acceptodds
Under review as a conference paper at ICLR 2027

Constrained Tabular Generation: Exact Conditioning of Arbitrary Markov Processes

26 views

Abstract

Tabular data generation supports privacy-preserving data sharing, minority-class augmentation, test-database generation, and missing-value imputation. These often require generated rows to satisfy constraints on arbitrary subsets of columns with probability one, while matching the distribution of real rows satisfying them. We study this as *constrained tabular generation*. Existing methods are limited to pinned values, sacrifice distributional fidelity for constraint satisfaction, or restrict exact conditioning to continuous diffusions. Using generator matching, we unify tabular flow, diffusion, and jump models through Markov generators and introduce , a pure-jump model for both numerical and categorical columns. We apply Doob's -transform to general Markov generators, deriving exact conditioning through additive drift corrections for diffusions and multiplicative reweighting of jump rates, both learnable with a consistent objective. This characterizes which processes admit corrections: deterministic flows do not, since the transform leaves the generator unchanged, although in special cases flows admit training-free but biased conditioning using a mean-field approximation of their endpoint posterior. We amortize conditioning through a trainable guide that takes constraints as input, keeps the unconditional generator frozen, and generalizes to unseen constraints across arities, modalities, types, and selectivities. Across eight datasets, our methods improve unconditional fidelity and outperform imputation-, penalty-, and LLM-based baselines in both constraint satisfaction and conditional fidelity.

Primary area:
generative models
Keywords:
Table Generation, Markov Processes, Generative Models

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

This paper studies constrained tabular generation, aiming to generate rows that satisfy arbitrary column-wise constraints while matching the corresponding conditional data distribution. The paper develops a unified generator-level view of mixed-type tabular generative models. Based on this framework, the authors propose TabJump, a pure-jump generator for both numerical and categorical columns, together with a column-symmetric network core that is permutation-equivariant by construction. The paper applies Doob’s hh-transform to derive conditioned generators for general Markov processes, showing that conditioning modifies diffusion drifts additively and jump rates multiplicatively, while pure deterministic flows require a different treatment. Building on these results, the authors develop a conditional generator matching objective and an amortized guidance module that learns conditioning corrections on top of a frozen unconditional generator and generalizes to unseen constraint sets. Experiments evaluate both unconditional and constrained generation on eight tabular datasets.

Critical strengths

1.The generator-level treatment is elegant and provides a unified view of existing tabular models through a common Markov-generator decomposition. This is consequential because it clarifies the relationship of existing approaches and provides a clean foundation for extending.

2.The authors leverage Doob’s hh-transform to derive an exact conditioning method for general Markov generators. They further propose a scalable, model-agnostic learning framework with a consistency guarantee. This substantially strengthens the paper by connecting a rigorous conditioning principle with a practical framework that can be applied across multiple classes of tabular generative models.

Critical weaknesses

1.The novelty of the architecture of column symmetric cores seems limited, and there is not sufficient context compared to previous table modeling work. For example, CM2 (Ye et al, WWW 2024, Section 4.2.2) has already utilized feature level self attention without positional encoding to respect column permutation symmetry. The core proposed in Section 3.2 adopts the same basic principles. Although identifying sequential dependencies in existing table generators and adapting equivariant backbones for generation and constraint guidance is useful, this article should clarify new architectural content beyond this adaptation and discuss relevant precedents such as CM2. This concern limits the novelty of architectural components.

2.The experiential advantage of TABJUMP seems to be concentrated on C2ST-XGB, rather than being consistently proven in evaluation criteria. In Table 2, it achieved the best C2ST-XGB scores on all six datasets, but did not achieve the best trend error or DCR balance on any of them, and only obtained the best MLE results on adults. These results support the strong advantages of XGBoost based discriminator metrics, but provide more limited evidence to demonstrate their broad high fidelity and downstream utility. Further analysis of these differences will help determine which aspects of data distribution TABJUMP performs better than competitive methods.

Questions for authors

1.Can you reconcile the noise correction in equations (13) and (129) with algorithm 4, line 4, where the noise reducer update seems to require −ςtΔχ-\varsigma_t\Delta_\chi instead of +Δχ+\Delta_\chi, and clarify whether this will affect the implementation or reported results? Appendix E.7 and E.8 also seem to have proof errors, whether they are spelling errors or affecting the proof.

2.Is there a situation where constraints are missing or even non-existent? What is the robustness of TABKUMP for this situation?

Reviewer_BhwFOpenReview ↗
Rating
4(+2)
Confidence
4
Informativeness
2
Presentation
2
Soundness
3(+1)
Read review
Paper summary

The paper proposes TabJump a pure-jump Markov-based method for conditional tabular diffusion. The method applies Doob's h-transform and a trainable guide module to satisfy constraints (exactly in theory) while keeping the unconditional backbone frozen. An architectural improvement made is to make the generative backbones column-invariant via self-attention. While the empirical results are promising, the evaluation makes unfair comparisons against the baselines and overstates its ability to handle constraints at inference, prompting me to give this work a weak reject.

Critical strengths
  1. The empirical results are strong

  2. The paper is generally well-written

  3. The problem is relevant, and the work is a well-engineered solution

Critical weaknesses
  1. According to Appendix I.3, the training guidance network trains on a cartesian product of constraints across columns (line 2737). Moreover, according to Appendix H, the numerical bounds for a given column are restricted to a single range e.g. [a, b]. The guide module does not seem capable of handling more complex kinds of constraints like disjunctions (OR). For example, how does the proposed method work if I have a constraint like "24 > Age > 18 OR 28 < Age < 36"? The cited work [1], however, does not have this drawback and can generalise to such forms of constraints. Moreover, it seems that the guidance network takes the constraints as inputs during training, leading me to believe that it cannot generalise to constraints that do not come from the settings it has specifically trained for. To validate the claims, it is necessary to evaluate against more general forms of constraints (e.g. disjunctions, conjunctions, inter-column differences e.g. X_i > X_j, etc. over a subset of columns).

  2. According to the constraint set construction (Appendix I.3), it seems that both training and test set constraints are drawn from a similar rule-based generator. Although the test-time constraints are not exactly the same as the train-time constraints, they are likely coming from a similar in-distribution scenario instead of being drastically different from what was observed during training.

  3. The guidance network needs to train on the constraints in advance, however this cost is not evaluated. Empirical runtime costs quantifying these overhead costs are necessary to give a complete picture of the proposed method's tradeoffs.

  4. The definition of conditional generation used in the paper is somewhat limited. Other works, e.g. [1], consider conditional generation to include imputation constraints, inequalities, conjunctions, or disjunctions of constraints.

  5. The code is not available for review

  6. The background on diffusion and flow matching models is somewhat limited for readers outside the field to follow.

References: [1] Shankar, Aditya, et al. "Harpoon: Generalised manifold guidance for conditional tabular diffusion." International Conference on Learning Representations. Vol. 2026. 2026.

Questions for authors

Please see the weaknesses above. Additionally, I have the following questions

Can the method generalise to arbitrary constraints without training on them? For example, can it handle disjunctions (S_i OR S_j), multiple range bounds on the same column (e.g. Age > 60 or Age < 18), or even constraints across columns (e.g. A - B > C) ?

Reviewer_U5RiOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a comprehensive theoretical framework and algorithmic implementation for tabular data generation, aimed at generating constrainted data that matches the true corresponding distribution. The authors first establish a unified formulation for generating both numerical and categorical data using a Markov process generator, and introduce a generator matching strategy to model the generation path. Based on this, they design a Doob’s h-transform kernel to model the degree of satisfaction for a specific set of conditions S, thereby constraining the generation process. Based on this theoretical framework, the paper adopts a straightforward implementation and training strategy, demonstrating the effectiveness of both unconditional and conditional generation across several datasets.

Critical strengths

The proposed theoretical framework is both versatile and instructive, offering deep insights for future work. First, the Markov process generator provides a unified theoretical perspective on existing methods, including tabular diffusion, flow-based, and jump-based approaches. Furthermore, the conditioned process generator establishes a unified formulation for constrained data generation. Although the method implementation does not fully strictly align with the theoretical framework, the theoretical analysis remains insightful and valuable.

Critical weaknesses
  • Theoretical Convergence is not well-preserved: How does the Markov process generator guarantee convergence under condition SS? Specifically, can the process described by Equation (9) strictly ensure that the generated data matches the target data distribution? Furthermore, Equation (9) essentially introduces a trade-off along the conditional path of the pre-sample generator, where the Doob’s hh-transform kernel modifies the generator's trajectory based on the probability of satisfying condition SS. Under such modifications, how does the model ensure that the final generated samples still adhere to the true underlying distribution?
  • Gap Between Theory and Implementation: There is a noticeable discrepancy between the proposed theory and its implementation. While the theory handles continuous and categorical data separately using continuous flow/diffusion and discrete jump terms respectively, the implementation discretizes continuous features into bins. Consequently, the implementation retains only the jump terms and their corresponding corrections. This discretization may lead to inaccurate modeling for continuous numerical features.
  • Sampling Bias in Sparse Conditional Training: The paper lacks discussion on potential training issues arising from conditional sampling. Sampling data that satisfies condition SS can be biased: while the model can effectively learn conditions well-represented in the dataset (where S has a lot of supporting records), it may struggle with sparse or rare conditions SS due to insufficient training samples (where S only contains limited records or even no record). As a result, the model might only succeed in generating common conditional patterns, while failing or performing poorly under rare or long-tail conditional constraints.
Questions for authors
  • Theoretical Clarification on Convergence: Could the authors provide a more detailed discussion on how the Markov process generator guarantees convergence to true data distribution? Specifically, it would be helpful to clarify whether the data generated via Equation (9) under conditional constraints strictly follows the target conditional distribution P(⋅∣S)P(\cdot \mid S). If exact distribution matching is not guaranteed, could the authors provide a theoretical upper bound or analysis on the worst-case gap between the two
  • Evaluation of Discretization Impact: The discretization of numerical data compromises feature resolution and introduces a noticeable gap with the theoretical formulation. The authors should include empirical analyses evaluating how closely the generated numerical features match the ground truth distribution, using fine-grained distance metrics such as MSE, MMD, or Wasserstein distance.
  • Fine-Grained Analysis on Long-Tail Conditions: To address the concern regarding sampling bias, could the authors split condition sets SS based on their support (sample frequency or number) in the training set? Reporting model performance across high-support versus low-support/sparse condition buckets, rather than presenting only average results. It would help demonstrate how robustly the model handles rare conditional patterns.

Comments on the paper

3 comments

Thank you for your response and most of my concerns are addressed. But considering the nature of ICLR and the contribution of the paper, I keep my score.

AuthorsOpenReview ↗

Unfair Reviewer Comments

We are shocked at reviewer’s vacuous response that undermines the entire purpose of the review process. If the reviewers’ concerns are addressed as they have explicitly acknowledged, a rating of 2 is unjustified. Further their comment “But considering the nature of ICLR and the contribution of the paper” is uncalled for given that Markov conditioning and Tabular Models represent two of the frontiers of modern (generative) ML.

It is also perplexing that just a few minutes before this response (as can be seen from the history), the reviewer in fact agreed and proceeded to updating their rating, only for some inexplicable reason, to go back on their word. We’re sorry to say but this just makes a mockery of the entire process, discrediting all the effort invested by the ICLR organizers, AC,  other reviewers, and the authors.

AuthorsOpenReview ↗

Final Comment by Authors

Final Comment by the Authors

We believe that our submission addresses an important gap in both the tabular generation and generative modelling community. Conditional generation is an important practical task and we believe that we have provided a very comprehensive framework that the entire community stands to gain from. We now summarize our contributions:

Theoretical Impact

First, we provide a unified generator level view of existing tabular models. This identified a gap in the current literature, which we addressed by proposing a pure-jump tabular generator (TabJump) which is seen to perform extremely well in practice. Additionally, this opens up the possibility for the community to build more sophisticated models down the road.

Second, we provide an extremely general theory of conditioning based on Doob's h-transform which, to our knowledge, is novel. We give a rigorous characterization of the processes which can be conditioned, and further provide a model agnostic framework to learn the conditioned parameters. We also believe that this has widespread usage even outside of tables, where exact conditioning is required, and the greater ML community stands to benefit from this work.

Practical Impact

Third, our strong experimental results demonstrate the practical efficacy of our theoretical framework. In particular, our methods outperform baselines both in terms of constraint satisfaction rate as well as fidelity to the conditional distribution across a wide variety of constraint types. We have also addressed the limitations of our methods, and believe that it is exciting future work to resolve them.

The critical strengths of the paper were also highlighted by the reviewers which we now summarize:

  1. Reviewer U5Ri mentions "The proposed theoretical framework is both versatile and instructive, offering deep insights for future work".
  2. Reviewer BhwF found the empirical results strong.
  3. Reviewer HqdF finds "This substantially strengthens the paper by connecting a rigorous conditioning principle with a practical framework that can be applied across multiple classes of tabular generative models."

We also would like to mention that we have provided a comprehensive rebuttal that has addressed the concerns of the reviewers, which is now reflected in their increased support of the work. We also note that our rebuttal addressed the reviewers' concerns. Reviewer BhwF wrote that we "resolved each of my concerns very clearly" and raised their score to a clear accept, Reviewer HqdF wrote that the revisions "addressed my main concerns" and raised their score, and Reviewer U5Ri wrote that "most of my concerns are addressed".

Finally, we would also like to thank the reviewers and the chairs, for their feedback and engagement throughout the process, and their comments which have substantially improved the quality of the paper.

Best, the Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.