acceptodds
Under review as a conference paper at ICLR 2027

Condition-Once Flow Matching for Conditional Multivariate Tabular Generation

Abstract

Tabular data often pair observed covariates with multiple continuous responses whose joint distribution matters beyond a point prediction. We study conditional multivariate tabular generation, where covariates remain fixed while a designated response block is generated jointly, preserving conditional location, uncertainty, and cross-response dependence. Conventional conditional flow models repeatedly supply the same context to the velocity throughout transport. With a learnable conditional source, however, context information can already be encoded into the evolving state, leaving open how much direct velocity conditioning remains useful. We characterize this residual predictive value and show that it vanishes as the state becomes sufficient for velocity prediction, yielding condition saturation. Guided by this result, we introduce Condition-Once Flow Matching (CoFM), which uses a learnable conditional source and a velocity shared across contexts, together with Conditional Source Forcing (CS-Forcing) to control local source expansion over heterogeneous tabular fields. Controlled studies support the predicted saturation behavior. Across four real-world benchmarks covering Beijing air quality, Superconductivity, GTEx, and UK Biobank proteomics, CoFM-CSF consistently surpasses ten competitive probabilistic and generative baselines in overall predictive and distributional performance.

Primary area:
generative models
Keywords:
Conditional generative modeling, Tabular data generation, Flow matching, Probabilistic machine learning, Multivariate prediction

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 2 reviews, above 27% of the others.
Reviewer_a6NBOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a way to conditionally generate continuous tabular columns subject to provided constraints (in the form of other tabular columns). The standard way to generate data via diffusion/flow matching is to start from a white noise prior and gradually shift this towards the data distribution by parameterising the learned backbone on the context/constraints C. This paper proposes an alternative wherein the parameter C is removed from the backbone by changing the initial prior distribution to be conditional on the context. This is achieved by introducing CS-forcing loss to improve training stability under such scenarios. There are theoretical and empirical results to validate the proposed approach and results against the baselines are strong. However, there are several key evaluations which are missing, prompting me to give this work a weak reject.

Critical strengths
  1. The empirical results are promising
  2. The idea is intuitive and the theoretical results offer useful insights.
  3. There are useful ablation studies which help to evaluate the contributions from several angles.
Critical weaknesses
  1. A key weakness of the method is that it assumes the context C will always be available during inference. Hence the method cannot do unconditional generation, i.e., when there is no context provided. There are alternative methods that can train on the full table distribution [1, 2] and constrain their generation on the available context at inference-time, allowing them to do both conditional and unconditional generation. Such approaches can treat the provided context as observed cells (like a tabular imputation task). Comparisons with such methods are absent.

  2. It is unclear if CS-Forcing is generalisable to other approaches. For example, a comparison against TabDDPM or TabDiff using the same CS-Forcing loss would help establish whether the loss function generalises to other approaches of training diffusion models.

  3. CS-Forcing introduces additional parameter complexity and runtime overheads when training the model. This casts doubt over the practical feasibility of the proposed method since these overhead costs are not empirically evaluated.

  4. The definition of conditional generation used in the paper is somewhat limited. Other works, e.g. [2], consider conditional generation to include imputation constraints, inequalities, conjunctions, or disjunctions of constraints. The proposed work is conditional in the sense that the generated columns are conditional on the fixed attributes, but it cannot handle the more general types 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] Zhang, Hengrui, et al. "Diffputer: Empowering diffusion models for missing data imputation." International Conference on Learning Representations. Vol. 2025. 2025.

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

Questions for authors

See weaknesses

Reviewer_9yHkOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The papers deals with learning conditional tabular data distributions with multivariate outputs. Instead of conditioning the velocity of a flow matching model directly on the context, the authors propose to only introduce such conditioning information via the source distribution. This constitutes a factorization of the conditional distribution as p(y∣c)=∫p(z∣c)p(y∣z)dzp(y | c) = \int p(z | c) p(y |z) dz. They construct an objective that encodes similarity in the data space into the conditional source distribution. Limited results indicate gains in predictive accuracy but the explanation for these benefits remains unclear.

Critical strengths
  • Conditional tabular data generation is often only studied in relation to imputation. Tailoring flow matching models even more towards tabular data is important.
  • Encoding similarity in data space into the source distributions seems like a good idea and the ablation on conditional source forcing shows that constraining source distributions to encode similarity of local neighborhoods is beneficial.
  • Results include extensive additional experiments in the appendix with detailed analyses.
Critical weaknesses
  • The paper is confusingly written. It is not obvious why conditioning the velocity field directly on the context is bad, and why doing the conditioning only through the source distribution would be better. Section 3.2 does not indicate that the direction conditioning of the velocity field is generally useless. The results in section 5.4 does indicate that CoFM + CSF performance is better than Dual + CSF performance but it remains unclear why and whether we should expect this to hold in general. On the other hand, results in B2 (Figure S1 Panel C) and B3 (Figure S2 Panel C and) show that the conditional generator remains closer to the target distributions at all times. Since these aspects are the main claim of the paper, I expected much clearer communication.
  • A conditional source distribution for tabular data is nothing new, as for example used in CARD [1] and TabCascade [2]. The current write-up also keeps the actual contribution over CSFM quite opaque.
  • The adaptation of the existing unconditional diffusion model baselines to conditional generation makes no sense to me. Why not just condition the denoiser on that information? That would be a much cleaner and direct mechanism. I do not see the value in wasting modeling capacity by adding context to the data to be denoised and introducing discretization errors, etc.. Also, the DiT variant of the proposed model is more powerful just because of the more powerful backbone compared to the baselines. I expect the baseline performances to be much better when the backbones are aligned and the denoiser is directly conditioned on the context.
  • The used evaluation metrics are very limited. For example, the very popular C2ST (classifier-two-sample-test) metric is missing.
  • Four different datasets are not enough to judge generalizable expected performance, given that tabular data is very heterogeneous.
  • Separating multi-output regression from the conditional multivariate tabular generation (as in Figure 1) is questionable. This is rather a difference of point estimates versus probabilistic estimates.

[1] CARD: Classification and Regression Diffusion Models. https://arxiv.org/abs/2206.07275

[2] Cascaded Flow Matching for Heterogeneous Tabular Data with Mixed-Type Features. https://arxiv.org/abs/2601.22816

Questions for authors
  • Why is direct conditioning of the velocity field bad? Do the claims in B2 (Figure S1 Panel C) and B3 (Figure S2 Panel C and) that show that the conditional generator remains closer to the target distributions at all times no go against these claims?
  • Why did you not add the context information to the baseline denoisers but instead relied on a rather complex, likely worse (through discretization error, etc.), conditioning mechanism?
  • Why does CoFM-CSF outperform Dual + CSF? What happens if you first train CoFM-CSF, fix the source distribution and train a flow matching model with a context-conditioned velocity on that same source distribution?
  • Why do we need the Lvar\mathcal{L}_{var} and Lalign\mathcal{L}_{align} terms for learning the source distribution and not just the KL divergence?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.