One Table Unifies Evaluation of Tabular Generation, Imputation, and Prediction

Summary:The One Table Benchmark evaluates tabular generation, missing-value imputation, and supervised prediction under aligned data splits and explicit task contracts. Covering 56 datasets, 46 registered methods, 20 capability axes, and controlled multi-objective interventions across five backbones, it shows that the tasks share useful conditional information but impose conflicting rankings: downstream utility and distributional fidelity are not interchangeable. The results also demonstrate that auxil

---

Core Finding: One Table Does Not Imply One Capability

The One Table Benchmark places three central tabular tasks under one evaluation framework: generation, which samples or synthesizes tabular distributions; imputation, which reconstructs missing cells; and prediction, which infers a target from observed features. It covers 56 datasets, 46 registered methods, and 20 capability axes, using aligned data splits and explicit task contracts to reduce inconsistencies in preprocessing, missingness handling, and metric definitions.

The central objective is not to identify one model that dominates every metric. Instead, the benchmark separates data utility from distributional fidelity. Utility asks whether generated or imputed tables remain useful for downstream prediction. Fidelity asks whether marginal distributions, feature dependencies, and categorical structure are preserved faithfully. These rankings can conflict: a model may synthesize samples that improve analytical workflows while fitting the joint table distribution less accurately.

How the Framework Connects the Tasks

Cross-task comparisons rely on ranks over aligned support:

• Imputation and prediction are connected through numerical recovery and preservation of variable dependencies.
• Generation and prediction are linked through the predictive utility of synthetic data: whether generated rows improve downstream training or evaluation.
• Generation and imputation jointly expose categorical reconstruction and joint-distribution fidelity.

This is preferable to naively averaging heterogeneous scores. Tasks differ in scale, difficulty, and data sparsity, so a raw mean can conceal capability structure. Rank-based comparisons reveal where models transfer information and where they enter conflict.

Aligned support does not make the tasks information-theoretically identical. Imputation may use random cell missingness, whole-column missingness, or domain-specific mechanisms, while prediction may expose a different feature set and target column. A unified protocol reduces evaluator inconsistency, but task contracts and missingness assumptions still need to be inspected before transferring rankings to production.

Multi-Objective Training in Practice

Across five datasets, five backbones, and three random seeds, the study trains all seven non-empty combinations of generation, imputation, and prediction objectives. It asks whether auxiliary objectives improve the model’s intended task and whether gains from different objectives are additive.

The results rule out a simple “more objectives are better” rule:

1. Including a task’s own native objective produces the largest pooled gains, indicating that direct optimization of the relevant conditional distribution or conditional mean remains essential.
2. Generation has the broadest coverage of positive pooled auxiliary effects, suggesting that learned joint structures transfer to several tabular capabilities.
3. Prediction can improve imputation and synthetic-data utility for some backbones, revealing a route from supervised signal to useful conditional representations.
4. Cross-addition benefits are not additive. The effect of an auxiliary objective depends on which objectives are already present and can vary by backbone.

For engineering teams, multi-objective training should therefore be monitored with task-specific metrics rather than a single aggregate score. Otherwise, stronger generation performance can hide degraded prediction, or a mean-based objective can conceal poor joint-distribution modeling.

Generalists Versus Specialist Portfolios

Comparisons between fixed models and specialist portfolios expose a consistent utility-fidelity trade-off. Some generalist models outperform task-specific portfolios on utility-oriented metrics, while specialization remains advantageous for fidelity.

The mechanism follows from the learning objectives. Prediction favors conditional means and stable decision boundaries, which are useful for classification and regression but may reduce within-class diversity. Generation must instead represent joint structure and row-level dependencies, which a single prediction loss cannot fully constrain. A shared model can reuse conditional information across tasks, but it may also compromise between several distribution objectives.

The practical value of unification is therefore broader than reducing model count. Shared representations can reduce repeated labeling and pretraining cost, simplify task switching, and expose one interface for training and inference. The costs are more complex loss balancing, harder attribution of capability regressions, and additional tuning to shift priorities between business tasks.

What Frozen Foundation Models Enable

Frozen tabular foundation models show strong performance in prediction, numerical point imputation, and the downstream utility of synthetic data. This suggests that general pretraining can transfer structural knowledge across datasets more effectively than learning from scratch when labels are scarce or data distributions shift frequently.

Freezing does not remove adaptation costs. Table schemas, type systems, and missingness semantics remain domain-specific. Deployment may require retrieval, example selection, output-head adaptation, or light fine-tuning. It also constrains how directly the model can satisfy a particular task contract; prompt changes alone may not repair categorical reconstruction errors or biased joint sampling.

Reading the Results Critically

Five datasets and three random seeds make the objective interventions stronger than a single-dataset ablation, but the abstract does not disclose dataset difficulty, backbone architectures, training budgets, or statistical significance details. Tabular results are also sensitive to missingness mechanisms, preprocessing, categorical encoding, and hyperparameter-search budgets.

Although 46 registered methods provide broad baseline coverage, unequal implementation quality, tuning effort, or access to pretraining data can still affect rankings. Reproduction should check for train/test leakage, especially leakage of missing values during pretraining; whether specialists received comparable search budgets; and whether fidelity metrics cover marginals, cross-feature dependencies, and categorical support—not only average prediction error.

Deployment Guidance

Teams evaluating a unified tabular model should track three layers of metrics:

• Task utility: predictive accuracy and calibration, downstream gains from imputation, and the effect of synthetic data during training.
• Distributional fidelity: per-column marginals, joint dependencies, categorical coverage, and generation of rare combinations.
• Operations: training and inference throughput, memory use, data-ingestion overhead, and the latency of sending generated rows into analytical pipelines.

For prediction-centered applications, start with the prediction objective and test whether adding generation improves imputation or data augmentation. For privacy-preserving synthesis or model training, do not select a model from prediction scores alone; inspect joint-sampling fidelity. Frozen foundation models are useful for rapid baselines and small-scale domain adaptation, multi-objective generalists for shared interfaces and representations, and specialist portfolios where distributional fidelity remains the priority.

---

Source: ArXiv Machine Learning (cs.LG) (2026-11-11)

— END —

Tags: #Tabular Foundation Models Follow #Unified Benchmark Follow #Tabular Generation Follow #Missing-Value Imputation Follow #Multi-Objective Learning Follow