acceptodds
Under review as a conference paper at ICLR 2027

TabTreeVFM: Endpoint Variational Flow Matching with Boosted Trees for Mixed-Type Tabular Generation

Abstract

Synthetic tabular data supports downstream analytics, machine learning, and data sharing when real data is scarce, sensitive, or costly to collect. Mixed-type synthesis must capture joint dependencies while respecting heterogeneous column structure, yet high-fidelity generators can reproduce training records rather than learn the underlying distribution. Gradient-boosted trees suit such heterogeneous data, but existing tree-based flow models regress ambient-space velocities, imposing a Euclidean geometry on inherently multinomial categorical variables. To address this, we propose TabTreeVFM, a tree-boosted endpoint-posterior model with Gaussian regression heads for continuous columns and multinomial classifier heads for categorical variables, in place of direct velocity regression. The resulting posterior means recover the flow-matching velocity analytically, and training decomposes into independent XGBoost subproblems across timesteps and feature groups. Across 10 standard mixed-type benchmarks, TabTreeVFM attains the best overall average rank among 12 generators across fidelity, utility, and empirical privacy. Moreover, we introduce Mem, a held-out-calibrated measure of excess full-row reproduction, and find that several strong neural baselines reproduce training rows substantially above the natural collision rate, while TabTreeVFM remains close to the held-out reference. Together, these results show that type-aware endpoint prediction combines the tabular inductive bias of boosted trees with principled flow-matching geometry, yielding strong fidelity and utility alongside a favorable empirical privacy and memorization profile.

Primary area:
generative models
Keywords:
tabular data generation, flow matching, mixed-type data, synthetic data

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_yRXoOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper takes the variational flow matching (VFM) framework and changes the neural network backbone to a tree-based gradient boosting model. Using VFM allows a reparameterization of the denoising objective, so the endpoint, instead of the velocity, can be predicted. The authors argue that this is a more suitable objective for tree-based models.

Critical strengths

The authors include many benchmarks and 10 different datasets in their experiments. They include extensive additional experiments about hyperparameter sensitivity and scaling in the appendix. Some insight is provided into why the endpoint prediction objective is better for tree-based model via likelihood geometry argument.

Critical weaknesses

The contribution of the paper is really marginal. The authors only change the backbone used in EF-VFM [1], keeping the entire rest of the generative model unchanged. This shows in section 3.2, which is just a reframing of the content in [1]. The idea of using a gradient-boosting model as the backbone is not novel either, as many models, such as ForestDiffusion [2] exist. The proposed memorization metric is of little value, as DCR-based metrics suffice and capture the same information. A threshold-based DCR for different chosen threshold could as be computed. Since the only contribution is the switch to a tree-based model, I would expected a more thorough comparison to other tree-based methods, instead of only Forest-Flow. Lastly, the proposed model, similar to Forest-Flow requires substantial memory. Currently, this fact is hidden in the appendix. It also implies that presented results only hold for small to medium-sized datasets, as the lack of scalability would not allow the generation of higher-dimensional datasets.


[1] Exponential Family Variational Flow Matching for Tabular Data Generation. https://arxiv.org/abs/2506.05940

[2] Generating and Imputing Tabular Data via Diffusion and Flow-based Gradient-Boosted Trees. https://arxiv.org/abs/2309.09968

Questions for authors
  • I do not understand the added value of the proposed memorization metric over the popular Distance-to-closest-record (DCR) or DCR Share, e.g., as explained in https://arxiv.org/pdf/2504.01908. Instead of L2 distance, you are just switching to L1 distance with a difficult-to-choose threshold parameter. You could also report DCR@1%, which would have the same meaning?
  • The training times for TabDiff and Tabbyflow seems very excessive for some datasets. I do understand that CDTD is particularly efficient, but it is still a diffusion model. Why do the other models train for more than 100 minutes on the adult datasets whereas for CDTD 3 minutes are enough? Are you training CDTD for long enough to converge to a good state? The high DCR values seem like overfitting to me?
  • Should the component ablation in Table 2 not start from EF-VFM (TabbyFlow) model? Since you keep the specification the same and switch out the backbone?
  • Appendix L lacks a deeper comparison to ForestDiffusion: What explains the time differences between ForestFlow and TabTreeVFM?
  • How does the sample quality scale with method complexity? Mentioning that setting K=10 improves training time and memory scaling is meaningless if sample quality drops significantly relative to the other methods.
  • Why do you use different samplers across the diffusion baselines? Would it not be a fairer benchmark if you used the Euler sampler for all?
7 comments
AuthorsOpenReview ↗

Response to Reviewer yRXo (Part 1 / 5)

W:

The contribution of the paper is really marginal. The authors only change the backbone used in EF-VFM [1], keeping the entire rest of the generative model unchanged. This shows in section 3.2, which is just a reframing of the content in [1]. The idea of using a gradient-boosting model as the backbone is not novel either, as many models, such as ForestDiffusion [2] exist. The proposed memorization metric is of little value, as DCR-based metrics suffice and capture the same information. A threshold-based DCR for different chosen threshold could as be computed. Since the only contribution is the switch to a tree-based model, I would expected a more thorough comparison to other tree-based methods, instead of only Forest-Flow. Lastly, the proposed model, similar to Forest-Flow requires substantial memory. Currently, this fact is hidden in the appendix. It also implies that presented results only hold for small to medium-sized datasets, as the lack of scalability would not allow the generation of higher-dimensional datasets.

Response to W:

Thank you for the detailed comments. We address each point below.

Contribution beyond EF-VFM, ForestDiffusion, and Forest-Flow. Section 3.2 presents the endpoint formulation of EF-VFM, on which TabTreeVFM builds; our contributions are in Sections 3.3–3.6. Moving this formulation to boosted trees changes how the objective is optimized, not only the function class. EF-VFM trains a single time-conditioned network on all timesteps and columns jointly with Adam, a first-order method that rescales parameter gradients by running moment estimates; it notes that the Fisher information of its endpoint likelihoods is the Hessian of the log-partition function, but leaves natural-gradient optimization to future work. TabTreeVFM instead decomposes the objective into independent XGBoost regression and multiclass heads at each timestep, whose posterior means define the velocity analytically, and fits each head by second-order boosting, whose leaf updates use the per-sample curvature of the endpoint likelihood. This design also differs from ForestDiffusion and Forest-Flow, which regress scores or velocities coordinate-wise on one-hot encodings. Under the same probability path, solver, and preprocessing, the EF-VFM network reaches a C2ST of 78.10 and tree-based velocity regression 78.62, whereas TabTreeVFM reaches 85.26 (Table 2 and Section 4.4; see Q3).

For the categorical heads, this curvature is proportional to the diagonal Fisher information of the endpoint posterior, and Proposition 3.1 shows that the leaf update equals the diagonal-Fisher step at the leaf-mean prediction rescaled by 1/(1−ρjk)1/(1-\rho_{jk}), where ρjk\rho_{jk} is the within-leaf dispersion ratio, so the two coincide when the predictions within a leaf agree. This curvature matters in practice (Appendix D): across the 46 non-binary categorical heads of nine datasets, a constant-curvature update raises held-out endpoint cross-entropy roughly a hundredfold for t≥0.90t\geq0.90 and worsens the categorical fidelity of generated tables, whereas the full Fisher matrix of natural-gradient boosting brings no consistent gain. The Fisher scaling thus comes from standard second-order boosting at no extra cost, without NGBoost's explicit Fisher inversion.

Comparison with other tree-based generators. Beyond Forest-Flow, we add ForestDiffusion and Adversarial Random Forests (Watson et al., 2023). Both are evaluated with the full Table 1 protocol on all ten datasets:

MethodC2ST ↑\uparrowShape ↓\downarrowTrend ↓\downarrowβ\beta-recall ↑\uparrowMem@1% ↓\downarrowAvg. rank ↓\downarrow
TabTreeVFM85.261.312.9447.560.144.06
Forest-Flow20.807.785.1642.420.036.62
ARF27.734.6921.7834.220.007.30
ForestDiffusion20.199.736.3536.240.007.71

Avg. rank is computed over all 12 metrics of Table 1 among 14 generators (Table 1 plus the two new baselines). TabTreeVFM remains first overall and has by far the highest fidelity among tree-based generators. The other tree-based generators reproduce fewer training rows, but at a much lower fidelity (C2ST at most 27.73).

Mem versus DCR. Please see our response to Q1.

AuthorsOpenReview ↗

Response to Reviewer yRXo (Part 2 / 5)

Response to W (continued):

Memory and scalability. Peak training memory grows linearly with the number of rows and is directly controlled by KK: across the ten datasets, it is 4.4–57.3 GiB at K=100K=100 and 3.1–10.7 GiB at K=10K=10 (Appendix L). With K=10K=10, TabTreeVFM already matches the leading generators, including on large datasets (see our response to Q5). On Diabetes (101,766 total rows, including 61,059 training rows; 9 numerical and 28 categorical columns with an average cardinality of approximately 79 in the training split, including three diagnosis-code columns with 670–716 categories), it ranks first on Shape (0.80) and WD (0.0016) and second on C2ST (66.97, after TabDiff).

Q1:

I do not understand the added value of the proposed memorization metric over the popular Distance-to-closest-record (DCR) or DCR Share, e.g., as explained in https://arxiv.org/pdf/2504.01908 (https://arxiv.org/pdf/2504.01908). Instead of L2 distance, you are just switching to L1 distance with a difficult-to-choose threshold parameter. You could also report DCR@1%, which would have the same meaning?

Response to Q1:

We appreciate the opportunity to clarify this point.

Mem is not an L2-to-L1 replacement. Our benchmark DCR already uses L1 distance (Appendix H.6), following TabSyn. Mem instead requires exact agreement on categorical attributes and ∣xj−zj∣≤τrj\lvert x_j-z_j\rvert\leq\tau r_j on every numerical attribute jj, where rjr_j is the attribute range. Its numerical criterion is therefore a range-normalized L∞L_\infty threshold.

DCR share and Mem answer different questions. DCR share measures whether synthetic records are closer to training or holdout records; Mem measures the positive excess of synthetic-to-training full-row matches over held-out-to-training matches. On Churn, TabSyn and TabTreeVFM have DCR shares of 53.29% and 51.26%, but Mem scores of 2.26 and 0.44 percentage points. On NMES, CoDi has a higher DCR share than CDTD (94.93% vs. 86.80%) but much lower Mem (1.37 vs. 18.75 percentage points). Thus, relative training-set proximity does not determine the amount of tolerance-bounded full-row reproduction.

Relationship to the cited framework. Sidorenko et al. (2025, Section 2.3) compute L2 distances in a language-model embedding space and, alongside DCR share, report identical-match statistics against a holdout set. We build on this holdout-based assessment principle, while specifying near-copy matching directly in the original mixed-type attributes and reporting the excess matching rate. This extension matters empirically: in our additional three-seed evaluation, CDTD's average excess exact-match rate is 0.85 percentage points, compared with 8.43 for Mem@1%; on Credit-g, TabbyFlow has an excess exact-match rate of 0.00, but Mem@1% of 20.20. Exact matching alone therefore misses substantial reproduction under the tolerance-based criterion.

Regarding DCR@1%. DCR@1% would not have the same meaning. A first-percentile DCR is a distance, not a reproduction rate. The share of synthetic records whose DCR falls below a threshold is a rate, but it differs from Mem in two ways. First, a threshold on the summed L1 distance becomes stricter per attribute as the number of columns grows, whereas Mem bounds the deviation on each attribute. Second, it has no held-out reference: at τ=0.01\tau=0.01, real held-out records already match training records at rates of 3.52% on Default and 2.50% on Adult. Mem explicitly measures the excess above this observed reference.

Regarding threshold selection. The tolerance τ\tau specifies the maximum numerical deviation as a fraction of each attribute's range, while categorical values must match exactly. Across the tolerances evaluated in Table 19 (0.1%–5%), CDTD and TabbyFlow have the highest dataset-averaged Mem, while TabTreeVFM remains at 0.14–0.28 percentage points, below TabDiff and TabSyn throughout.

AuthorsOpenReview ↗

Response to Reviewer yRXo (Part 3 / 5)

Q2:

The training times for TabDiff and Tabbyflow seems very excessive for some datasets. I do understand that CDTD is particularly efficient, but it is still a diffusion model. Why do the other models train for more than 100 minutes on the adult datasets whereas for CDTD 3 minutes are enough? Are you training CDTD for long enough to converge to a good state? The high DCR values seem like overfitting to me?

Many thanks for examining the training budgets so carefully. We strictly followed CDTD's official configuration and completed 30,000 optimization steps on all ten datasets. CDTD's original paper also reports that 30,000 steps are sufficient for convergence in its evaluated settings ( Appendix M of CDTD paper).

The runtime difference reflects different training workloads and implementations. On Adult, TabDiff and TabbyFlow perform approximately 64,000 optimization updates, and their official training loops additionally evaluate EMA loss over the full training set after every epoch. They also use multi-worker PyTorch data loading, whereas CDTD employs a lightweight tensor-based loader.

Q3:

Should the component ablation in Table 2 not start from EF-VFM (TabbyFlow) model? Since you keep the specification the same and switch out the backbone?

We are grateful for this good question. Following your suggestion, Table 2 in the revised manuscript now starts from TabbyFlow, and a new paragraph in Section 4.4 (Estimator comparison) discusses this comparison (both highlighted in blue). We add TabbyFlow (Aligned), which keeps the network and training recipe of TabbyFlow but follows the specification of TabTreeVFM, including a probability path with the same residual terminal noise in training and sampling, a unit-variance Gaussian objective with a fixed weight for numerical columns rather than a time-dependent variance with a weight annealed to zero, and the sampling grid. The remaining differences thus lie in the estimator: one network amortized over continuous time and trained with Adam, or independent boosted-tree heads at each grid time fit by second-order boosting, whose categorical leaf updates are diagonal-Fisher-scaled (Section 3.5):

ModelEstimatorC2ST ↑\uparrowShape ↓\downarrowTrend ↓\downarrow∣DCR−50∣\lvert \mathrm{DCR}-50 \rvert ↓\downarrowMIA ↑\uparrowMem@1% ↓\downarrow
TabbyFlow (Aligned)Network78.101.793.0121.020.96365.38
Direct velocity regressionTrees78.622.514.891.470.98430.07
TabTreeVFMTrees85.261.312.942.480.98370.14

With the specification fixed, switching to boosted trees raises C2ST from 78.10 to 85.26 and lowers Shape error on every dataset. TabbyFlow (Aligned), in contrast, has the largest DCR deviation and Mem@1% in Table 2, 21.02 and 5.38 percentage points against 2.48 and 0.14 for TabTreeVFM, and its DCR share is farther from 50% on every dataset. Lower reproduction thus accompanies higher fidelity rather than coming at its expense. Trees alone are not sufficient either: with direct velocity regression, C2ST stays at 78.62, and it reaches 85.26 only with the boosted-tree multinomial heads of TabTreeVFM (Table 2).

Q4:

Appendix L lacks a deeper comparison to ForestDiffusion: What explains the time differences between ForestFlow and TabTreeVFM?

We thank the reviewer for this question. The runtime differences in Table 21 largely reflect differences in training implementations. We use the scalable Forest-Flow implementation released by Cresswell and Kim (2024). The two implementations differ in their data handling and model organization. For example, Forest-Flow fits separate XGBRegressor models across timesteps, classes, and output columns, rebuilding the data matrix for each fit and repeatedly incurring preprocessing overhead, including quantile sketch construction and GPU data transfers. In contrast, TabTreeVFM constructs a QuantileDMatrix once per timestep and reuses it across prediction heads, with GPU data prepared directly in device memory. Moreover, TabTreeVFM uses a single multinomial classifier per categorical column rather than separate regressors for individual one-hot dimensions, reducing the number of independent model fits. These implementation and modeling differences help explain the observed runtime difference.

Reviewer_wCTTOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents TabTreeVFM, an XGBoost-based generator for mixed-type tabular data. It predicts clean data endpoints with Gaussian regression for continuous variables and multinomial classification for categorical variables, then derives the flow velocity from these predictions. Across 10 datasets and 11 baselines, TabTreeVFM achieves the best average rank of 3.56. It also delivers strong data fidelity and downstream utility, with low excess reproduction of training rows.

Critical strengths
  1. The problem is of practical significance. Mixed-type tabular data is ubiquitous in finance, healthcare, and business; the joint generation of continuous and categorical variables is a critical challenge.
  2. The methodology is clearly designed, with thorough ablation studies. The paper evaluates components such as velocity prediction, endpoint prediction, simplex projection, and the multinomial classification head, effectively demonstrating the individual contributions of endpoint modeling and the multinomial head to overall performance.
  3. The experimental evaluation is comprehensive. The study covers 10 datasets, 12 generative models, and a wide range of metrics for fidelity, utility, and privacy. It further analyzes the replication of training rows, stability across random seeds, and sensitivity to hyperparameters.
Critical weaknesses
  1. Limited technical contribution and privacy interpretation regarding "Mem"

The definition of "Mem" is relatively simplistic, primarily measuring whether the rate of near-duplicate rows (relative to a held-out dataset) exceeds what would be expected by chance. Consequently, it only reflects "full-row replication" and cannot support broader claims such as "the model lacks memory" or "the model offers general privacy advantages." Specifically, "Mem" fails to account for risks such as reconstructing data by combining multiple synthetic rows, or scenarios where a small number of high-risk individuals are replicated, even if the overall proportion remains low. Furthermore, this metric is highly sensitive to the number of columns, the cardinality of categorical variables, numerical outliers, the range of values, and the overall scale of the dataset.

  1. Lack of statistical significance in aggregate rankings

The paper ranks models based on various metrics within each dataset and then calculates an unweighted average of these ranks. While this aggregation method is convenient for presentation, its statistical interpretation remains limited. First, metrics such as Shape, WD, and JSD may be highly correlated and cannot be treated as independent pieces of evidence. Second, the relative importance of fidelity, utility, and privacy varies by application; assigning equal weights to them implicitly introduces strong subjective assumptions. Finally, average rankings may mask significant differences across datasets and fail to capture the Pareto trade-offs between different metrics.

Questions for authors
  1. Could the authors report confidence intervals for the "repetition rate" on the hold-out set? Given the current low values ​​for the "Mem" metric, the observed differences might be smaller than the fluctuations caused by data partitioning or finite sample sizes.

  2. Does a low "Mem" score necessarily stem from the proposed method's design? Low "Mem" values ​​could also result from limited tree capacity, regularization, underfitting, or over-smoothing of the generative distribution. It is recommended to analyze the impact of tree depth, boosting rounds, and regularization strength on the "Mem" metric.

  3. Could the authors include experiments using real-world sensitive data or scenarios with more complex business constraints? While current public benchmarks facilitate reproducibility, they are insufficient for evaluating privacy risks and the ability to satisfy constraints in high-stakes domains such as healthcare or finance.

Following the discussion during the rebuttal period, most of my concerns have been addressed, and I have raised my score to 3.

4 comments
AuthorsOpenReview ↗

Response to Reviewer wCTT (Part 1 / 4)

W1:

Limited technical contribution and privacy interpretation regarding "Mem"

Response to W1:

We are grateful for the careful discussion of what Mem can and cannot measure.

Regarding the contribution and interpretation of Mem. Mem is designed to measure excess full-row reproduction rather than general privacy risk. Fang et al. (2025, Table 1) report that TabSyn achieves a DCR share of 52.50% on Shoppers while exhibiting a memorization ratio of 27.68%, demonstrating that a near-balanced DCR does not necessarily imply low memorization. Compared with Fang et al.'s relative nearest-neighbor distance ratio, Mem offers two advantages. First, it uses an explicit mixed-type matching criterion, requiring exact categorical matches and bounded numerical differences. Second, it subtracts the matching rate observed in real held-out data, distinguishing excess training-row reproduction from naturally occurring matches.

We agree that Mem does not cover reconstruction from multiple synthetic rows, which is hard to define: if values can come from arbitrary rows, almost any record can be assembled. Mem can, however, be extended to any subset of columns to test whether partial records are reproduced more often than in held-out data. Moreover, Mem's criterion is evaluated row by row, so beyond the aggregate rate it identifies exactly which training records are reproduced, even when the rate is low. Practitioners can therefore inspect these records, for example those with rare attribute combinations, and use domain knowledge to judge whether high-risk individuals are among them. Because the tolerance applies to each attribute, the criterion does not tighten as columns are added. The number of columns, categorical cardinality, and dataset size change how often real records naturally match the training set, which is exactly the rate Mem subtracts; our response to Q1 quantifies its finite-sample variability. The range-normalized tolerance is invariant to affine rescaling of numerical features, and robust percentile-based ranges can replace min–max ranges for outlier-heavy attributes without changing the framework.

Our contribution is therefore an interpretable, held-out-calibrated measure that directly quantifies excess full-row reproduction in mixed-type tabular data.

W2:

Lack of statistical significance in aggregate rankings

Response to W2:

The point about interpreting average ranks is well taken, and we thank the reviewer for it. We address its four aspects below.

Statistical reliability. Treating datasets as the unit of replication, we resample the ten datasets with replacement (5,000 bootstrap draws) and recompute the overall ranking: TabTreeVFM ranks first in 96.9% of the resamples and never below second. The ranking is also stable under training randomness: with three training seeds for TabTreeVFM and for each of the four closest baselines, TabTreeVFM remains first, also when each of its seeds is paired with the corresponding baseline seed.

Correlated metrics and weighting. We recompute the ranking under different aggregations:

AggregationTabTreeVFM Avg. RankPosition
Fidelity only2.9291st
Fidelity + Utility3.0001st
Fidelity + Utility + Privacy (equal group weights)3.8041st
One metric per distributional level (Shape, Trend, C2ST)2.0671st

Equal group weights count the correlated fidelity metrics as a single group, and the last row keeps one metric per distributional level, removing the overlap among Shape, WD, and JSD. TabTreeVFM ranks first under every aggregation, and the overall ranking keeps it first when any single metric or dataset is removed, including Mem, the metric we propose (average rank 3.39).

Differences across datasets. Averaging the 12 metric ranks within each dataset, TabTreeVFM ranks first on six datasets and second on the remaining four. Per-dataset results for every metric are reported in Appendix I.

Pareto trade-offs. On average, no generator attains both higher C2ST and lower Mem than TabTreeVFM (Figure 4 in the revised manuscript). Higher coverage comes only with higher reproduction: CDTD and TabbyFlow reach β\beta-recall of 62.70 and 58.29 but Mem@1% of 8.37 and 3.51 percentage points, against 47.56 and 0.14 for TabTreeVFM.

AuthorsOpenReview ↗

Response to Reviewer wCTT (Part 2 / 4)

Q1:

Could the authors report confidence intervals for the "repetition rate" on the hold-out set? Given the current low values for the "Mem" metric, the observed differences might be smaller than the fluctuations caused by data partitioning or finite sample sizes.

Response to Q1:

Thank you for suggesting confidence intervals. We have added bootstrap analyses covering all 10 datasets and seven tolerances, with 500 resamples per setting. With the fitting set fixed, the held-out repetition rates have the following central 95% bootstrap ranges at the primary tolerance τ=0.01\tau=0.01:

Dataset95% bootstrap range (%)
Adult[2.301, 2.693]
Beijing[0.000, 0.000]
Churn[0.701, 1.766]
Credit-g[0.000, 0.000]
Default[3.192, 3.830]
Lending[0.000, 0.000]
Magic[0.000, 0.053]
News[0.000, 0.000]
NMES[0.000, 0.342]
Shoppers[1.008, 1.606]

To directly assess uncertainty in the method differences, we additionally resample both the holdout and whole synthetic tables, recomputing the final Mem scores using the original evaluation procedure. All methods share the same holdout resample.

We define D=Membaseline−MemTabTreeVFMD=\mathrm{Mem}_{\mathrm{baseline}}-\mathrm{Mem}_{\mathrm{TabTreeVFM}}, where positive values indicate lower Mem for TabTreeVFM. The differences averaged across the ten datasets are:

BaselineMean difference DD (percentage points)Conservative envelope of the 95% bootstrap range (percentage points)
TabDiff+0.1556[0.1194, 0.1845]
CDTD+8.2230[7.8751, 8.2577]
TabbyFlow+3.3654[3.1758, 3.4284]
TabSyn+0.1685[0.1407, 0.1947]

The difference ranges for TabDiff, CDTD, TabbyFlow, and TabSyn are all above zero, confirming that TabTreeVFM consistently achieves lower Mem scores in these comparisons.

AuthorsOpenReview ↗

Response to Reviewer wCTT (Part 3 / 4)

Q2:

Does a low "Mem" score necessarily stem from the proposed method's design? Low "Mem" values could also result from limited tree capacity, regularization, underfitting, or over-smoothing of the generative distribution. It is recommended to analyze the impact of tree depth, boosting rounds, and regularization strength on the "Mem" metric.

Response to Q2:

We appreciate this insightful question. We conducted additional experiments on five datasets to investigate whether TabTreeVFM's low Mem scores arise from limited tree capacity, strong regularization, underfitting, or over-smoothing. We varied L2 regularization λ\lambda over 0.1,1,10,20\\{0.1, 1, 10, 20\\}, boosting rounds over 50,100,200\\{50, 100, 200\\}, and tree depth over 3,7,9\\{3, 7, 9\\}, changing one hyperparameter at a time. Each configuration was trained on the 50/50 split, and we report Mem@1% together with coverage (β\beta-recall) for the same trained models, both averaged over all 20 synthetic tables.

Datasetβ\beta-recall across λ\lambdaMem across λ\lambdaβ\beta-recall, rounds 50→100→20050 \rightarrow 100 \rightarrow 200Mem, rounds 50→100→20050 \rightarrow 100 \rightarrow 200β\beta-recall, depth 3→7→93 \rightarrow 7 \rightarrow 9Mem, depth 3→7→93 \rightarrow 7 \rightarrow 9
Credit-g52.5–53.30.0049.2→53.1→56.349.2 \rightarrow 53.1 \rightarrow 56.30.00→0.00→0.000.00 \rightarrow 0.00 \rightarrow 0.0053.1→62.4→64.453.1 \rightarrow 62.4 \rightarrow 64.40.00→0.03→0.060.00 \rightarrow 0.03 \rightarrow 0.06
NMES47.9–48.00.0046.8→48.0→48.846.8 \rightarrow 48.0 \rightarrow 48.80.00→0.00→0.000.00 \rightarrow 0.00 \rightarrow 0.0048.0→53.2→55.248.0 \rightarrow 53.2 \rightarrow 55.20.00→0.06→0.280.00 \rightarrow 0.06 \rightarrow 0.28
Lending50.1–50.70.0049.0→50.7→52.049.0 \rightarrow 50.7 \rightarrow 52.00.00→0.00→0.000.00 \rightarrow 0.00 \rightarrow 0.0050.7→55.4→56.850.7 \rightarrow 55.4 \rightarrow 56.80.00→0.00→0.000.00 \rightarrow 0.00 \rightarrow 0.00
Adult48.8–49.00.0048.0→49.0→49.648.0 \rightarrow 49.0 \rightarrow 49.60.00→0.00→0.000.00 \rightarrow 0.00 \rightarrow 0.0045.4→49.0→50.045.4 \rightarrow 49.0 \rightarrow 50.00.00→0.00→0.000.00 \rightarrow 0.00 \rightarrow 0.00
Churn14.5–14.70.54–0.6111.5→14.7→18.511.5 \rightarrow 14.7 \rightarrow 18.50.08→0.62→1.010.08 \rightarrow 0.62 \rightarrow 1.0114.7→36.7→44.714.7 \rightarrow 36.7 \rightarrow 44.70.62→2.95→5.360.62 \rightarrow 2.95 \rightarrow 5.36

Increasing model capacity improves coverage without necessarily increasing memorization. If low Mem were a by-product of underfitting or over-smoothing, higher capacity should raise coverage and Mem together. Instead, more boosting rounds and deeper trees consistently increase β\beta-recall, e.g., from 53.1 to 64.4 on Credit-g and from 50.7 to 56.8 on Lending as depth grows, while Mem remains at or below 0.28 percentage points on four of the five datasets. The same settings also substantially improve C2ST, e.g., from 5.1 to 53.7 on Lending with depth and from 19.0 to 44.8 on NMES with boosting rounds. Varying λ\lambda changes β\beta-recall by less than 1 percentage point and has little effect on Mem. These results indicate that low Mem is not simply a consequence of shallow trees, insufficient boosting rounds, or strong regularization.

Churn reveals a dataset-dependent coverage–reproduction trade-off. It is a small dataset with predominantly integer-valued features and the highest natural repetition in our benchmark. Records occurring more than once constitute 11% of its training set but account for 34–44% of detected replica matches, indicating that reproduction is concentrated on recurring records. As tree depth increases from 3 to 9, β\beta-recall improves from 14.7 to 44.7, while Mem increases from 0.62 to 5.36 percentage points. This pattern is consistent with deeper trees capturing recurring training patterns more closely, thereby increasing the frequency of samples within the matching tolerance.

Overall, these experiments show that improved fidelity and coverage do not inherently require higher full-row reproduction, while Mem remains sensitive to capacity-associated reproduction when it occurs. The relationship depends on the dataset rather than being determined solely by tree capacity or regularization strength.

AuthorsOpenReview ↗

Response to Reviewer wCTT (Part 4 / 4)

Q3:

Could the authors include experiments using real-world sensitive data or scenarios with more complex business constraints? While current public benchmarks facilitate reproducibility, they are insufficient for evaluating privacy risks and the ability to satisfy constraints in high-stakes domains such as healthcare or finance.

Response to Q3:

Thanks for encouraging evaluation in more realistic, high-stakes settings. To address this concern, we additionally include Diabetes 130-US Hospitals, a real-world healthcare dataset containing 101,766 hospital encounters from 130 U.S. hospitals, covering clinical diagnoses, discharge dispositions, and readmission outcomes.

We evaluate eight cross-column consistency rules across three domains: finance (Lending), e-commerce (Shoppers), and healthcare (Diabetes). The rules are derived from dataset field definitions and verified against real training and held-out records:

  • Lending: Consistency between loan grades and sub-grades, sub-grade-specific interest rates, principal repayment bounds, and open versus total credit lines.
  • Shoppers: Zero Administrative or Informational page visits must correspond to zero accumulated visit duration.
  • Diabetes: The recorded diagnosis count must cover the observed diagnosis codes, and patients discharged with an expired status must not be marked as readmitted.

We evaluate all 20 unmodified synthetic tables per method, without constraint-specific training or post-processing. The violation rates (%) are reported below.

DatasetConstraintTabTreeVFMTabDiffCDTDTabbyFlowTabSynTabVVFMForest-Flow
LendingGrade–sub-grade consistency0.7561.3616.4112.2312.4261.120.93
LendingSub-grade–interest-rate consistency12.3691.3828.6629.6744.2178.2376.16
LendingPaid principal ≤\leq loan amount1.492.012.051.882.9521.032.03
LendingOpen credit lines ≤\leq total credit lines1.2817.911.380.461.4039.610.33
ShoppersZero Administrative visits ⇒\Rightarrow zero duration0.161.031.140.910.931.7456.32
ShoppersZero Informational visits ⇒\Rightarrow zero duration0.020.200.390.370.420.0766.23
DiabetesDiagnosis-count consistency4.7012.7381.6246.9965.3091.04—
DiabetesExpired discharge ⇒\Rightarrow no readmission1.1010.630.0411.5516.690.21—

TabTreeVFM achieves the lowest violation rate on six of the eight rules and remains close to the best on the other two. Every baseline, in contrast, violates at least one rule in a large fraction of its synthetic rows, and the baselines that lead on a single rule perform poorly on others. TabTreeVFM thus preserves domain-specific relationships consistently across finance, e-commerce, and healthcare, without explicit constraint enforcement.

Reviewer_dS9iOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes TabTreeVFM, which applies exponential-family variational flow matching to mixed-type tabular generation using boosted trees. At each timestep, separate heads predict numerical endpoint means and categorical endpoint probabilities, which determine the flow velocity analytically. All heads condition on the complete noisy row, allowing the model to capture dependencies between columns despite the factorized endpoint posterior.

The paper also relates XGBoost’s multiclass leaf updates to diagonal Fisher curvature, with a correction determined by within-leaf prediction dispersion. Finally, it introduces Mem, which measures full-row reproduction above a held-out reference, accounting for collisions that occur naturally in the data.

Experiments cover 10 datasets and 11 baselines. TabTreeVFM achieves the best reported overall average rank, with additional component ablations, alternative fidelity discriminators, training-seed checks, and time/memory measurements.

Critical strengths

Stengths

  1. The modeling change is well motivated, and the ablations show that it helps. The method makes sensible use of EF-VFM by fitting appropriate endpoint predictors for numerical and categorical variables while conditioning every head on the complete state. I liked the endpoint-regression baseline with simplex projection in Table 2. Relative to this baseline, multinomial heads reduce Trend error from 4.18 to 2.94 and the classification utility gap from 2.85% to 1.77%. This supports the value of the categorical training objective beyond simply enforcing valid probabilities during sampling.
  2. The empirical evaluation makes a convincing case for practical usefulness. Table 1 supports the best average ranks on Shape, Trend, WD, C2ST, and classification utility. The alternative C2ST discriminator and the ten-training-seed study for TabTreeVFM are nice robustness checks. The controlled direct-velocity comparison is useful: C2ST improves from 78.62 to 85.26 under matched preprocessing and sampling settings (§4.4). Together, these results give reasonably strong evidence that the proposed adaptation improves tree-based generation.

In short, the paper is clear and effective, though perhaps modest in its contribution.

Critical weaknesses

Weaknesses

  1. My main reservation is that the methodological contribution is relatively modest. The endpoint formulation and type-specific objectives follow EF-VFM, while boosted-tree predictors have already been used for tabular flow models. Proposition 3.1 adds a useful interpretation of the existing multiclass boosting update. I therefore see the main contribution as bringing these ideas together effectively and demonstrating their practical benefits. The ablations make a convincing case for those benefits, but the amount of new methodology is somewhat limited.
  2. Mem’s contribution needs clearer positioning relative to existing held-out matching measures. Ganev & De Cristofaro (§2.3) and Yao et al. (§3.1), both already cited in the paper, discuss Identical Match Share compared against a held-out reference. At τ=0\tau=0, Mem appears to take the positive part of the corresponding synthetic-minus-held-out match-rate difference. The tolerance-based extension and reporting the excess as a continuous score may be useful, but the relationship to this existing comparison should be explained explicitly. Please clarify exactly what is new and revise the distinction drawn in §2. This matters because Mem is presented as a separate main contribution.
  3. Maybe I am misunderstanding the evaluation, but it seems the combination of strong quality and low reproduction is not directly evaluated on the same fitted models. Fidelity, utility, and coverage use the standard benchmark splits, 90/10 except for Adult, whereas DCR and Mem use separately retrained models on 50/50 splits (Appendices G, H.6, and H.8). This substantially changes the amount of training data; on Credit-g, for example, it changes from 900 to 500 rows. Since §4.3 interprets reproduction jointly with coverage, and §4.4 concludes that quality gains are not obtained through memorization, the different protocols leave a gap in the evidence. Reporting quality and coverage for the existing samples from the 50/50-trained models would directly address this concern.
Questions for authors

Questions For Authors

  1. Could you report C2ST, coverage, and Mem together for the same 50/50-trained generators, at least for TabTreeVFM and the strongest relevant baselines? Whether the quality advantage persists alongside low reproduction is the main empirical question affecting my assessment.
  2. A minor point, but how frequent are distance ties in Eq. (H.4)? Ties contribute zero, but results are ranked by their distance from 50%. Under an exchangeable-reference null, the expected strict-comparison score is 1−Pr⁡[tie]2\frac{1-\Pr[\mathrm{tie}]}{2}. Do the DCR and overall rankings change if ties receive half credit, or if the score is calibrated against a suitable real-data reference? Reporting negligible tie frequencies would also resolve this concern.
  3. How many independent training runs are used for the results for the closest baselines?
3 comments
AuthorsOpenReview ↗

Response to Reviewer dS9i (Part 1 / 3)

W1:

My main reservation is that the methodological contribution is relatively modest. The endpoint formulation and type-specific objectives follow EF-VFM, while boosted-tree predictors have already been used for tabular flow models. Proposition 3.1 adds a useful interpretation of the existing multiclass boosting update. I therefore see the main contribution as bringing these ideas together effectively and demonstrating their practical benefits. The ablations make a convincing case for those benefits, but the amount of new methodology is somewhat limited.

Response to W1:

We are glad that the reviewer finds the ablations convincing, and we would like to clarify two methodological contributions of the paper.

First, our tree-specific design makes the endpoint formulation substantially more effective than its neural implementation in EF-VFM. Each timestep decomposes into independent XGBoost regression and multiclass heads whose posterior means define the velocity analytically. Unlike EF-VFM, which trains one time-conditioned network across all timesteps and columns with Adam and leaves natural-gradient optimization of its likelihoods to future work, each head is fit by second-order boosting, whose leaf updates use the curvature of that head's endpoint likelihood. Each existing component alone falls well short: under the same probability path, solver, and preprocessing, a neural network trained on the same endpoint objective reaches a C2ST of 78.10, and boosted trees regressing velocities, as in Forest-Flow, reach 78.62, whereas TabTreeVFM reaches 85.26 (Table 2 and Section 4.4 of the revised manuscript).

Second, Proposition 3.1 goes beyond interpreting the multiclass boosting update: it gives an exact identity, Δjk=Δjkrep/(1−ρjk)\Delta_{jk}=\Delta^{\mathrm{rep}}_{jk}/(1-\rho_{jk}), between the implemented leaf update and the diagonal-Fisher step at the leaf-mean prediction, where the within-leaf dispersion ratio ρjk\rho_{jk} determines when the two coincide and how far they differ. Appendix D tests the practical role of this curvature on all 46 non-binary categorical heads from nine datasets: replacing it with a constant increases held-out endpoint cross-entropy by about two orders of magnitude at late times (t≥0.90t\geq0.90) and degrades categorical fidelity in generation, while the full Fisher matrix used in natural-gradient boosting gives no consistent improvement. Standard second-order boosting therefore already provides the required Fisher scaling, without the explicit Fisher inversion of NGBoost.

W2:

Mem’s contribution needs clearer positioning relative to existing held-out matching measures. Ganev & De Cristofaro (§2.3) and Yao et al. (§3.1), both already cited in the paper, discuss Identical Match Share compared against a held-out reference. At τ=0\tau = 0, Mem appears to take the positive part of the corresponding synthetic-minus-held-out match-rate difference. The tolerance-based extension and reporting the excess as a continuous score may be useful, but the relationship to this existing comparison should be explained explicitly. Please clarify exactly what is new and revise the distinction drawn in §2. This matters because Mem is presented as a separate main contribution.

Response to W2:

Thanks for pointing out the relationship to existing held-out IMS comparisons. At τ=0\tau=0, Mem is indeed the positive part of the synthetic-minus-held-out IMS difference.

We have clarified this relationship in Section 2 (highlighted in blue). Our contribution lies in two extensions. First, we extend exact matching to tolerance-aware full-row matching for mixed-type data, combining exact categorical matches with range-normalized numerical tolerances. Second, we quantify excess reproduction as a continuous score rather than a binary pass/fail test. Applying Mem across ten datasets and twelve generators reveals substantial near-duplicate reproduction in high-fidelity generators (CDTD and TabbyFlow).

AuthorsOpenReview ↗

Response to Reviewer dS9i (Part 2 / 3)

W3 & Q1:

W3: Maybe I am misunderstanding the evaluation, but it seems the combination of strong quality and low reproduction is not directly evaluated on the same fitted models. Fidelity, utility, and coverage use the standard benchmark splits, 90/10 except for Adult, whereas DCR and Mem use separately retrained models on 50/50 splits (Appendices G, H.6, and H.8). This substantially changes the amount of training data; on Credit-g, for example, it changes from 900 to 500 rows. Since §4.3 interprets reproduction jointly with coverage, and §4.4 concludes that quality gains are not obtained through memorization, the different protocols leave a gap in the evidence. Reporting quality and coverage for the existing samples from the 50/50-trained models would directly address this concern.

Q1: Could you report C2ST, coverage, and Mem together for the same 50/50-trained generators, at least for TabTreeVFM and the strongest relevant baselines? Whether the quality advantage persists alongside low reproduction is the main empirical question affecting my assessment.

Response to W3 & Q1:

We appreciate this concrete suggestion. We now report C2ST, coverage (β\beta-recall), and Mem on the same 50/50-trained generators for all ten datasets. C2ST and β\beta-recall are averaged over 20 synthetic releases per generator.

MethodC2ST ↑\uparrowβ\beta-recall ↑\uparrowMem@1% ↓\downarrowAvg. rank ↓\downarrow
TabTreeVFM55.6149.400.152.57
TabbyFlow41.6758.713.513.30
CDTD49.3565.268.373.37
TabDiff40.4043.940.303.75
Forest-Flow14.0442.950.034.00
TabSyn36.8646.390.254.02

The quality advantage persists alongside low reproduction. TabTreeVFM has the highest average C2ST and the best C2ST on 7 of the 10 datasets; against each baseline, its C2ST is higher on at least 8 of the 10 datasets. Its Mem stays below 1 percentage point on every dataset. CDTD and TabbyFlow reach higher coverage, but with 8.37 and 3.51 percentage points of excess reproduction versus 0.15 for TabTreeVFM. Among generators whose average Mem is below 1 percentage point (TabTreeVFM, TabDiff, TabSyn, and Forest-Flow), TabTreeVFM attains both the highest C2ST and the highest β\beta-recall.

Q2:

A minor point, but how frequent are distance ties in Eq. (H.4)? Ties contribute zero, but results are ranked by their distance from 50%. Under an exchangeable-reference null, the expected strict-comparison score is 1−Pr⁡[tie]2\frac{1-\Pr[\mathrm{tie}]}{2}. Do the DCR and overall rankings change if ties receive half credit, or if the score is calibrated against a suitable real-data reference? Reporting negligible tie frequencies would also resolve this concern.

Response to Q2:

Thank you for raising the question of distance ties. Our original DCR evaluation follows the established TabSyn/TabDiff/TabbyFlow benchmark protocol, using equal-sized training and holdout sets. TabSyn's official implementation counts a synthetic record only when its distance to the training set is strictly smaller than its distance to the holdout set, so ties receive zero credit, consistent with our Eq. (H.4).

Tie frequencies depend on the dataset. On Adult, Beijing, Credit-g, Default, Lending, News, and NMES, ties are negligible: at most 0.17% of synthetic rows for any method. They are more frequent on Churn (1.2–9.4% across methods) and, to a lesser extent, on Shoppers (0.1–1.4%). These two datasets are also the only ones on which the half-credit recomputation below changes any DCR rank.

Following the reviewer's suggestion, we additionally recomputed DCR with half credit for ties. The overall ranking of all 12 methods remains unchanged, and TabTreeVFM remains first (average rank: 3.5583→3.56673.5583 \rightarrow 3.5667). Its DCR ranking changes from second to third (3.90→4.003.90 \rightarrow 4.00).

AuthorsOpenReview ↗

Response to Reviewer dS9i (Part 3 / 3)

Q3:

How many independent training runs are used for the results for the closest baselines?

Response to Q3:

This is an important question, and we thank the reviewer for raising it. Table 1 follows the protocol of Appendix G: every generator, including TabTreeVFM, is fitted once. For TabTreeVFM, Appendix J additionally repeats training with 10 seeds.

To quantify training randomness for the closest baselines, we retrained the four closest baselines (TabDiff, TabSyn, CDTD, and TabbyFlow) with three independent training seeds (0, 1, and 2) on all ten datasets, and compare them with TabTreeVFM trained with the same three seeds (from Appendix J). Each run generates 20 synthetic tables, which we evaluate with all Table 1 metrics.

MetricTabTreeVFMCDTDTabDiffTabbyFlowTabSyn
C2ST (%) ↑\uparrow85.51 ±\pm 0.1875.45 ±\pm 0.2271.92 ±\pm 0.4979.25 ±\pm 0.2070.77 ±\pm 2.15
β\beta-recall (%) ↑\uparrow47.41 ±\pm 0.0363.36 ±\pm 0.0645.22 ±\pm 1.1957.83 ±\pm 0.0745.85 ±\pm 0.82
α\alpha-precision (%) ↑\uparrow96.85 ±\pm 0.0297.30 ±\pm 0.0496.46 ±\pm 0.4092.34 ±\pm 0.1895.82 ±\pm 0.79
Shape error (%) ↓\downarrow1.30 ±\pm 0.001.95 ±\pm 0.031.88 ±\pm 0.052.37 ±\pm 0.012.57 ±\pm 0.14
Trend error (%) ↓\downarrow2.89 ±\pm 0.033.18 ±\pm 0.064.90 ±\pm 0.233.38 ±\pm 0.034.34 ±\pm 0.41
WD ↓\downarrow0.0039 ±\pm 0.00000.0062 ±\pm 0.00010.0069 ±\pm 0.00030.0078 ±\pm 0.00010.0076 ±\pm 0.0007
JSD ↓\downarrow0.0142 ±\pm 0.00000.0126 ±\pm 0.00040.0153 ±\pm 0.00010.0256 ±\pm 0.00030.0354 ±\pm 0.0011
MLE-C (avg. rank) ↓\downarrow3.174.003.333.504.33
MLE-R (avg. rank) ↓\downarrow3.753.254.003.254.25
∣DCR−50∣\lvert \mathrm{DCR}-50 \rvert (pp) ↓\downarrow2.69 ±\pm 0.0826.53 ±\pm 0.075.20 ±\pm 0.6415.84 ±\pm 0.364.83 ±\pm 0.51
MIA score ↑\uparrow0.9835 ±\pm 0.00020.9634 ±\pm 0.00070.9776 ±\pm 0.00320.9721 ±\pm 0.00280.9774 ±\pm 0.0025
Mem@1% (pp) ↓\downarrow0.14 ±\pm 0.008.43 ±\pm 0.150.27 ±\pm 0.012.89 ±\pm 0.210.31 ±\pm 0.06
Avg. rank (12 methods) ↓\downarrow3.535.244.475.664.87

The conclusions of Table 1 are robust to training randomness. With repeated training runs for all five methods, TabTreeVFM remains first in average rank, and TabDiff, TabSyn, CDTD, and TabbyFlow remain second to fifth, as in Table 1. Seed-to-seed variation is small relative to the differences between methods. TabTreeVFM also has the best seed-averaged C2ST, Shape, Trend, WD, DCR, MIA, and Mem scores, as well as the best MLE-C rank.

Comments on the paper

3 comments
AuthorsOpenReview ↗

Follow-Up on Our Rebuttal and Addressed Concerns

Dear Reviewers,

With only two days left in the discussion period, we would like to kindly follow up on our rebuttal. We believe that our responses have adequately addressed your concerns and clarified the questions raised in your review. We would greatly appreciate it if you could take a moment to review our responses and let us know whether your concerns have been resolved.

If any questions remain or further clarification is needed, we would be more than happy to address them.

Thank you for your time and thoughtful feedback. Your comments have been invaluable in helping us improve our work.

AuthorsOpenReview ↗

Summary of Rebuttal (Part 1 / 2)

Dear AC and Reviewers,

We thank the AC and reviewers (dS9i, wCTT, and yRXo) for their constructive feedback. We are encouraged by their recognition of our well-motivated method, convincing ablations, extensive experiments, and insightful likelihood-geometry analysis. We also appreciate Reviewer yRXo's constructive follow-up and positive reassessment.

Below, we highlight our core contributions and summarize how the additional experiments, manuscript revisions, and detailed responses address the reviewers' main concerns.

Core Contributions

  1. Type-aware endpoint learning with boosted trees. Building on EF-VFM, TabTreeVFM replaces its time-conditioned neural network with independently trained XGBoost heads at each timestep: Gaussian regression for numerical columns and multinomial classification for categorical columns. All heads condition on the full noisy row, and their posterior means analytically recover the flow velocity, reducing training to standard tree-boosting problems.
  2. An exact link between boosting updates and Fisher geometry. We prove that the multiclass leaf update of XGBoost is a diagonal-Fisher-scaled step on the categorical endpoint likelihood, with an exact correction given by within-leaf dispersion (Proposition 3.1). This result connects to a point raised in EF-VFM, which notes that the Fisher information of its endpoint likelihoods is the Hessian of the log-partition function but leaves natural-gradient optimization to future work. With boosted trees, Fisher scaling comes from standard second-order boosting, without a custom optimizer. Appendix D shows that this curvature matters in practice, whereas the full Fisher matrix of natural-gradient boosting brings no consistent gain.
  3. An interpretable audit of training-row reproduction. Mem quantifies excess near-copy reproduction relative to real held-out data, in percentage points. Unlike DCR share, which measures relative proximity to training and held-out records, Mem can reveal substantial reproduction even when DCR share appears balanced. Across 12 generators, it exposes excess reproduction in several high-fidelity models, motivating the evaluation of reproduction alongside fidelity.
  4. Strong empirical performance. Across 10 datasets and 11 baselines, TabTreeVFM achieves the best overall average rank and the highest C2ST while keeping excess reproduction low. These results show that, with type-specific endpoint targets, boosted trees can outperform strong neural generators in mixed-type tabular generation, a setting where earlier tree-based generators have lagged behind.

Main Concerns and Our Responses

1. Contribution beyond EF-VFM and tree-based generators (dS9i W1; yRXo W, Q3)

The central question was whether TabTreeVFM amounts to a backbone swap. To answer it, we extended the component ablation in Section 4.4 with two neural variants: TabbyFlow, the original implementation of EF-VFM, and TabbyFlow (Aligned), which keeps the same network but adopts the specification of TabTreeVFM. Under the same probability path and solver setting, TabbyFlow (Aligned) reaches a C2ST of 78.10, and boosted trees that regress velocities, as in Forest-Flow, reach 78.62. TabTreeVFM, which combines the endpoint objective with boosted-tree heads, reaches 85.26, so neither component accounts for the gain alone. We also evaluated two additional tree-based generators, ForestDiffusion and Adversarial Random Forests, on all ten datasets. Both fall well behind TabTreeVFM: among the 14 generators, they rank 9th and 7th overall, whereas TabTreeVFM remains first.

2. Positioning of Mem (dS9i W2; wCTT W1, Q1, Q2; yRXo Q1)

We clarified Mem's relationship to existing held-out identical-match comparisons and thresholded DCR. Its contribution is an interpretable, tolerance-aware measure of excess full-row reproduction for mixed-type data, revealing near-copies that exact matching misses. Paired bootstrap analyses confirm lower Mem than TabDiff, TabSyn, CDTD, and TabbyFlow. To investigate capacity effects, we varied tree depth, boosting rounds, and regularization across five datasets. Mem remains at or below 0.28 percentage points on four datasets despite substantial fidelity gains.

AuthorsOpenReview ↗

Summary of Rebuttal (Part 2 / 2)

3. Reliability of the conclusions (dS9i W3, Q1–Q3; wCTT W2)

We jointly evaluated C2ST, coverage, and Mem using the same fitted generator for each method under the 50/50 train–holdout split. TabTreeVFM retains the highest average C2ST and, among generators with average Mem below 1 percentage point, the highest coverage. The new Figure 4 additionally shows that no generator in the original benchmark achieves both higher C2ST and lower Mem. Robustness checks further confirm the conclusions:

  • Training randomness: Three independent training seeds for the four closest baselines; TabTreeVFM remains first.
  • Ranking stability: First in 96.9% of dataset-bootstrap resamples, never below second, and first under the tested alternative metric weightings.
  • DCR ties: Giving ties half credit leaves the overall ranking unchanged.

4. Fairness, scalability, and practical settings (yRXo W, Q2, Q4–Q6; wCTT Q3)

Reviewer yRXo questioned whether CDTD's substantially shorter training time indicated insufficient training. We confirmed that CDTD completed all 30,000 optimization steps prescribed by its official configuration on all ten datasets. To further assess fairness, we evaluated the generators using a common Euler sampler, under which TabTreeVFM retains the highest C2ST and lowest Shape error. For scalability, reducing KK to 10 lowers peak memory to 3.1–10.7 GiB while preserving its lead on C2ST, Shape, and WD. On the larger, higher-dimensional Diabetes dataset (101,766 records, with three categorical columns containing 670–716 categories each), it also ranks first on Shape and WD. Across eight cross-column constraints in finance, e-commerce, and healthcare, TabTreeVFM achieves the lowest violation rate on six. We also clarified the Forest-Flow runtime differences in Appendix L.

Summary

In response to the reviewers' feedback, we have made the following revisions to the manuscript (highlighted in blue):

  • Added TabbyFlow and TabbyFlow (Aligned) to the ablation study, with a new paragraph on the estimator comparison (Table 2 and Section 4.4; yRXo Q3, dS9i W1).
  • Added a plot of C2ST versus Mem for all 12 generators, with the Pareto front (Figure 4 and Section 4.3; wCTT W2, dS9i W3).
  • Clarified the relationship of Mem to held-out identical-match and DCR measures (Section 2; dS9i W2, wCTT W1, yRXo Q1).
  • Added the implementation differences behind the training-time gap to Forest-Flow (Appendix L; yRXo Q4).

In our responses, we have additionally:

  • Evaluated two further tree-based generators, ForestDiffusion and Adversarial Random Forests.
  • Evaluated fidelity, coverage, and Mem on the same fitted models, and tested the ranking under dataset resampling, alternative metric weights, three training seeds, and half credit for DCR ties.
  • Confirmed the training budget of CDTD and the robustness of our conclusions to a common Euler sampler.
  • Showed that TabTreeVFM keeps its fidelity lead with reduced memory (K=10K=10), and evaluated it on a larger dataset with high-cardinality columns and on eight business rules.

We believe these revisions and clarifications substantially strengthen the manuscript and directly address the reviewers' concerns. We appreciate your consideration and hope this summary will be helpful during the discussion and final decision process.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.