p. 4Constraint induction is essentially pairwise, leaving dependencies over three or more columns invisible; relational evaluation covers one two-table schema, so deeper hierarchies, composite keys, and cycles remain untested.
Source: §4 · pp. 4–5
Abstract
from arXiv · showhide
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.
1 Introduction
GENSCRIPT argues that schema-driven data can be generated without training a model, using a profile, inferred rules, and an executable sampler. This approach addresses the per-dataset cost, modality fragmentation, and opacity of fit-then-sample generation.
1 Introduction
- Synthetic surrogates are used to share sensitive data, test systems, and develop downstream models, while common table, sequential, and relational methods fit models to private data.Examples include conditional GANs, VAEs, diffusion models, and large-language-model-based generators.Source: §1 · p. 1
- Fit-then-sample methods require a new training run for each dataset, separate model families across modalities, and offer no direct place to inspect or repair rule violations.The paper illustrates opacity with a generated refund on an unpaid order.Source: §1 · pp. 1–2
Our position
- GENSCRIPT replaces a trained generator with an LLM-authored sampling program built from a deterministic profile and explicit, reviewable constraints.The paper limits its claim to schema-driven operational data whose structure can be summarized by a profile.Source: §1 · p. 2
2 Method
GENSCRIPT profiles source data deterministically, uses an LLM to infer semantics and constraints, and compiles these into a sampler. The same profile-based interface supports single-table, temporal, and relational data.
2 Method
- The pipeline ingests source tables, constructs a generator through profiling, analysis, and assembly, then generates and evaluates synthetic data.The source dataset may contain K tables, with K = 1 for the single-table case.Source: §2 · p. 2
2.1 Ingest and Profile: deterministic statistical profiling
- A fixed script builds a reproducible profile with column types, ranges or category frequencies, missingness, summary statistics, and table correlations.Relational profiles add keys and parent-child cardinalities; temporal profiles add timestamp granularity, inter-arrival gaps, sequence lengths, and coverage.Source: §2.1 · p. 2
2.2 Analyze: semantic analysis and constraint induction
- An LLM uses the profile and correlations to infer field semantics, cross-column rules, and conditional sampling structure, with candidate predicates checked against real rows.Predicates holding on fewer than τ of real rows are demoted or discarded; experts can review and correct the resulting confidence-weighted rules.Source: §2.2 · pp. 2–3
2.3 Assemble, Generate and Evaluate: synthesizing and running the sampler
- A coding agent compiles the profile and constraints into a generator that samples candidates and accepts them only when hard predicates hold.For deterministic constraints such as arithmetic identities or derived fields, the agent computes dependent values directly rather than rejecting candidates.Source: §2.3 · p. 3
2.4 Extension to time series and relational data
- For temporal data, profiles add intervals, sequence lengths, trends, seasonality, and lag autocorrelations; relational generation follows schema order and assigns foreign keys by construction.This construction preserves referential integrity by design, while the practitioner selects no model.Source: §2.4 · p. 3
3 Experiments
Across four single-table benchmarks and a building dataset, experiments show near-leading marginal fidelity, faster construction and sampling, exact preservation of a functional dependency, and a closer temporal match than two baselines.
Setup and metrics
- The single-table tests used CoverType, Credit, Intrusion, and Adult against TabTreeFormer [7], TabDiff, and REaLTabFormer [15]; utility was train-on-synthetic/test-on-real XGBoost AUC.The first three datasets used 50,000 rows and Adult used 32,561; GENSCRIPT's “Train” time covers profiling, constraint induction, and program synthesis, with no gradient step.Source: §3 · p. 3
Fidelity and utility
- MLE was 0.894 on Credit and 0.828 on Adult, versus 0.572 on Intrusion and 0.467 on CoverType, degrading with more target classes.The first two datasets are binary; Intrusion has twenty classes and CoverType has seven.Source: §3 · p. 3
- Shape scores were 0.984 on CoverType, 0.936 on Adult, 0.930 on Credit, and 0.874 on Intrusion; GENSCRIPT was last on Trend across all four datasets (0.748 to 0.897).It ranked second of four on CoverType Shape and third elsewhere.Source: §3 · p. 3, Table 1 · p. 4
Cost
- GENSCRIPT built generators in 73 to 115 seconds versus 175 to 6760 seconds for baselines, and sampled in 0.9 to 5.4 seconds versus 8.6 to 534 seconds.Table 2 reports construction and sampling times in seconds.Source: Table 2 · p. 4, §3 · p. 4
Constraint preservation
- GENSCRIPT emitted none of the invalid education/education-num combinations, while every trained baseline violated the dependency; the real table has exactly 16 valid pairs.The aggregate metrics do not register these violations.Source: §3 · p. 4
Time series and relational structure
- On the building dataset, all methods preserved key integrity, but GENSCRIPT identified the key from the profile and its temporal samples interleaved with real data while both baselines formed disjoint clusters.The dataset is an office-building time series linked to an access-control table; its categorical source.id frequencies are recorded exactly in the profile.Source: §3 · p. 4
4 Conclusion and future work
GENSCRIPT demonstrates useful synthetic generation without training, combining fast construction and sampling with strong marginal fidelity and explicit constraint enforcement. The paper identifies pairwise constraint induction and limited relational evaluation as important boundaries.
Conclusion
- Across four single-table benchmarks, GENSCRIPT was fastest to construct and sample while staying close to leaders on marginal fidelity; it also matched building-telemetry distributions more closely than two temporal models.It alone emitted no rows violating a functional dependency, because the rule was named and enforced rather than fitted.Source: §4 · p. 4
Limitations
- Because correlation drives constraint induction, dependencies over three or more columns are invisible, limiting the approach for wide tables and the Trend gap.Extending induction beyond pairwise structure is identified as the most direct route to closing that gap.Source: §4 · pp. 4–5
- Relational evaluation covers one two-table schema; deeper hierarchies, composite keys, and cycles remain untested.Source: §4 · p. 5
A Extended related work
Prior work either trains modality-specific generators or uses language models to extract structure before handing generation to a fitted model. GENSCRIPT instead targets cross-modal coverage with a program that retains explicit structure.
Related work: single tables
- Single-table methods train on private data, while TabDiff and TabTreeFormer [7] remain within that modality and weight space; PrivBayes and SDV fit dependencies but cannot express logical constraints.Source: A Extended related work · p. 7
Related work: sequential and relational
- Sequential and relational work uses separate trained architectures, whereas GENSCRIPT aims to cover all three modalities without training and asks the language model to write a sampler program, not rows.The program is described as a short verifiable artifact that makes generation deterministic given a seed.Source: A Extended related work · p. 7
Related work: LLMs making sense of tabular data
- LLM-TabFlow [10] and SPADA extract structure but hand generation to fitted statistical models; GENSCRIPT hands it to a program, while its Analyze phase also infers generation-relevant semantics and cross-column constraints.Related semantic annotation work includes TabLLM, Narayan et al., and ArcheType.Source: A Extended related work · p. 7
B Dataset statistics
The benchmark statistics cover four public single tables and two partner tables with distinct class counts and scales.
B Dataset statistics
- CoverType, Credit, and Intrusion use 50,000 target-stratified rows, while Adult uses its standard training split; all six tables generate as many rows as their sources contain.The partner tables are acc, a 100-row reference table, and aicamera, with 138,341 detections averaging roughly 1,400 per registered user.Source: Table 3 · p. 7, B Dataset statistics · p. 8
- Intrusion has 20 synthesized classes after rare labels are dropped, so its macro-averaged AUC spans more classes than CoverType’s seven.The 10% KDD file has 23 labels; stratified subsampling removes spy, perl, and phf, each occurring fewer than five times.Source: B Dataset statistics · p. 8
- Aicamera’s source.id has 242 distinct values, whose frequencies the profile records exactly and the generator reproduces by construction.The acc and aicamera tables form a pair of partner tables with substantially different sizes.Source: B Dataset statistics · p. 8
C Per-dataset fidelity and utility
GENSCRIPT stays competitive on marginal Shape fidelity but trails on Trend and utility, particularly for multi-class targets.
C Per-dataset fidelity and utility
- 0.984 on CoverType places GENSCRIPT second of four on Shape, ahead of TabDiff and REaLTabFormer and within 0.7 points of TabTreeFormer.Its Shape scores are 0.936 on Adult and 0.930 on Credit, within 5 and 6 points of the leaders, respectively.Source: C Per-dataset fidelity and utility · p. 8
- GENSCRIPT is last on Trend across all four datasets, scoring 0.748–0.897 because it reproduces only dependencies captured by induced rules or proposal conditionals.Pairwise correlations not captured by those dependencies are not modelled.Source: C Per-dataset fidelity and utility · p. 8
- 0.894 AUC on Credit and 0.828 on Adult trail the best baselines’ 0.999 and 0.925 on the binary-target datasets.These results are described as usable but behind the best baseline.Source: C Per-dataset fidelity and utility · p. 8
- 0.572 AUC on Intrusion and 0.467 on CoverType mark a sharp multi-class utility decline, with CoverType below chance while no baseline falls below chance.The sampler draws the target from its marginal without conditioning on the determining features, which costs more for multi-class targets.Source: C Per-dataset fidelity and utility · p. 8
F Constraint preservation on Adult
The Adult functional dependency reveals a constraint-preservation advantage that aggregate fidelity metrics miss: GENSCRIPT enforces the rule and emits no violating rows.
F Constraint preservation on Adult
- GENSCRIPT emits zero violations of the Adult education–education-num dependency, which maps 16 diploma labels to year counts; no trained baseline reproduces it exactly.The dependency is materialized as a hard predicate, making zero violations a construction guarantee rather than a post-generation filtering result.Source: Table 4 · p. 8, F Constraint preservation on Adult · pp. 8–9
- The completed Adult run recovered nine constraints, enforced two structurally, passed 30 of 33 validation checks, and produced an exportable generator.The exported generator can be audited and re-run independently of the pipeline that wrote it.Source: Figure 2 · p. 9
- In the aicamera t-SNE embedded by source.id, GENSCRIPT lies inside the real manifold while TabularARGN and TimeAutoDiff occupy disjoint regions.The figure encodes GENSCRIPT in orange, real sequences in blue, TabularARGN in red, and TimeAutoDiff in green.Source: Figure 3 · p. 9
- A correct count of distinct pairs does not ensure correct diploma–year matches; the row-level violation rate is the diagnostic that matters.A row with an impossible year count violates the functional dependency even if such rows are rare.Source: F Constraint preservation on Adult · pp. 8–9, Table 4 · p. 8
- Rare invalid pairs can evade spot-checks and surface later in downstream jobs that join or filter on the columns; Shape, Trend, and MLE do not detect them.The dependency allows no exceptions, regardless of how infrequently a violation occurs.Source: F Constraint preservation on Adult · p. 9