TabResFlow: Probabilistic Tabular Regression

• TabResFlow is a probabilistic regression model for univariate continuous targets that couples deep feature extraction with a conditional rational-quadratic spline flow, addressing overconfident predictions from Gaussian regressors.
• It integrates per-feature 2-layer MLP encoders, a ResNet backbone, and an analytic density model to achieve exact likelihood evaluation and efficient inversion without costly ODE computations.
• Empirical benchmarks show TabResFlow delivers top likelihood performance and faster inference, making it effective for uncertainty quantification and selective regression in practical applications.

TabResFlow is a probabilistic regression model for univariate continuous targets on tabular data that combines tabular-specific deep feature extraction with a conditional normalizing spline flow to learn flexible, non-Gaussian predictive distributions. It was introduced to address a recurrent limitation of point-estimation and Gaussian-head regressors on tabular tasks: they often yield overconfident predictions and can fail on skewed, heavy-tailed, or multi-modal targets. The model couples feature-wise encoders and a fully connected ResNet backbone with a conditional rational-quadratic spline normalizing flow, thereby retaining exact likelihood evaluation and analytic inversion while avoiding the ODE costs associated with continuous normalizing flows (Madhusudhanan et al., 23 Aug 2025).

1. Problem setting and statistical objective

TabResFlow is formulated for datasets of the form

$$(x,y)\sim p^\text{data}(x,y), \qquad x\in \mathcal{X}^F,\; y\in\mathbb{R},$$

where $x$ is a tabular feature vector with mixed numerical and categorical fields and $y$ is a single continuous target. The target of learning is the full conditional distribution $\hat p(y\mid x)$, not merely a point estimate $\hat y(x)$. Training quality is evaluated by expected negative log-likelihood (NLL), with the optimization objective described as minimizing $-\log \hat p(y\mid x)$ in expectation over the data distribution (Madhusudhanan et al., 23 Aug 2025).

The motivation is practical as well as statistical. Point predictors optimized under MSE or MAE return a single scalar for each input and therefore do not provide a principled abstention mechanism, explicit uncertainty quantification, or a direct basis for selective automation. The paper situates this as especially important in industrial automation, finance, used-car pricing, industrial quality prediction, customer behavior scoring, and healthcare risk scoring, where trustworthy decisions require uncertainty-aware outputs rather than uniform confidence.

A central criticism of conventional probabilistic tabular regressors is their reliance on a fixed parametric family such as a Gaussian, or on distribution families selected in advance as in NGBoost or PGBM. In that setting, the model predicts parameters such as $\mu(x)$ and $\sigma(x)$, but the family itself remains constrained. The paper emphasizes that real tabular targets may be non-Gaussian, skewed, heavy-tailed, or multi-modal, and specifically notes clearly multi-modal conditional distributions on the Protein and Wine datasets. Mixture-based alternatives mitigate some of this rigidity, but require a chosen number of mixture components and introduce training complexity and mode-dropping risk.

2. Architectural design and conditional spline flow

TabResFlow consists of three main components: feature-wise encoders for tabular inputs, a fully connected ResNet backbone, and a conditional rational-quadratic spline normalizing flow over the scalar target (Madhusudhanan et al., 23 Aug 2025).

Component	Form	Role
Feature encoders	2-layer MLPs for numerical features; embeddings for categorical features	Per-feature representation learning
Backbone	Fully connected ResNet	Nonlinear feature interaction extraction
Density model	Conditional RQ-NSF	Flexible tractable $p(y\mid x)$

For numerical inputs, each scalar feature $x_i$ is mapped through a 2-layer MLP,

$x$0

while categorical features are encoded by standard embedding matrices. The resulting feature embeddings are stacked into

$x$1

flattened into $x$2, and passed to the backbone.

The backbone is a fully connected ResNet. Each residual block has the form

$x$3

with

$x$4

and Dropout after each Linear. After $x$5 stacked blocks, the final hidden representation $x$6 serves as the conditioning vector for the flow.

The conditional density model is a Rational Quadratic Neural Spline Flow (RQ-NSF) specialized to the univariate case. The conditional change-of-variables formula is written as

$x$7

or equivalently through

$x$8

with base $x$9, typically $y$0. Because $y$1 is one-dimensional, the Jacobian determinant reduces to $y$2, which makes likelihood evaluation especially simple relative to multivariate flow architectures.

The spline transform is a monotonic rational-quadratic spline with strictly increasing knots $y$3 and positive derivatives $y$4. Its key properties are exactly those needed for conditional density regression: monotonicity, continuity and differentiability, analytic inverse, and analytic Jacobian. Spline parameters—bin widths, heights, and derivatives—are generated conditionally from the ResNet representation $y$5, with constraints such as softmax for widths and softplus for derivatives to maintain monotonicity. Multiple spline flow steps are stacked sequentially, with the number of steps tuned per dataset.

A defining design decision is the rejection of multivariate flow architectures such as RealNVP and MAF for this problem. The paper argues that those constructions are inherently multivariate and unsuitable when the regression target is scalar. It likewise contrasts TabResFlow with residual flows and continuous normalizing flows, which require numerical approximation or ODE solving and therefore impose higher computational cost.

3. Training objective, optimization, and uncertainty outputs

Training minimizes conditional NLL over a batch $y$6: $y$7 where

$y$8

Because the RQ-NSF transformation has a closed-form derivative, this objective is computed exactly, without numerical inversion, approximate Jacobians, or ODE integration during training or inference (Madhusudhanan et al., 23 Aug 2025).

The implementation details reported for TabResFlow include PyTorch, batch sizes up to 2048, and training on an NVIDIA GTX 1080 with 32 GB RAM. Hyperparameters tuned by Optuna include ResNet depth $y$9, embedding dimension $\hat p(y\mid x)$0, number of flow steps, number of spline bins, dropout rates, and learning rate. Early stopping is based on validation NLL.

The model’s probabilistic output supports several uncertainty summaries. The paper uses NLL as the primary proper scoring rule, reports CRPS for selected datasets in supplementary material, and evaluates selective regression through a newly introduced Area Under Risk Coverage (AURC) metric. For confidence ranking, the study considers inverse standard deviation for Gaussian baselines and both inverse standard deviation and inverse entropy for flow-based models. This places TabResFlow in the class of models intended not only for calibrated density estimation but also for downstream uncertainty-sensitive deployment.

4. Empirical results on benchmarks

On nine public benchmark datasets, TabResFlow is compared with Deep Ensembles, CatBoost, NGBoost, PGBM, TreeFlow, and NodeFlow. The reported likelihood results show that it achieves best or second-best NLL on 7/9 datasets, is best on 4 datasets, and attains an average NLL rank of 2.22, compared with 3.11 for TreeFlow and 3.67 for NodeFlow. The paper reports a 9.64% improvement over TreeFlow in NLL across benchmarks and highlights especially strong behavior on the multi-modal Protein and Wine datasets, where the model is described as accurately recovering multi-modal shapes (Madhusudhanan et al., 23 Aug 2025).

Although optimized for NLL rather than squared-error loss, TabResFlow remains competitive in point prediction. In the RMSE table, it is reported to be often close to or better than PGBM and CatBoost, and it ties PGBM for the largest number of RMSE wins across datasets. This suggests that improved conditional density modeling does not necessarily come at the cost of degraded point performance.

The runtime comparison with NodeFlow is one of the clearest efficiency claims. Average inference speed is reported as 5.6× faster than NodeFlow. On the large Year MSD dataset, the paper lists approximately 8211 s for NodeFlow and 561 s for TabResFlow, corresponding to about 14.6× speed-up. The stated explanation is architectural: NodeFlow relies on ODE solves in a CNF, whereas TabResFlow’s spline flow uses analytic inverse and derivative computations.

The ablation studies isolate the contribution of each design choice. A Gaussian-head version, TabResNet+Gauss, performs worse in NLL, especially on Naval, where NLL changes from -5.30 to 0.74; this is presented as evidence that the flexible flow layer is crucial. TabResFlow-Emb, which removes numerical MLP embeddings, is generally slightly worse than the full model, indicating that feature-wise numerical embeddings help. TabResFlow+CNF reaches similar NLL but is 2–3× slower at inference, supporting the spline-flow trade-off. A comparison of NodeFlow and TabResFlow+CNF further suggests that the ResNet backbone is a more efficient feature extractor than NODE in this setting.

5. Selective regression and the AURC metric

A major application domain in the paper is selective regression, in which the model predicts both a value and an uncertainty estimate, then abstains on high-uncertainty cases. The intended operational logic is explicit: a system may automatically act on low-risk inputs while escalating uncertain cases for manual review. In the paper’s running example, this is framed as used-car price prediction, where only sufficiently confident pricing decisions should be automated (Madhusudhanan et al., 23 Aug 2025).

The evaluation protocol uses risk-coverage curves. For a given test set, predictions are ranked by a confidence score; the system then computes risk among the most confident $\hat p(y\mid x)$1 of cases. The risk metric is MAPE for used-car pricing and RMSE for public datasets. For Gaussian baselines, confidence is defined as inverse standard deviation,

$\hat p(y\mid x)$2

For flow models, both inverse standard deviation and inverse entropy are examined.

To summarize the entire risk-coverage curve in a scalar, the paper introduces AURC: $\hat p(y\mid x)$3 where $\hat p(y\mid x)$4 is coverage and $\hat p(y\mid x)$5 is risk at coverage $\hat p(y\mid x)$6. Lower AURC indicates lower risk across coverage levels and therefore better selective behavior.

The used-car pricing case study contains 2.24M records with 64 features spanning numerical, categorical, and date fields, and uses chronological train/validation/test splits for Dec-2023, Jan-2024, and Feb-2024. Against CatBoost with virtual ensembles, NGBoost, ProbSAINT, and ProbResNet, TabResFlow achieves the best AURC in each month: 4.90 in Dec-2023, 4.73 in Jan-2024, and 4.54 in Feb-2024. For Feb-2024, the listed comparison values are CatBoost 5.13, NGBoost 5.10, ProbSAINT 4.69, ProbResNet 4.88, and TabResFlow 4.54. The accompanying risk-coverage plots are described as lying below those of the baselines, particularly at medium-high coverage, which the paper interprets as evidence of better uncertainty ordering.

6. Comparative context, naming ambiguity, and limitations

Within probabilistic tabular regression, TabResFlow is positioned against several model families. Relative to tree-based probabilistic methods such as NGBoost, PGBM, and CatBoost-based uncertainty estimation, its main distinction is that it does not restrict the target distribution to a fixed parametric family. Relative to TreeFlow, which uses CatBoost as encoder and a CNF for density modeling, TabResFlow is end-to-end trainable and avoids the need to pass an entire CatBoost ensemble to a CNF at inference. Relative to NodeFlow, it retains end-to-end learning but replaces the NODE encoder and CNF with a ResNet backbone and spline flow, thereby removing ODE-solve costs. TreeFlow itself belongs to a broader line of conditional density learners based on tabular flows and can also address multivariate outcomes, but its architecture is based on covariate-dependent tree-CDF transforms rather than a ResNet-conditioned spline flow (Wang et al., 2024).

A separate source of confusion is terminological. In later literature on synthetic tabular data generation, the phrase “tabular latent flow” or even “TabResFlow” has sometimes been used informally for latent-space flow-matching systems. One empirical study explicitly states that it does not introduce a single new algorithm called “TabResFlow”, but instead studies latent flow models for tabular synthesis under different objectives, paths, and samplers (Nasution, 18 Jun 2026). This matters because the 2025 paper named "TabResFlow" refers specifically to probabilistic univariate tabular regression, not to generic tabular flow matching or synthetic table generation.

The limitations reported for TabResFlow are correspondingly specific. The model, as presented, supports univariate targets only; multivariate extension is described as possible but is not studied empirically. The uncertainty ranking analysis relies primarily on inverse standard deviation, with entropy-based scoring noted as an area for deeper investigation. The paper also notes that systematic scaling analysis across dataset size and feature complexity is incomplete, that performance depends on careful tuning of flow steps, ResNet size, and embeddings, and that deep learning training may be heavier than gradient-boosted trees on small tabular datasets.

The future directions identified by the authors include multivariate TabResFlow, more extensive study of uncertainty metrics and their relationship to NLL, CRPS, AURC, and business metrics, integration with active learning and more elaborate selective prediction policies, and robustness analyses under dataset size variation, noise, covariate shift, and out-of-distribution conditions. Taken together, these directions suggest that TabResFlow is best understood not as a general tabular flow framework, but as a specific design point in probabilistic tabular regression: a ResNet-conditioned, spline-based conditional density estimator optimized for flexible univariate uncertainty modeling and selective deployment (Madhusudhanan et al., 23 Aug 2025).