TabularARGN: Autoregressive Tabular Data Synthesis

• TabularARGN is a discretization-first autoregressive model that generates synthetic tabular data by factorizing records into ordered discrete sub-columns.
• It employs any-order training with random permutation masking to support unconditional synthesis, arbitrary conditional generation, and effective imputation.
• Empirical evaluations show high fidelity, computational efficiency, and robust privacy guarantees compared to GAN, VAE, and diffusion-based models.

Tabular Auto-Regressive Generative Network (TabularARGN) is a discretization-first neural framework for synthetic tabular data generation in which each record is modeled as an ordered sequence of conditionally generated discrete sub-columns. In the 2025 literature, the name appears in two closely related formulations: a privacy-preserving flat-table generator for sensitive data, and a broader framework that extends the same any-order autoregressive principle to mixed-type, multivariate, and sequential settings. In both cases, the defining idea is explicit autoregressive density estimation over transformed tabular attributes rather than adversarially learned implicit generation, with support for unconditional synthesis, arbitrary conditional generation, and imputation through random-order training (Sidorenko et al., 8 Aug 2025, Tiwald et al., 21 Jan 2025).

1. Conceptual scope and problem setting

TabularARGN addresses synthetic tabular data generation as a density-estimation problem over heterogeneous records. The target is not merely realistic-looking rows, but synthetic tables that remain statistically representative, analytically useful, structurally consistent, and privacy-protective. The privacy-focused formulation emphasizes sensitive domains such as healthcare and argues that traditional anonymization often fails to preserve privacy adequately, while the broader framework positions TabularARGN as a practical alternative to heavier GAN-, VAE-, diffusion-, and transformer-based synthesizers for mixed-type, high-dimensional, relational, and sequential data. The broader formulation also states that the model is closest in spirit to NADE, but adapts autoregression to categorical sub-column encodings, arbitrary conditioning sets, and sequence-aware generation (Sidorenko et al., 8 Aug 2025, Tiwald et al., 21 Jan 2025).

What “auto-regressive” means in this context is explicit columnwise factorization. After preprocessing, a record is represented as $D$ discrete sub-columns, and the model learns

$p(\mathbf{x})=\prod_{i=1}^{D} p(x_i \mid x_{<i}).$

A notable feature is that TabularARGN need not be restricted to a single left-to-right order. In its any-order form, feature order is randomly shuffled per batch, so training minimizes negative log-likelihood under random permutations and exposes the network to conditionals over many subsets rather than one fixed prefix structure. This makes the method a tabular-native autoregressive model rather than a direct import of text serialization practices (Sidorenko et al., 8 Aug 2025, Tiwald et al., 21 Jan 2025).

2. Representation, discretization, and any-order probability modeling

The preprocessing layer is central to TabularARGN. Every original attribute is converted into one or more categorical sub-columns so that the entire model can be trained with categorical cross-entropy. The common probabilistic statements reported for the framework are

$p(\mathbf{x})=\prod_{i=1}^{D} p(x_i \mid x_{<i}),$

and, for any-order training,

$$\max_{\theta}\; \mathbb{E}_{\sigma}\Biggl[\sum_{i=1}^{D}\log p_{\theta}\bigl(x_{\sigma(i)}\mid x_{\sigma(<i)}\bigr)\Biggr],$$

with the broader framework also describing support for arbitrary conditioning sets through

$$\hat{p}(x_i \mid \{x_j:j\in S\}),\quad \forall S\subseteq \{1,\dots,D\}\setminus\{i\}.$$

After sampling, the discretization mapping is reversed to reconstruct rows in the original schema (Sidorenko et al., 8 Aug 2025, Tiwald et al., 21 Jan 2025).

Reported transformations across the two TabularARGN papers are summarized below.

Original field type	Reported transformation	Decoding note
Categorical / binary	Kept discrete; binary treated as categorical	Direct categorical reconstruction
Numerical	Percentile binning or digit splitting; alternatively numeric-discrete, numeric-binned, or numeric-digit	Binned values may be sampled uniformly within intervals
Datetime	Split into year, month, day, time; sequential model also uses relative datetime encoding	Components reassembled after sampling
Geospatial	Discretized into categorical quadtiles or quad-tile sequences	Inverse mapping restores geographic representation
Missing / empty values	Treated as separate categories in the broader framework	Privacy-focused paper does not fully specify missing-value handling

The broader framework states that missing or empty values are modeled as separate categories and therefore can be reproduced by the generator. The privacy-focused paper, by contrast, explicitly notes that exact handling of missing values is not specified there. Both papers stress that preprocessing choices are not incidental: performance depends materially on discretization design, and larger numerical datasets can inflate model size because digit-based encodings may expand one original column into many sub-columns (Sidorenko et al., 8 Aug 2025, Tiwald et al., 21 Jan 2025).

3. Flat-table architecture and training mechanics

The flat TabularARGN model is shallow and modular. For each discretized sub-column it comprises an embedding layer, a regressor layer, and a predictor layer, connected by a permutation masking layer that enforces the current autoregressive order. All sub-column embeddings are concatenated; for a target $x_i$, the mask zeros out embeddings that should be invisible under the current permutation, so only valid predecessors are passed to the target-specific regressor. The predictor then outputs a softmax over the target sub-column categories. Training uses teacher forcing, so visible predecessors are ground-truth values rather than sampled values. The privacy-focused paper reports size heuristics

$$\text{Embeddings}=3\cdot (d_{\text{in}})^{0.25},\qquad \text{Regressors}=16\cdot \max(1,\ln(d_{\text{in}})),\qquad \text{Predictors}=d_{\text{in}},$$

while the broader framework reports default regressor layers $[16]$, ReLU activations, and dropout in the regressor block (Sidorenko et al., 8 Aug 2025, Tiwald et al., 21 Jan 2025).

Optimization and stopping logic are likewise simple. Both formulations use a 10% validation split, compute validation loss each epoch, halve the learning rate after $K=3$ stagnant epochs, stop after $N=5$ epochs without improvement, and retain the best checkpoint. The privacy-focused paper states a 25% dropout rate in regressor layers. At sampling time, generation is strictly autoregressive: choose an order, remove the training-time permutation mask, start from zeros, predict the first feature from its marginal, embed the sampled value, then continue sequentially until all sub-columns are generated and decoded back into original-format records (Sidorenko et al., 8 Aug 2025, Tiwald et al., 21 Jan 2025).

4. Conditional generation, imputation, and sequential or relational extensions

Any-order training is what gives TabularARGN its flexibility beyond unconditional sampling. Because the model is exposed to random permutations during training, it can be used at inference time with arbitrary subsets of columns fixed as evidence. Conditional generation is performed by placing observed columns before unobserved ones in the chosen order, or equivalently by holding observed embeddings fixed and sampling the rest autoregressively. Imputation is the same mechanism with missing fields designated as targets. The broader framework also claims fairness-aware generation through this conditioning flexibility, but does not formalize a fairness loss or a dedicated constrained optimization procedure; fairness is presented as an enabled downstream use of arbitrary conditional sampling rather than as a separate architectural component (Tiwald et al., 21 Jan 2025).

The broader 2025 formulation extends TabularARGN beyond i.i.d. flat rows. For sequential data, it adds a history encoder whose encoded history $p(\mathbf{x})=\prod_{i=1}^{D} p(x_i \mid x_{<i}).$0 is the last hidden state of an LSTM with dropout. At each current time step, $p(\mathbf{x})=\prod_{i=1}^{D} p(x_i \mid x_{<i}).$1 is concatenated with visible current-time-step embeddings and fed to the regressor, yielding a factorization described operationally as autoregressive over time and autoregressive within each time step over columns. Variable sequence lengths are handled by adding two categorical features during preprocessing: a sequence-length column and a counting-index column. In two-table settings, a context processor compresses a flat parent table into $p(\mathbf{x})=\prod_{i=1}^{D} p(x_i \mid x_{<i}).$2, and child-sequence generation conditions on both $p(\mathbf{x})=\prod_{i=1}^{D} p(x_i \mid x_{<i}).$3 and this flat-table context. The framework explicitly states that no equidistant time steps are required, and generation proceeds by sampling the flat table first, then the sequential child table (Tiwald et al., 21 Jan 2025).

5. Empirical performance, efficiency, and privacy evaluation

The broader framework reports strong fidelity and efficiency on both flat and sequential benchmarks. On flat data, aggregate accuracy is reported as 97.9% on Adult and 98.5% on ACS-Income, with Adult just behind TabSyn at 98.2% and ACS-Income ahead of TabSyn at 97.5%. On sequential data with flat context, TabularARGN reports 88.4% on Baseball and 98.1% on California, exceeding REaLTabFormer, RC-TGAN, and ClavaDDPM on both benchmarks. The same paper emphasizes compute efficiency: Adult training time is reported as 138 s versus 2,316 s for TabSyn, 4,588 s for STaSy, and 2,684 s for CTGAN; Baseball training time is 174 s versus 86,031 s for REaLTabFormer and 40,564 s for RC-TGAN. Parameter counts are correspondingly small: on Adult, TabularARGN is reported at 129,768 parameters versus 10,616,892 for TabSyn, 11,768,942 for TabDDPM, and 10,695,154 for STaSy. Its DCR share stays near the ideal 50% on the reported datasets: 50.3% on Adult, 50.2% on ACS-Income, 50.5% on Baseball, and 51.1% on California. Differentially private variants remain competitive, with Adult DP TabularARGN at 93.8% for $p(\mathbf{x})=\prod_{i=1}^{D} p(x_i \mid x_{<i}).$4, ACS-Income at 96.7% for $p(\mathbf{x})=\prod_{i=1}^{D} p(x_i \mid x_{<i}).$5, Baseball at 79.0% for $p(\mathbf{x})=\prod_{i=1}^{D} p(x_i \mid x_{<i}).$6, and California at 93.4% for $p(\mathbf{x})=\prod_{i=1}^{D} p(x_i \mid x_{<i}).$7 (Tiwald et al., 21 Jan 2025).

The privacy-focused paper evaluates TabularARGN on 11 datasets within the CDTD benchmarking pipeline against SMOTE, ARF, CTGAN, TVAE, TabDDPM, CoDi, TabSyn, and CDTD. Its initial average ranks are RMSE 3.6, F1 3.5, AUC 3.0, $p(\mathbf{x})=\prod_{i=1}^{D} p(x_i \mid x_{<i}).$8 distance 3.1, detection score 1.8, JSD 3.1, and Wasserstein distance 2.4. After penalizing methods with positive DCR-CDF integrals, TabularARGN becomes the strongest overall method, with privacy-adjusted average ranks of RMSE 2.4, F1 2.0, AUC 2.0, $p(\mathbf{x})=\prod_{i=1}^{D} p(x_i \mid x_{<i}).$9 distance 1.9, detection score 1.5, JSD 2.3, and Wasserstein distance 1.5; the paper states that it achieves five first-place rankings and two second-place rankings across the seven evaluated metrics. Dataset-specific examples include AUC utility of 0.912 on adult versus 0.915 on the training set, 0.943 on bank versus 0.947 on the training set, and 0.761 on default versus 0.768 on the training set; F1 utility of 0.796 on adult versus 0.797 on the training set; JSD values of 0.010 on bank, 0.021 on adult, and 0.006 on news; and Wasserstein values of 0.003 on adult, 0.002 on bank, and 0.006 on news. The same paper also performs a detailed privacy analysis on Adult using 1,000 records and 1,000 shadow models. Baseline TabularARGN with only dropout and early stopping yields most MIA AUCs between about 0.48 and 0.64, with Correlation Groundhog as the strongest attack at AUC 0.635 and accuracy about 60.1%. Rare-category and extreme-value protection reduce most AUCs to around 0.50–0.55 and drop CorrGroundhog to about 0.501; with DP at $p(\mathbf{x})=\prod_{i=1}^{D} p(x_i \mid x_{<i}),$0, all attacks are reported near chance, with AUC about 0.50–0.52 and accuracy about 50–51% (Sidorenko et al., 8 Aug 2025).

6. Relation to adjacent models and acknowledged limitations

TabularARGN belongs to a broader shift from one-shot implicit tabular generation toward explicit conditional structure. TGAN already demonstrated column-sequential generation with an LSTM generator, but remained an implicit GAN without tractable autoregressive likelihoods (Xu et al., 2018). CasTGAN pushed further toward feature-wise sequential generation with one generator per feature and a shared latent vector, yet still optimized a GAN objective rather than explicit conditional probabilities (Alshantti et al., 2023). TabMT rejected fixed-order factorization in favor of masked bidirectional conditioning and random iterative unmasking, thereby avoiding commitment to one directed decomposition (Gulati et al., 2023). TabDAR can be read as a more elaborate arbitrary-order autoregressive model in which categorical features use softmax heads while continuous conditionals are parameterized by diffusion rather than discretization (Zhang et al., 2024). Causal-TGAN and DATGAN instead impose graph-structured conditioning, effectively replacing full-prefix dependence with dependence on graph parents or expert-defined DAG structure (Wen et al., 2021, Lederrey et al., 2022). Against that background, TabularARGN is distinctive for combining explicit autoregressive factorization, permutation masking for any-order conditioning, and a deliberately shallow tabular-specific architecture rather than adversarial training, masked-transformer denoising, or diffusion-nested continuous heads.

Its limitations are correspondingly clear in the papers themselves. All variants depend on discretization, which may lose information and makes preprocessing a first-order design choice. The privacy-focused paper notes that larger numerical datasets can produce much larger models because some numerics are digit-split into many sub-columns, and that rare or high-cardinality categories can create privacy problems, particularly on the lending dataset, where the emp_title column causes DCR issues unless rare-category protection is enabled. It also states that exact handling of missing values and some decoding details are not specified. The broader framework likewise acknowledges that fairness-aware generation is underspecified, since no explicit fairness objective or benchmark is provided; that global conditional consistency across all orderings is not analyzed formally; and that autoregressive sequential sampling remains slower than fully parallel decoders even when practical runtime is strong. Both papers therefore present TabularARGN as a pragmatic and efficient autoregressive framework with strong empirical behavior, but not as a complete solution to missing-data semantics, high-cardinality rarity, or formal consistency and privacy guarantees (Sidorenko et al., 8 Aug 2025, Tiwald et al., 21 Jan 2025).