acceptodds
Under review as a conference paper at ICLR 2027

Rhythm: Synchronised Corruptions for Continuous Tabular Diffusion

26 views

Abstract

Mainstream tabular diffusion methods use mixed-noise formulations that separately handle discrete and continuous features. While effective at capturing overall distribution trends, these methods have limited control over constrained data generation. We propose Rhythm, a new tabular diffusion paradigm built entirely on a continuous-noise formulation, enabling theory-grounded conditional generation via training-free guidance. Rhythm incorporates a key property of mixed-noise models: when noise is added, different features deviate from their ground-truths at synchronised rates, making all features equally informative. We theoretically characterise this property as corruption alignment and incorporate it into Rhythm by computing feature-wise noise schedules, via an offline, training-free algorithm. Results on diverse datasets show that Rhythm matches existing models on unconditional generation and significantly outperforms mixed-noise methods under conditional generation, highlighting the benefits of continuous diffusion for high-quality and controllable generation. Code: https://anonymous.4open.science/r/BestDiffusion-3B12/README.md.

Primary area:
generative models
Keywords:
diffusion models, tabular data

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_nSqkOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The papers proposes a corruption alignment of features, which tries to ensure that informativeness of each feature is equal at a given timestep. The authors extend results for text diffusion models to include numerical features in tabular data. This is achieved by simple binning of numerical features. Interestingly, these schedules are not learned alongside the model but analytically computed from meta data before training as a correction to a linear reference noise schedule. Results indicate some gains mostly in modeling correlations but this comes at the cost of much lower coverage, probably due to the model relying now on the cooccuring of feature values because of the noise schedule alignment.

Critical strengths
  • It is known that noise schedules are important for the sample quality of diffusion models and this paper deals with noise schedule particularly designed for tabular data, for which there is little existing work.
  • I like that the authors align the MLP backbones for a fairer benchmark comparison.
Critical weaknesses
  • The motivation that a shared noise schedule leads to different levels of corruption among the features is weak. This is the reason why image diffusion models work so well (coarse to fine structure). It is not clear why this is a problem for tabular data and there is no evidence given that this hypothesis leads to an increased difficulty in learning the joint distribution. It could also be beneficial for the model to generate "easy-to-generate" features first, and then the more difficult features. This would require a non-aligned noise schedule.
  • The code base does not use the official implementations of the baseilnes but is fully AI generated. This leads to major deviations from the actual models. For example, continuous time models, like CDTD, are still confined to a grid of only 500 timesteps, instead sampling tt uniformly. This is a major shortcoming, in particular since the code for these models is readily available on github. Visual results in Figures 16, 17, 20 and 21 clearly show that Vanilla, CDTD, CDCD, TabDDPM and TabSyn have not converged, since the generated distribution often stay close to Gaussian. Hence, I do not trust the main benchmark results.
  • No evaluation of C2ST (detection score), despite it being the most important metric to evaluate the quality of the joint distribution and being included in all recent papers. Compared to other papers, the selection of reported metrics is very limited. There is also no metric about privacy.
Questions for authors
  • Why are different levels of corruption bad for tabular data? Is there any evidence that sample quality decreases when not aligning corruption?
  • From Figure 2 and Equation 4c, I see that discrete-noise schedules are always (approximately) aligned. So following that logic, I should always see that models with discrete-noise schedules outperform Gaussian-based diffusion for tabular data?
  • You mention in line 204 that each feature has the same proportion of its maximum achievable entropy. How does entropy come into play here? In Definition 1 the corruption is not defined in terms of entropy. How does this compare to the entropy-based noise schedules (see https://arxiv.org/abs/2504.13612)?
  • How is CDCD implemented? Since the original implementation is only for categorical data. Is the CDTD implementation reused with the CDCD time-warping?
  • Why do you not rely on the official model implementations and meaningful changed the default hyperparameters of the baselines?
Reviewer_iywPOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes RHYTHM, a new tabular diffusion model built entirely on a continuous noise formulation. Its key idea is corruption alignment: features with different cardinalities should lose information at the same relative rate during diffusion. RHYTHM combines two advantages: closes the gap to both mixed-noise methods and methods that learn feature-wise noise schedules during training; its continuous formulation remains adaptive to inference-time constraints, significantly outperforming mixed-noise methods on conditional generation.

Critical strengths
  1. The core idea is simple to understand: different features should lose information at the same relative rate. Instead of assuming that categorical features and continuous features require fundamentally different noise processes
  2. The proposed RIC is a useful diagnostic tool for tabular diffusion. It makes feature-wise corruption dynamics directly comparable and can reveal information imbalance across features.
Critical weaknesses

None

Questions for authors

It would also be interesting to compare RHYTHM with directly using a state-of-the-art LLM to complete the same task. Can the latest generation of LLMs perform tabular data generation or constrained generation without a dedicated diffusion model

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

This paper proposes Rhythm, a continuous diffusion model for tabular data. Its main idea is to align corruption rates across features by computing feature-specific noise schedules without additional training. Experiments on nine datasets show competitive unconditional generation and strong performance on conditional generation.

Critical strengths
  1. I liked the theoretical motivation. The corruption alignment principle provides an interesting way to design feature-wise noise schedules without additional training.

  2. The experiments are reasonably comprehensive. Rhythm performs competitively against strong baselines, and its conditional generation results are promising, especially for categorical constraints.

  3. The structure is very easy to follow.

Critical weaknesses
  1. I noticed a possible error in Algorithm 1. The paper states that corruption increases as the signal level decreases, but the bisection updates in lines 8–11 appear to follow the opposite direction. This should be clarified since it affects the core algorithm.

  2. The extension to continuous features relies on estimated effective cardinality. However, Table 3 shows almost identical performance with fixed and adaptive cardinalities. Also, the conditional generation gains over other continuous diffusion methods, such as CDCD, are relatively small. I am not fully convinced how much improvement comes specifically from corruption alignment.

Questions for authors
  1. Could the authors clarify the bisection update in Algorithm 1? Is this a pseudocode error, or does the implementation follow the same update rule?

  2. Given the limited sensitivity to cardinality estimation in Table 3, what evidence supports effective cardinality as an important part of the proposed alignment principle?

  3. In Table 2, Rhythm and CDCD have similar constraint satisfaction rates for categorical and conjunctive constraints. What is the main practical advantage of Rhythm over this continuous diffusion baseline?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.