Research map / Image generation / Autoregressive and latent generation

BSTabDiff: Block-Subunit Diffusion Priors for High-Dimensional Tabular Data Generation

NeurIPS 2026 · Image generation › Autoregressive and latent generation · approach: Latent diffusion with learned tokenizers · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes BSTabDiff, a block-subunit generator for high-dimensional low-sample-size tabular data that learns a diffusion or flow prior over M low-dimensional block latents and decodes to features via copula emissions, marginals and missingness masks, evaluated by ML efficiency against GAN, VAE and diffusion tabular baselines on eight omics-style datasets.

Problem

Generating realistic synthetic tabular data in the high-dimensional low-sample-size regime (n much smaller than m), where dense dependence learning in the full feature space is ill-conditioned and features show heavy tails, heteroscedasticity and structured missingness.

Approach

Partition features into M latent blocks, generate each block from a shared low-dimensional subunit latent drawn from a diffusion/flow prior, and decode features with Gaussian-copula emissions and learned inverse marginal CDFs, so global dependence is learned in R^M rather than R^m.

Key results

• On eight HDLSS datasets with logistic regression MLE, BSTabDiff reports the best mean accuracy among synthetic methods on all eight, e.g. 95.96% on LNG vs 96.54% TRTR and 72.08% on SMK vs 72.34% TRTR.
• On Colon with TANDEM, TabPFN-2.5 and CatBoost, BSTabDiff synthetic data gives accuracy and AUC comparable to real-data training, e.g. TabPFN-2.5 AUC 89.25 vs 90.53 TRTR.
• Fidelity diagnostics show C2ST near chance on COL, AML and GLI but much higher on LNG (ACC 0.74).

Surprising findings

• BSTabDiff's synthetic data sometimes yields higher downstream accuracy than real data on some datasets, e.g. TANDEM accuracy 81.97 vs 79.44 on Colon, which the authors interpret as utility transfer rather than overfitting.
• The paper states that the reported gains hold even though the model is trained with a single fixed configuration and little per-dataset tuning, with only the block size M changed per dataset.

Limitations

• Some multi-class datasets such as LNG are harder to match closely, with C2ST strongly above chance.
• Only numerical features without missing values are evaluated, so the missingness and mixed-type components are not tested empirically.
• The identifiability and sample-complexity results are stated as informal or proof sketches, and the paper treats feature permutation as fixed at identity during training.

Experimental setup

Datasets: Colon GLI-85 Lung Prostate SMK TOX171 ALLAML Arcene

Benchmarks: Colon GLI-85 Lung Prostate SMK TOX171 ALLAML Arcene

Metrics: accuracy (MLE with logistic regression, TSTR vs TRTR) AUC KS Wasserstein-1 C2ST correlation discrepancy

Baselines: SMOTE CTGAN CTAB-GAN CTAB-GAN+ TVAE TabDDPM TabDiff BinaryDiffusion ForestDiffusion

Builds on: TabDDPM TabDiff CTGAN

Closest papers by meaning

• TabDiff: a Mixed-type Diffusion Model for Tabular Data Generation (ICLR 2025 · similarity 0.87)
• BGDB: Bernoulli-Gaussian Decision Block with Improved Denoising Diffusion Probabilistic Models (AAAI 2025 · similarity 0.85)
• Continuous Diffusion for Mixed-Type Tabular Data (ICLR 2025 · similarity 0.85)
• Tabular Foundation Model for Generative Modelling (NeurIPS 2026 · similarity 0.85)
• TabPrep: Closing the Feature Engineering Gap in Tabular Benchmarks (NeurIPS 2026 · similarity 0.84)
• TabArena: A Living Benchmark for Machine Learning on Tabular Data (NeurIPS 2025 · similarity 0.84)