One Table, Many Queries: Benchmarking Generalist Tabular Models Across Generation, Imputation, and Prediction
Organizations: StableAI · Beijing University of Posts and Telecommunications · Southeast University · Tsinghua University
Abstract
Generation, imputation, and prediction query different aspects of the same tabular distribution, but are usually evaluated separately. This leaves unclear which capabilities they share, how their training objectives interact, and whether one model can perform all three well. We introduce the One Table Benchmark (\ours{}), spanning 56 datasets, 46 registered methods, and 20 capability axes under aligned data splits and explicit task contracts. We compare task ranks on aligned support, train five backbones under all seven non-empty objective combinations on five datasets with three seeds, and contrast fixed-model profiles with specialist portfolios. Rank comparisons connect imputation and prediction through numerical recovery and dependence preservation, and link generation to both through predictive utility; distributional fidelity follows different, sometimes opposing, rankings. Objective interventions show the largest pooled gains from including the recipient's own objective. Generation has the broadest coverage of positive pooled auxiliary effects, while prediction shows potential to improve imputation and synthetic-data utility, with responses that vary across backbones. Cross-addition effects reveal that objective benefits need not add and depend on the objectives already present. Fixed-model comparisons expose corresponding utility--fidelity trade-offs: some generalists surpass specialist portfolios on utility, while fidelity continues to favor specialization. Frozen foundation models show strong prediction and numerical point imputation, together with promising synthetic-data utility, highlighting their potential across all three tasks. These findings motivate unified tabular models that share useful conditional information while improving categorical reconstruction and faithful joint sampling.
Figures & tables
| Task | Observed | Queried | Target distribution |
| Generation | |||
| Imputation | |||
| Prediction |
| Cohort | Metrics | Clusters | PCs | PCA mass (PC1) |
| Generation | 29 | 17 | 5 | 80% (53%) |
| Imputation | 10 | 8 | 3 | 74% (41%) |
| Prediction (cls.) | 10 | 4 | 2 | 88% (75%) |
| Prediction (reg.) | 5 | 2 | 1 | 86% (86%) |
| Task | Retained axes (member metrics) |
| G | Fidelity : numerical_ks_error , shape_error , one_way_error , two_way_error , three_way_error ; C2ST : gbdt_c2st_auroc , logistic_c2st_auroc ; -precision : alpha_precision ; association : mixed_association_error , mixed_association_similarity ; TSTR gap : tstr_gap ; global-utility compromise : global_utility_ratio , beta_recall_coverage , authenticity ; privacy-risk diagnostics : mean_min_euclidean_dcr , normalized_median_gower_dcr , distance_mia_auroc , distance_mia_tpr_at_1pct_fpr , delta_presence . |
| I | numerical point accuracy : macro_nmae , macro_nrmse ; categorical point accuracy : categorical_balanced_accuracy , categorical_macro_f1 ; categorical distribution : categorical_jsd ; numerical distribution : numeric_wasserstein1 ; association : association_matrix_error ; downstream utility gap : utility_gap . |
| P (cls.) | hard-decision performance : accuracy , balanced_accuracy , macro_f1 , mcc ; proper scores : brier_score , log_loss ; ranking/selective risk : roc_auc , average_precision , aurc ; calibration : ece . |
| P (reg.) | squared-error family : r2 , nrmse , rmse ; mean absolute error : mae ; median error : median_absolute_error . |
Appendix figures & tables44 assets
Supplementary material from the paper’s appendix.
Appendix
| Category | Literature sources |
| Imputation and cross-task studies | Shades-of-Null ( Khan et al., 2025 ) ; TabImpute ( Feitelberg et al., 2025 ) ; ForestDiffusion ( Jolicoeur-Martineau et al., 2024 ) ; DiffPuter ( Zhang et al., 2025 ) ; TabCSDI ( Zheng and Charoenphakdee, 2022 ) ; MissDiff ( Ouyang et al., 2023 ) ; ReMasker ( Du et al., 2024 ) ; HyperImpute ( Jarrett et al., 2022 ) ; TabGenDDPM ( Villaizán-Vallelado et al., 2025 ) . |
| Generation and structural evaluation | TabStruct ( Jiang et al., 2026 ) ; Systematic Assessment of Tabular Data Synthesis ( Du and Li, 2025 ) ; Synthcity ( Qian et al., 2023 ) ; Structured Evaluation of Synthetic Tabular Data ( Yang et al., 2024 ) ; CauTabBench ( Senane et al., 2024 ) ; SynthEval ( Lautrup et al., 2025 ) ; TabDDPM ( Kotelnikov et al., 2023 ) ; TabDiff ( Shi et al., 2025 ) . |
| Prediction and robustness | TabArena ( Erickson et al., 2025 ) ; TALENT ( Liu et al., 2025 ; Ye et al., 2024 ) ; TabZilla ( McElfresh et al., 2023 ) ; Why Trees ( Grinsztajn et al., 2022 ) ; TabReD ( Rubachev et al., 2025 ) ; CARD ( Han et al., 2022 ) ; TabularBench ( Simonetto et al., 2024 ) . |
| Group | Original | Retained | Retained meta-features |
|---|---|---|---|
| A | 9 | 9 | cat_feature_ratio ; frac_binary_features ; frac_numeric_lowcard ; log_attr_to_inst ; log_n_features ; log_n_samples ; max_log_cat_cardinality ; mean_log_cat_cardinality ; mean_numeric_unique_ratio . |
| B | 68 | 29 | duplicate_row_ratio ; frac_bounded_01 ; frac_continuous_cols ; frac_heavy_tail ; frac_integer_valued ; frac_multimodal ; frac_nonneg ; frac_outlier_rows ; frac_skewed ; hopkins ; kmeans_silhouette_max ; mean_n_modes ; pymfe_attr_conc.mean ; pymfe_attr_conc.sd ; pymfe_attr_ent.mean ; pymfe_attr_ent.sd ; pymfe_cor.mean ; pymfe_cor.sd ; pymfe_frac_bin ; pymfe_frac_norm ; pymfe_frac_outliers ; pymfe_kurtosis.mean ; pymfe_kurtosis.sd ; pymfe_nr_cor_attr ; pymfe_skewness.mean ; pymfe_skewness.sd ; pymfe_sparsity.mean ; pymfe_sparsity.sd ; zero_ratio . |
| C | 16 | 16 | cat_num_association ; frac_cols_with_missing ; frac_rows_with_missing ; frac_strong_cor_pairs ; max_abs_cor ; max_col_missing_ratio ; max_norm_mutual_info ; mean_abs_cor ; mean_missing_per_missing_row ; mean_norm_mutual_info ; missing_cell_ratio ; missing_mar_score ; missing_pattern_count_log ; missing_pattern_entropy ; pca_dims_95_ratio ; top_eigen_ratio . |
| D | 20 | 18 | class_entropy_norm ; class_imbalance_log ; is_binary ; is_multiclass ; is_regression ; lm_best ; lm_knn ; lm_linear ; lm_rf ; lm_tree ; log_n_classes ; minority_class_ratio ; n_targets_log ; noise_level ; nonlinearity_gain ; target_skew_abs ; target_unique_ratio ; tree_vs_knn . |
| Total | 113 | 72 | – |
| Dataset | Source | OML ID | Task | Num. | Cat. | ID | Cls. | Class min | Class max | |||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| adult-census | OpenML | 1119 | Cls. | 32,561 | 15 | 2170.73 | 6 | 8 | 1 | 2 | 7,841 | 24,720 |
| allbp | OpenML | 40707 | Cls. | 3,772 | 29 | 130.07 | 6 | 23 | 0 | 3 | 14 | 3,609 |
| Amazon_employee_access | OpenML | 46905 | Cls. | 32,769 | 9 | 3641.00 | 0 | 9 | 0 | 2 | 1,897 | 30,872 |
| appendicitis_test | OpenML | 45589 | Cls. | 106 | 7 | 15.14 | 7 | 0 | 0 | 2 | 21 | 85 |
| Australian | OpenML | 40981 | Cls. | 690 | 14 | 49.29 | 6 | 8 | 0 | 2 | 307 | 383 |
| auto_price | OpenML | 745 | Cls. | 159 | 15 | 10.60 | 14 | 1 | 0 | 2 | 54 | 105 |
| Methods | G | I | P |
|---|---|---|---|
| ForestDiffusion-VP (joint) ( Jolicoeur-Martineau et al., 2024 ) ; Gaussian copula ( Patki et al., 2016 ) ; GReaT ( Borisov et al., 2023 ) ; LimiX-2 ( Zhang and others, 2026 ) ; TabDDPM ( Kotelnikov et al., 2023 ) ; TabDiff ( Shi et al., 2025 ) ; TabICLv2 ( Qu et al., 2026 ) ; TabPFN-3 + unsupervised ( Grinsztajn and others, 2026 ; Prior Labs, 2026 ) ; TabSyn ( Zhang et al., 2024 ) . | |||
| ForestDiffusion-VP (label-conditioned) ( Jolicoeur-Martineau et al., 2024 ) . | |||
| ARF ( Watson et al., 2023 ) ; Bayesian network ( Qian et al., 2023 ) ; CTGAN ( Xu et al., 2019 ) ; Independent marginals; NFlow ( Durkan et al., 2019 ; Qian et al., 2023 ) ; NRGBoost ( Bravo, 2025 ) ; SCM (DirectLiNGAM) ( Shimizu et al., 2011 ) ; SMOTE ( Chawla et al., 2002 ) ; SMOTE-NC ( Chawla et al., 2002 ) ; SMOTER-style interpolation ( Torgo et al., 2013 ) ; TabEBM ( Margeloiu et al., 2024 ) ; Training-row bootstrap ( Efron, 1979 ) ; TVAE ( Xu et al., 2019 ) . | |||
| DiffPuter ( Zhang et al., 2025 ) ; HyperImpute ( Jarrett et al., 2022 ) ; kNN imputation ( Troyanskaya et al., 2001 ; Pedregosa et al., 2011 ) ; Mean/mode imputation ( Pedregosa et al., 2011 ) ; Median/mode imputation ( Pedregosa et al., 2011 ) ; MICE ( van Buuren and Groothuis-Oudshoorn, 2011 ) ; MissForest ( Stekhoven and Bühlmann, 2012 ) ; ReMasker ( Du et al., 2024 ) ; TabCSDI ( Zheng and Charoenphakdee, 2022 ) ; TabImpute ( Feitelberg et al., 2025 ) . | |||
| CARD ( Han et al., 2022 ) ; CatBoost ( Prokhorenkova et al., 2018 ) ; FT-Transformer ( Gorishniy et al., 2021 ) ; LightGBM ( Ke et al., 2017 ) ; LightGBM-Quantile ( Ke et al., 2017 ) ; Logistic regression ( Pedregosa et al., 2011 ) ; MLP ( Pedregosa et al., 2011 ) ; NGBoost ( Duan et al., 2020 ) ; Random forest ( Breiman, 2001 ) ; Ridge regression ( Hoerl and Kennard, 1970 ; Pedregosa et al., 2011 ) ; TabM ( Gorishniy et al., 2025 ) ; XGBoost ( Chen and Guestrin, 2016 ) . |
| Metric (exact leaf) | Core meaning |
|---|---|
| generation.fidelity.* ( SDV, 2026 ) | |
| categorical_tvd_error | Mismatch in categorical marginal distributions. |
| mixed_association_error | Mismatch in pairwise association strengths across mixed column types. |
| numerical_ks_error | Mismatch in numerical marginal distributions. |
| pairwise_trend_error | Mismatch in signed numerical trends and categorical joint distributions. |
| shape_error | Overall numerical and categorical marginal mismatch. |
| Metric (exact leaf) | Core meaning |
|---|---|
| imputation.point.* | |
| categorical_balanced_accuracy | Class-balanced accuracy of recovered categorical values. |
| categorical_macro_f1 | Class-balanced precision and recall of recovered categorical values. |
| macro_nmae | Normalized absolute error of numerical reconstruction. |
| macro_nrmse | Normalized numerical reconstruction error, emphasizing larger errors. |
| imputation.structure.* | |
| Metric (exact leaf) | Core meaning |
|---|---|
| prediction.classification.* | |
| accuracy | Fraction of correctly classified examples. |
| average_precision | Precision-recall quality of probability-based class rankings. |
| balanced_accuracy | Classification accuracy weighted equally across classes. |
| macro_f1 | Balance of precision and recall, weighted equally across classes. |
| mcc | Correlation between predicted and true class labels. |
| Task pair | Candidate pool | Datasets | Range of | Candidate methods |
| I P-cls | 9 | 13–38 | 92–287 | Forest-VP Joint, Gaussian Copula, GReaT, LimiX-2, TabDDPM, TabDiff, TabICLv2, TabPFN-3, TabSyn |
| I P-reg | 9 | 6–18 | 44–139 | Forest-VP Joint, Gaussian Copula, GReaT, LimiX-2, TabDDPM, TabDiff, TabICLv2, TabPFN-3, TabSyn |
| G I | 10 | 12–56 | 98–482 | Forest-VP, Forest-VP Joint, Gaussian Copula, GReaT, LimiX-2, TabDDPM, TabDiff, TabICLv2, TabPFN-3, TabSyn |
| G P-cls | 9 | 30–38 | 246–300 | Forest-VP Joint, Gaussian Copula, GReaT, LimiX-2, TabDDPM, TabDiff, TabICLv2, TabPFN-3, TabSyn |
| G P-reg | 9 | 18 | 138–141 | Forest-VP Joint, Gaussian Copula, GReaT, LimiX-2, TabDDPM, TabDiff, TabICLv2, TabPFN-3, TabSyn |
| Task profile | Exact frontier | frontier |
| Generation, six specified axes | ARF | ARF |
| Imputation, numerical targets | MissForest ; DiffPuter ; KNN Imputer ; MICE | MissForest ; DiffPuter ; KNN Imputer |
| Imputation, categorical targets | MissForest ; MICE ; DiffPuter ; KNN Imputer ; ReMasker | MissForest ; MICE ; DiffPuter |
| Prediction, classification | CatBoost; TabM | CatBoost; TabM |
| Prediction, regression | LightGBM; LightGBM Quantile; Random Forest | LightGBM Quantile; Random Forest |
| Method | Configuration |
|---|---|
| Methods with generation or conditional-query interfaces | |
| TabDDPM | MLP widths ; time embedding 128; 1,000 diffusion steps; cosine schedule; 100,000 updates, LR , WD , ; quantile normalization. |
| TabDiff | Two layers, token width 4, one head, factor 32, time width 1,024; 50 sampling steps; 8,000 training epochs, LR , EMA 0.997, ; main and target-only prediction-guidance models. |
| TabSyn | Two-layer, width-4 VAE tokens; latent diffusion width 1,024; VAE 4,000 epochs and diffusion 10,001 epochs, LR , , patience 500; 50 sampling steps, 20 imputation resampling rounds. |
| Forest-VP / Joint | 50 time points, 100 noisy duplicates, 100 trees, depth 7, shrinkage 0.3; histogram trees; RePaint , ; inference row batches 256 (I) and 1,024 (G). |
| Gaussian copula | At most 1,000 marginal quantiles; 100 conditional samples for regression and 200 class-conditional samples. |
| Dataset | OpenML ID | Type | Features | Train | Test |
| JapaneseVowels | 976 | C | 14 | 7,968 | 1,993 |
| MiamiHousing2016 | 43093 | R | 16 | 11,145 | 2,787 |
| cleveland | 786 | C | 13 | 242 | 61 |
| credit_card_clients_default | 46919 | C | 23 | 24,000 | 6,000 |
| website_phishing | 46963 | C | 9 | 1,082 | 271 |
| Backbone | Architecture and loss | Optimizer and configured budget |
| Gaussian Copula | Empirical marginals with at most 1,000 stored quantiles; joint/conditional Gaussian likelihood for the selected query. | Adam, LR , batch 1,024, 2,000 updates; gradient norm clipped at 10. |
| Forest-VP Joint | 50 diffusion times; 100 duplicated exposures; query-masked multi-output squared-error trees. | 100 boosting rounds per time, depth 7, shrinkage , histogram trees; one job and one XGBoost thread per worker. |
| TabDDPM | MLP widths , time embedding 128, no dropout; 1,000 diffusion times, cosine noise schedule; masked numerical MSE plus multinomial loss. | AdamW, LR with linear decay, WD , batch 4,096, 100,000 updates. |
| TabDiff | Two layers, token width 4, one head, factor 32, time width 1,024; learnable power-mean numerical and log-linear categorical schedules. Discrete loss weight 1; continuous weight at epoch . | AdamW, fixed LR , WD 0, batch 4,096, 8,000 epochs; EMA and final EMA checkpoint; no separate target-only guidance model. |
| TabSyn | Shared VAE: two layers, token width 4, one head, factor 32; reconstruction plus -KL. Latent EDM: width 1,024 and query-masked loss. | Adam, LR , WD 0, batch 4,096; VAE 4,000 epochs; latent diffusion 10,001 epochs at fixed LR, saving the final model. |
Tabular Foundation Models
Tabular foundation models learn reusable representations or in-context prediction skills for structured data such as spreadsheets and databases.
See all 209 papers