acceptodds
Under review as a conference paper at ICLR 2027

DiffLTM: Continuous Diffusions Large Tabular Models for Conditional Generation

42 views

Abstract

Conditional generation is a key capability to be supported by Large Tabular Models, enabling a number of applications. Accurate conditional generation has to contend with two key challenges: Data comprise a mixture of numerical and categorical entries, the latter often with high cardinality and heavily unbalanced proportions; Conditional generation becomes increasingly difficult for rare sub-populations. The best models in this class use a hybrid of continuous and discrete diffusions. In this paper, we design a continuous diffusion model with strong generation quality. An important element is the use of cross-entropy loss for categoricals, for which we provide a first theoretical justification. We introduce metrics that focus on rare subpopulations and extend conditional generation techniques. Finally, we show that our model (trained on labor data) can capture classical insights from econometrics.

Primary area:
generative models
Keywords:
continuous diffusion models, tabular diffusion, conditional generation

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

The paper develops a diffusion model for tabular generation that is capable of conditional generation, and a variant of the Shape metric that is not dominated by common subpopulations (seemingly similar to balanced accuracy in classification). The paper also has a theoretical analysis showing that under suitable assumptions, cross entropy is a better loss than mean-squared error (MSE) for generating categorical variables, contradicting the common idea that MSE is the loss of choice for diffusion models. The paper has experiments comparing the model to baselines on both conditional and unconditional generation, and two case studies showing that that the synthetic data from the model replicates findings in econometric studies.

Critical strengths

The paper is easy to follow and presents clear arguments. The case study showing how the model could be used as a synthetic control is especially informative.

Critical weaknesses

The comparisons experiments are not very comprehensive: the paper only uses four datasets which are all derived from US Census data. For a comparison, Mueller et al. (2026) use 12 datasets from several domains. The selection of baselines is also limited: two for the unconditional case and one (excluding the idealised baseline and variants of the paper's model) for conditional generation. For example, ARF (Watson et al. 2023) is typical baseline missing from the paper. In addition, current tabular in-context supervised learning models like TabPFN (Grinsztajn et al. 2026) are very strong and quickly handle large datasets. Any such model can be used to generate synthetic data unconditionally, or with equality conditions, by sequentially predicting each column from already generated columns (for unconditional generation the first column can be sampled from the data). Synthpop (Nowok et al. 2016) is an example of this approach with tree-based predictors. These are easy enough to deploy due to in-context learning that at least one model of this type should be included as a baseline.

The proposed conditional shape metric is motivated by arguing that it does not dominated by common subpopulations, but this argument is not verified in any way. The definition of conditional shape for numerical columns needs to be clarified. It is not clear how it would be computed based on the description for categorical columns, since that description relies on the fact that there is a finite number of categories.

The paper does not clearly distinguish novel components of the model and training setup, so it is difficult to evaluate how interesting they are to the research community.

Minor points:

  • Figure 3: the lines are hard to distinguish.
  • Figure 3: to my understanding, the wage numbers are inflation-adjusted, but minimum wages are not, so the plotted minimum wage line is not directly comparable with the other lines.

References:

  • Nowok, Beata, Gillian M. Raab, and Chris Dibben. “Synthpop: Bespoke Creation of Synthetic Data in R.” Journal of Statistical Software 2016
  • Watson, David S., Kristin Blesch, Jan Kapar, and Marvin N. Wright. “Adversarial Random Forests for Density Estimation and Generative Modeling.” AISTATS 2023
  • Grinsztajn, Léo, Klemens Flöge, Oscar Key, et al. “TabPFN-3: Technical Report.” arXiv:2605.13986. 2026
Questions for authors
  • The paper weights the numerical and categorical losses equally. Have you experimented with other weightings?
  • What does Figure 3 look like if the y-axis is not cut off at 0.3 and plotted all the way to 1, as CDFs are usually plotted?
4 comments

We thank the reviewer for the remarks and respond to their questions below :

  1. Additional datasets. First of all, note that the datasets we consider are significantly larger scale than in earlier work (of the order of 5M). As such we consider them much more realistic for actual application than earlier studies. Second, we carried out comparisons on four additional datasets. On three out of four of these new datasets, our approach outperforms all alternatives. (This is added to Table 10 in Appendix H)
DatasetMethodShapeTrendCShapeCShape RareMLEDCR ratio
Consumer ComplaintsTabRep88.5975.8873.9371.7598.121.336
TabCascade95.1687.5583.1781.4492.250.969
TabDiff96.19‾\underline{96.19}90.38‾\underline{90.38}84.97‾\underline{84.97}83.34‾\underline{83.34}99.37‾\underline{99.37}1.161
DiffLTM97.9193.0587.0185.3599.461.157
SF Permit TimeTabRep83.6969.5269.3360.6179.541.516
TabCascade87.9178.7072.8463.5161.130.992
TabDiff95.45‾\underline{95.45}90.09‾\underline{90.09}85.05‾\underline{85.05}78.55‾\underline{78.55}81.59‾\underline{81.59}1.217
DiffLTM96.5292.2387.1180.6388.331.018
Diabetes 130-USTabRep89.3980.3383.7776.2190.411.135
TabCascade93.5787.4482.2669.3397.55‾\underline{97.55}0.897
TabDiff98.5996.3393.8188.4593.471.014
DiffLTM98.18‾\underline{98.18}95.82‾\underline{95.82}92.82‾\underline{92.82}86.47‾\underline{86.47}100.231.001
5G Energy ConsumptionTabRep88.6895.8884.3878.3578.735.975
TabCascade84.0390.8480.3278.0793.05‾\underline{93.05}0.933
TabDiff89.15‾\underline{89.15}96.50‾\underline{96.50}86.14‾\underline{86.14}81.53‾\underline{81.53}90.031.143
DiffLTM98.3696.8395.7592.7599.471.002
  1. Additional baselines.

    We carried out new comparisons with two new diffusion based models: CDTD and TabDiff on ACSLabour Dataset (with additional CShape Rare metric). The conclusions are unchanged. See table below. This is also added to Table 1 of the revised paper.

MethodShapeTrendCShapeCShape RareMLEDCR ratio
TabRep93.53 ± 0.0587.00 ± 0.0988.57 ± 0.0587.10 ± 0.0698.43 ± 0.031.000 ± 0.000
CDTD98.86 ± 0.0197.31 ± 0.0294.02 ± 0.0291.87 ± 0.0599.18‾\underline{99.18} ± 0.021.000 ± 0.000
TabDiff98.64 ± 0.1596.00 ± 0.2192.78 ± 0.1590.81 ± 0.1197.29 ± 1.711.000 ± 0.000
TabCascade99.00‾\underline{99.00} ± 0.0197.56‾\underline{97.56} ± 0.0194.10‾\underline{94.10} ± 0.0492.04‾\underline{92.04} ± 0.0698.81 ± 0.031.000 ± 0.000
DiffLTM99.01 ± 0.0397.67 ± 0.0395.40 ± 0.0293.95 ± 0.0199.43 ± 0.021.000 ± 0.000
  1. Sequential generation. We carried out a comparison with one the best tabular foundation model (TabPFN-3.5).

    • DiffLTM achieves substantially better metrics, see Figure 9 in Appendix H of the revised paper.
    • TabPFN is significantly slower at inference time.

    It is incorrect to claim that TabPFN has comparable inference time performance.

  2. Comparison to ARF and Synthpop. We deliberately decided to focus our comparisons to state-of-the-art continuous/discrete diffusion models. We do not think that it is possible to be entirely exhaustive in such a publication.

  3. Conditional shape. This question is very puzzling. It is mathematically trivial that conditional shape is less dominated by common subpopulations Consider for instance a categorical variable taking 4 values, with frequencies (0.7,0.2,0.05,0.05)(0.7,0.2,0.05,0.05). Shape weighs the 3 subpopulations by (0.7,0.2,0.05,0.05)(0.7,0.2,0.05,0.05). Conditional shape weighs them by (0.25,0.25,0.25,0.25)(0.25,0.25,0.25,0.25).

    We also use 'conditional shape on rare' that weighs them by (0,0,0.5,0.5)(0,0,0.5,0.5).

  4. The definition for conditional shape on rare for numerical columns can be given by binning. We will add this in the next iteration.

  5. Weights for numerical and categoricals. We carried out some experiments using different weights, but did not observe an obvious improvement. We believe this is in fact one advantage of using continuous diffusions both on numerical and categorical variables.

  6. CDFs plots. We zoomed into the lower tail since that is where minimum-wage related changes are typically most concentrated. Plotting the full CDF makes these changes harder to visually distinguish, and adds nothing substantial.

Thank you for the response and the additional experiments. They make the performance improvement of DiffLTM compared to other diffusion models much clearer. However, I disagree with some of the points in the response.

We deliberately decided to focus our comparisons to state-of-the-art continuous/discrete diffusion models.

If all tabular diffusion papers did this, the field could miss another model family surpassing them, and get stuck developing suboptimal models. The baselines should be chosen based on the specific problem and operational constraints like required compute, but not the model family they represent.

(I'm not suggesting you should compare against synthpop, I cited it to show that sequential generation is not a new idea.)

It is mathematically trivial that conditional shape is less dominated by common subpopulations.

I don't think it is trivial at all because of possible interactions between columns. For example, consider data with two categorical variables A∈0,1A \in \\{0, 1\\} and B∈1,…,1000B\in \\{1, \dotsc, 1000\\}, where 1% of the population has A=1A = 1, everyone with A=1A = 1 has B=1B = 1, and the BB values for people with A=0A = 0 are uniformly distributed. It's true that the A=1A = 1 subpopulation has equal weight when conditioning on AA, but they only have 0.1% of the weight when conditioning on BB.

The paper also does not explicitly state that a subpopulation is defined based on the value of exactly one column. A subpopulation based on a rare combination of values, where none of the those values is by itself rare, would not have equal weight.

As a result, the claim needs supporting evidence. For example it could be a theorem that formalizes what exactly does "all subpopulation receive equal weight" mean and proves CShape does that. Or it could be a controlled experiment where it is clear what a metric weighting all subpopulations equally should do, and CShape does that, but Shape does not.

Plotting the full CDF makes these changes harder to visually distinguish, and adds nothing substantial.

You could still add the full CDF plots to the Appendix so readers can evaluate whether they have anything substantial themselves.

  1. We agree with the general methodological point that comparisons should not be limited to a single approach. On the other hand, comparisons cannot be exhaustive, and we tried to strike a middle ground. The TabCascade paper has carried out comaparisons with ARF in earlier work and obtained better results with diffusion-based methods. We will test ARF on our datasets as well in the final version of the paper.

  2. We will add a formal statement characterizing the reweighting factor that is introduced by the conditional shape metric, and the sense in which it is more focused on subpopulations.

  3. We will report the full CDFs in Appendix as suggested.

Thank you, this addresses the remaining concerns I had, so I've raised my score.

Reviewer_L8MCOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

Paper Summary

In one sentence: DiffLTM is a diffusion model that generates realistic rows of large census tables, including rows that meet a requested condition (e.g., "speaks no English"), even for rare groups.

Problem:

Diffusion models are built for continuous data, but tables also have discrete (categorical) columns. The paper addresses how to model these discrete columns with diffusion.

A second problem is rare groups: standard quality scores weight every row equally, so poor results on a small group barely change the score.

Goal:

Model discrete columns well with diffusion, generate rows that meet a requested condition, and measure quality so that poor results on rare groups are not hidden.

Method:

The base is standard: discrete values become learned vectors trained with cross-entropy (CE), and conditions use classifier-free guidance (CFG). The new parts are support for range and linear conditions (not just "column = value"), a theorem that CE beats squared error (MSE) in a simplified setting, and a score (CShape) that weights every group equally so rare-group failures show up.

Results:

  • Without conditions, DiffLTM beats the two baselines on all three tables, but often only slightly.
  • CE beats MSE by a wide margin on the largest table, and barely matters on the other two.
  • With conditions, CFG gives results close to filtering DiffLTM's own unconditional samples, even for groups as small as 0.1% of the data.

Difference between this paper and the closest baseline:

Two prior methods are closest. TabDDPM ("TabDDPM: Modelling Tabular Data with Diffusion Models", Kotelnikov et al., ICML 2023, arXiv:2209.15421) is an early, widely used diffusion model for mixed tables; it adds Gaussian noise to numerical columns and uses a separate discrete diffusion for categorical ones. CDTD ("Continuous Diffusion for Mixed-Type Tabular Data", Mueller, Gruber & Fok, ICLR 2025, arXiv:2312.10431) already does what DiffLTM does at its core: one continuous model, with categories turned into learned vectors and trained with CE. DiffLTM's main changes over CDTD are a single noise schedule, range and linear conditions, the CE theorem, and the rare-group score. The paper compares against neither: it does not cite TabDDPM, and it cites CDTD but leaves it out of the experiments. So it is unclear how much DiffLTM actually improves on them.

Critical strengths

Critical Strengths

S1: The paper generates synthetic rows for rare groups and checks that a standard economics analysis gives the same answer on them as on the real data. For the ≈1% of workers who speak no English, the synthetic rows correctly show a much lower return to schooling than in the full population.

Why it matters: It shows the synthetic data are useful for studying rare groups, not just good on summary scores.

Critical weaknesses

Critical Weaknesses

W1: The CE contribution is weak. Cross-entropy (CE) is not new here. The paper itself notes that earlier work, including CDTD, already uses it. The new part is a theorem saying CE is the best loss, but it only covers a toy case (one discrete column, fixed embeddings, a very simple denoiser).

Why it matters: The CE theorem is presented as a main contribution, but it adds no new method.

W2: Key baselines and benchmarks are missing. The paper compares against only two continuous models (TabRep and TabCascade) and tests only on US census tables. Missing baselines:

  • TabDDPM (Kotelnikov et al., ICML 2023): an early, widely used diffusion model for mixed tables.
  • TabSyn (Zhang et al., ICLR 2024): latent diffusion for mixed tables; the paper cites it only for metrics.
  • CDTD (Mueller et al., ICLR 2025): the closest method, cited but not compared (TabCascade is a later model by the same authors, not CDTD itself).
  • TabDiff (Shi et al., ICLR 2025): the state of the art, which the paper itself names.

No standard tabular benchmark (e.g., Adult, Default, Shoppers, Magic, Beijing, News) is used. For conditional generation, the only baseline, Harpoon, is run on the weakest DiffLTM variant, not the one used by CFG.

Why it matters: Without these comparisons, the state-of-the-art claim is not supported.

Questions for authors

Questions For Authors

See weaknesses.

1 comment

We thank the reviewer for the remarks and respond to their questions below :

  1. Theory for Cross-Entropy loss.

    The referee argues that our analysis of CE loss is not valuable because CE loss is already used in some applied papers. By this argument, the vast majority of theory papers that are regularly published in top ML conferences should be rejected, because they typically analyze existing methods.

    To the best of our knowledge, this is the first result showing that cross-entropy is asymptotically optimal for categorical variables in continuous diffusions reconciling the empirical preference for CE with theory. Earlier theoretical results in fact favored MSE loss (Chewi et al., 26). We solve this conundrum by identifying the key role of misspecification.

    Just this result, by itself, would justify acceptance.

  2. We carried out additional comparisons with CDTD and TabDiff. Our conclusion stand. As an example, below we report the result of such a comparison for the ACS Labour Dataset (with additional CShape Rare metric). This is also added to Table 1 of the revised paper.

    The referee correctly point out that our approach is methodologically related to CDTD. However the comparison show a clear advantage for our approach. | Method | Shape | Trend | CShape | CShape Rare | MLE | DCR ratio | |---|---:|---:|---:|---:|---:|---:| | TabRep | 93.53 ± 0.05 | 87.00 ± 0.09 | 88.57 ± 0.05 | 87.10 ± 0.06 | 98.43 ± 0.03 | 1.000 ± 0.000 | | CDTD | 98.86 ± 0.01 | 97.31 ± 0.02 | 94.02 ± 0.02 | 91.87 ± 0.05 | 99.18‾\underline{99.18} ± 0.02 | 1.000 ± 0.000 | | TabDiff | 98.64 ± 0.15 | 96.00 ± 0.21 | 92.78 ± 0.15 | 90.81 ± 0.11 | 97.29 ± 1.71 | 1.000 ± 0.000 | | TabCascade | 99.00‾\underline{99.00} ± 0.01 | 97.56‾\underline{97.56} ± 0.01 | 94.10‾\underline{94.10} ± 0.04 | 92.04‾\underline{92.04} ± 0.06 | 98.81 ± 0.03 | 1.000 ± 0.000 | | DiffLTM | 99.01 ± 0.03 | 97.67 ± 0.03 | 95.40 ± 0.02 | 93.95 ± 0.01 | 99.43 ± 0.02 | 1.000 ± 0.000 |

  3. We also note that TabDDPM and TabSyn have already been shown in the literature to be dominated by the methods we compare with. We carried out preliminary experiments with those, and conclude that they are not representative of the state-of-the-art.

  4. Datasets such as Adult are very small, and not representative of large-scale applications. The results we observe on such datasets are consistent with the ones on large datasets but much noisier.

    [There are strong methodological arguments for actually not using Adult, see: Ding F, Hardt M, Miller J, Schmidt L. Retiring adult: New datasets for fair machine learning. Advances in neural information processing systems. 2021]

    Below we carry out comparisons on 4 additional small-scale datasets. For three out of four datasets our approach is superior to all the others. (This is added to Table 10 in Appendix H)

DatasetMethodShapeTrendCShapeCShape RareMLEDCR ratio
Consumer ComplaintsTabRep88.5975.8873.9371.7598.121.336
TabCascade95.1687.5583.1781.4492.250.969
TabDiff96.19‾\underline{96.19}90.38‾\underline{90.38}84.97‾\underline{84.97}83.34‾\underline{83.34}99.37‾\underline{99.37}1.161
DiffLTM97.9193.0587.0185.3599.461.157
SF Permit TimeTabRep83.6969.5269.3360.6179.541.516
TabCascade87.9178.7072.8463.5161.130.992
TabDiff95.45‾\underline{95.45}90.09‾\underline{90.09}85.05‾\underline{85.05}78.55‾\underline{78.55}81.59‾\underline{81.59}1.217
DiffLTM96.5292.2387.1180.6388.331.018
Diabetes 130-USTabRep89.3980.3383.7776.2190.411.135
TabCascade93.5787.4482.2669.3397.55‾\underline{97.55}0.897
TabDiff98.5996.3393.8188.4593.471.014
DiffLTM98.18‾\underline{98.18}95.82‾\underline{95.82}92.82‾\underline{92.82}86.47‾\underline{86.47}100.231.001
5G Energy ConsumptionTabRep88.6895.8884.3878.3578.735.975
TabCascade84.0390.8480.3278.0793.05‾\underline{93.05}0.933
TabDiff89.15‾\underline{89.15}96.50‾\underline{96.50}86.14‾\underline{86.14}81.53‾\underline{81.53}90.031.143
DiffLTM98.3696.8395.7592.7599.471.002
  1. Harpoon and DiffLTM variant. We repeated our experiment using the stronger version of DiffLTM to benchmark Harpoon, see Figure 6 in Appendix D.4 of the revised paper. We observe only a slight improvement for Harpoon, and no change in the overall conclusion.
Reviewer_b3LPOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents a method for conditional tabular diffusion. It proposes a continuous noise architecture that combines MSE loss (for numerical features) with cross entropy loss (categoricals) for training the backbone. Theoretical results are presented to justify the use of cross-entropy over MSE. Conditional generation is performed via classifier-free guidance by training the model on a large set of pre-defined condition types. Results show that the method achieves high constraint satisfaction rates under the evaluated condition scenarios while retaining the overall distribution quality. However, the work has several limitations regarding the novelty, evaluation scope, and model design, prompting me to give this a reject.

Critical strengths
  1. Although I have not checked in depth, the theorem justifying the use of cross-entropy over MSE for a continuous backbone seems useful.

  2. The empirical results are promising.

  3. The paper tackles a difficult yet relevant problem.

Critical weaknesses
  1. Minimal novelty over prior work: Methodologically the proposed approach is not novel. It combines a continuous diffusion backbone with cross-entropy and MSE loss, similar to prior work [1]. It also does not present a novel approach for conditional generation, since it simply applies the well-studied classifier-free guidance approach.

  2. The paper claims in Sec. 5.2 that Harpoon [2] cannot handle categorical and numerical constraints. I checked that paper and found this to be not true. In Sec. 4.3, Table 3, the authors show that Harpoon handles conjunctions and disjunctions of categorical and numerical constraints. I therefore additionally checked the claim made in Appendix D.4 that Harpoon cannot handle linear constraints. However, this is not true either. In the appendix of Harpoon, the paper presents a ReLU-based loss formulation to handle some constraints, leaving it to the user to formulate a loss for their custom inference-time constraints. I believe linear inequalities can be represented via a loss function of the form presented in Harpoon, hence there is insufficient justification to not compare against it on all settings.

  3. The proposed approach to handle constraints uses classifier free guidance, which trains on a vast set of pre-defined conditional scenarios (specified in Appendix D). This is very expensive compared to inference-time guidance methods, so I find the justification weak. Moreover, the proposed approach works by sampling 1, 2, 3, or 4 continuous or categorical columns randomly to impose a constraint. Why only 1-4 columns and not something higher or all possible columns? Moreover, the proposed approach can only handle conjunctive constraints, e.g. (25 < Age < 30) AND Job = 'military', etc. What if the constraint is disjunctive, e.g. (25 < Age < 30 OR 10 < Age < 20)? The cited work, Harpoon, appears to be able to flexibly handle such constraints too via an inference-time loss. Hence, in terms of generalisability and capability to handle arbitrary types of constraints, I would argue that the proposed method is not very flexible.

  4. Continuing point 3, why use classifier free guidance instead of directly applying inference-time guidance [2]? While training using a cross-entropy version of the proposed method

  5. The paper states in the introduction that a Large Tabular Model's (LTM) key component is the ability to learn across datasets. The paper claims its method (DiffLTM) is an LTM. However, in the limitations, it is also mentioned that cross-dataset capabilities are not present in the proposed method. Therefore, calling it an LTM is misleading.

  6. The proposed approach of training a conditional generator seems prohibitively expensive. It would be useful to compare the runtimes (total training and sampling) of DiffLTM against the baselines.

  7. The structure of the paper can be improved. Currently, it presents metrics and results before the main method section which disrupts flow.

  8. The background on diffusion and flow models are very little, making it hard for a reviewer outside this area to follow.

  9. 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.

References

  1. Mueller, Markus, Kathrin Gruber, and Dennis Fok. "Continuous diffusion for mixed-type tabular data." 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 above weaknesses.

5 comments

We thank the reviewer for the remarks and respond to their questions below :

  1. Methodological novelty.

    We agree that the approach we follow is quite simple, but we consider this an achievement rather than a drawback. In fact we simplified away so-called 'methodological innovations' introduced in earlier papers. We find that removing those innovations results in a simpler, more robust design that outperforms those.

    It is unfortunate that AI conferences have a strong bias towards 'innovations' that have no solid backing.

  2. Harpoon (equality and ranges).

    It is factually false that we claim that Harpoon cannot handle categorical or numerical constraints. We only claimed that it cannot handle arbitrary such constraints.

    In our experiments, Harpoon is substantially suboptimal in equality and range conditions. (see Figure 1 and other results in Appendix H)

  3. Harpoon (linear inequalities).

    As suggested, we carried out experiments with linear inequality condition.

    Harpoon is substantially suboptimal on linear inequality conditions as well. (Fig 6 in App D.4 of the revised paper)

  4. Inference-time versus train-time intervention.

    It is factually false that we only use train-time interventions: we also identify an inference time intervention that performs comparably, SMC.

    SMC is inference-time and performs better than Harpoon while affording all the abilities provided by Harpoon (Fig 6, App D.4 of the revised paper), and categorical OR/NOT conditions (Table 9, App H of the revised paper). | Slice (ACSLabour) | Rarity (%) | Shape (SMC) | Shape (Ideal) | |---|---:|---:|---:| | ST≠CA\mathrm{ST} \neq \mathrm{CA} | 88.47 | 98.37 | 99.08 | | ST=TX∨SEX=female\mathrm{ST} = \mathrm{TX} \lor \mathrm{SEX} = \mathrm{female} | 52.85 | 98.37 | 99.11 | | ST≠CA∧SEX=female\mathrm{ST} \neq \mathrm{CA} \land \mathrm{SEX} = \mathrm{female} | 43.13 | 98.41 | 99.12 | | (ST=AZ∧SEX=female)∨(ST=TX∧SEX=male)(\mathrm{ST} = \mathrm{AZ} \land \mathrm{SEX} = \mathrm{female}) \lor (\mathrm{ST} = \mathrm{TX} \land \mathrm{SEX} = \mathrm{male}) | 5.20 | 98.09 | 99.12 | | ST=MT∨COW=7\mathrm{ST} = \mathrm{MT} \lor \mathrm{COW} = 7 | 4.45 | 97.63 | 98.89 | | ST=AZ∨ST=NC\mathrm{ST} = \mathrm{AZ} \lor \mathrm{ST} = \mathrm{NC} | 4.98 | 97.95 | 99.15 |

Additionally, SMC enables conditioning on arbitrary numerical conditions (Table 8, App H of the revised paper).

Slice (ACSLabour)Rarity (%)Shape (SMC)Shape (Ideal)
(AGEP−4310)2+(log⁡wage−10.60.5)2≤0.64\left(\frac{\mathrm{AGEP}-43}{10}\right)^2 + \left(\frac{\log \mathrm{wage}-10.6}{0.5}\right)^2 \leq 0.649.4897.5699.09
0.25≤(AGEP−4310)2+(log⁡wage−10.60.5)2≤0.640.25 \leq \left(\frac{\mathrm{AGEP}-43}{10}\right)^2 + \left(\frac{\log \mathrm{wage}-10.6}{0.5}\right)^2 \leq 0.645.9197.3298.97
(AGEP−43)(WKHP−40)<−200(\mathrm{AGEP}-43)(\mathrm{WKHP}-40) < -2006.9096.7599.08
60≤exp⁡(log⁡wage)52 WKHP≤9060 \leq \frac{\exp(\log \mathrm{wage})}{52\,\mathrm{WKHP}} \leq 903.6396.7098.73
  1. We trained using a cross-entropy version of the model in conjunction with Harpoon (see Fig 6 in App D.4 of the revised paper). We observe a slight improvement in Harpoon performance.

    However, Harpoon is still substantially inferior to CFG (train time) and SMC (inference time), and hence our conclusions are unchanged.

  2. The number of columns to condition on is an hyperparameter during training and can be optimized.

  3. It is factually false that training a conditional generator seems prohibitively expensive.

    The additional complexity of training using CFG is is negligible, only 16% increase in train time, see table below (we report ms per gradient step.

    Modelms/updateParams
    DiffLTM-CFG40.830 ± 0.01110,365,901
    DiffLTM34.969 ± 0.0099,839,322

    We use exactly the same number of samples for training CFG than the original model, each time sampling a random condition. The only increase in training time is due to the fact that the model is slightly more complex.

    We observe that Harpoon is 2x slower at inference time, which is a serious issue in large scale deployments. In table below we show inference-time comparison. We report the time to generate 1,000 rows(both tables are added to Appendix D.4 of the revised paper.)

    MethodSec/1,000 rows
    DiffLTM0.213 ± 0.001
    DiffLTM-CFG0.237 ± 0.001
    Harpoon0.460 ± 0.003
  4. Let us also mention actual enterprise-scale applications it is often much preferable to invest compute at train-time rather than at inference time.

  5. In actual applications, good accuracy is of paramount importance. We observe that Harpoon is substantially surpassed both by SMC and CFG.

    In particular, our implementation of SMC covers all of the conditions mentioned by the referee, achieving excellent accuracy.

  6. We will improve the writing as suggested.

  7. We agree that multi-table training is important. This can be achieved by DiffLTM with some architectural adjustements, and we plan to presenting them in the next paper in this series.

Reviewer_b3LPOpenReview ↗

Reviewer response

1: We agree that the approach we follow is quite simple, but we consider this an achievement rather than a drawback. In fact we simplified away so-called 'methodological innovations' introduced in earlier papers. We find that removing those innovations results in a simpler, more robust design that outperforms those. It is unfortunate that AI conferences have a strong bias towards 'innovations' that have no solid backing.

Thanks, I agree that simplicity is indeed a strength. Regarding your remark on AI conferences, I find it somewhat ironic given that you criticise the community while making an equally sweeping, unsubstantiated generalisation about that very community.

2: It is factually false that we claim that Harpoon cannot handle categorical or numerical constraints. We only claimed that it cannot handle arbitrary such constraints. In our experiments, Harpoon is substantially suboptimal in equality and range conditions. (see Figure 1 and other results in Appendix H)

Your evidence only supports the claim that the current Harpoon implementation may be suboptimal at handling arbitrary constraints, but not that it cannot handle them. After all, Harpoon is an inference-time method that provides the necessary theoretical framework for incorporating arbitrary constraints. A better implementation can still use its theoretical framework to handle arbitrary constraints, as you show in Appendix D.

3: As suggested, we carried out experiments with linear inequality condition.

Thanks!

4: It is factually false that we only use train-time interventions: we also identify an inference time intervention that performs comparably, SMC. SMC is inference-time and performs better than Harpoon while affording all the abilities provided by Harpoon (Fig 6, App D.4 of the revised paper), and categorical OR/NOT conditions (Table 9, App H of the revised paper).

My concern is that the current experiments confound the choice of inference-time intervention with the underlying model architecture. In particular, the comparison is between Harpoon's architectural implementation and DiffLTM with SMC, making it difficult to attribute the observed performance differences to the intervention itself. For a fairer comparison, I would like to see what happens when applying Harpoon- or DPS-style guidance to DiffLTM as well. This would help establish whether the reported improvements are due to SMC itself or differences in the underlying implementations.

Regarding the results in Table 9, could you also include the results for the CFG approach? I would be interested to see whether the pre-trained CFG model can also accommodate these newly introduced conditions, without requiring additional training or fine-tuning.

5: We trained using a cross-entropy version of the model in conjunction with Harpoon (see Fig 6 in App D.4 of the revised paper). We observe a slight improvement in Harpoon performance.

Thanks! This goes back to my point above that the suboptimal performance of the original Harpoon implementation does not imply that the approach itself is inadequate. A better implementation that adopts the theoretical framework for handling arbitrary constraints could still do better.

6: The number of columns to condition on is an hyperparameter during training and can be optimized.

How and when? It is impossible to foresee all possible conditions that may arise in the future. One can pretrain the CFG model on X numer of columns, but the inference-time requirement may substantially diverge from this. How do you propose to scale the approach if this is the case?

7: Results on efficiency

Thanks for the additional results.

8: Let us also mention actual enterprise-scale applications it is often much preferable to invest compute at train-time rather than at inference time.

This does not address my concern about scalability to previously unseen constraints, e.g. Table 9 (App. H). If accommodating new constraints requires retraining the CFG model, then the computational expense is incurred repeatedly as new requirements arise. How do the authors propose to address this?

9: In actual applications, good accuracy is of paramount importance. We observe that Harpoon is substantially surpassed both by SMC and CFG. In particular, our implementation of SMC covers all of the conditions mentioned by the referee, achieving excellent accuracy.

I agree that the SMC implementation can handle these constraints. I have my doubts of the pre-trained CFG implementation's capability at handling these constraints.

11: We agree that multi-table training is important. This can be achieved by DiffLTM with some architectural adjustements, and we plan to presenting them in the next paper in this series.

If cross-dataset learning is future work, then calling the current method an LTM is misleading.

Overall, thanks for your answers! I will gladly reconsider my score if my remaining concerns are satisfactorily addressed.

  1. The main point raised by the referee is that Harpoon is compared to our methods using a different implementation of the denoiser backbone, or the training loss. This is incorrect. As we wrote explicitly in Appendix D.4 `... we apply Harpoon (Shankar et al., 2026) guidance to the MSE-loss, representation-mode variant of DiffLTM. This allows controlled comparison of conditional generation capability in which the denoiser architecture and diffusion process....'. Also, as we wrote explicitly in the rebuttal, we replaced the loss as well to use CE loss, and this did not change our conclusions. We presented results in the rebuttal backing this claim.

  2. Summarizing:

(a) Harpoon is a variant of DPS applied to tabular data.

(b) We have an inference time method: SMC for tabular data.

(c) Our experiments are carried out with the same denoiser implementation and same loss. We did our best to be as fair and careful as possible.

(d) Our conclusion is that the new inference-time approach performs better all other factors being equal.

Of course there could be completely different implementations on which this order is reversed, but this is always the case with empirical science.

  1. Scalability of CFG. TL;DR: In all of our experiments, with varying number of condition, training CFG took less than 2020% more GPU-hours than training the unconditional model. Irrespective of the number of conditions.

(a) As mentioned in the rebuttal, we do not scale the number of gradient steps or the batch size with the number of conditions. If nn is the number of training samples, and ncondn_{\textrm{cond}} is the number of conditions, each condition is applied to n/ncondn/n_{\textrm{cond}} samples.

(b) Further, we do not scale the number of conditions at train time with the number of conditions at test time. Say that the table has mm columns. Even if we can condition to any of the 2m2^m subsets, at train time we do not need to train on each of 2m2^m conditions.

(c) The previous point might appear `magic,' but this is the standard magic of machine learning. If we want to learn a function {+1,-1}m→R^m\to {\mathbb R}, we only need 2m2^m samples in the worst case, but not for simple enough functions.

  1. Applicability of CFG. We agree that CFG cannot adapt to new `classes' of conditions not seen in training, for instance disjunctions if it was trained only on conjunctions. However

(a) CFG can cover an important predefined set of constraints that are of most practical importance. (In particular, linear and imputation.)

(b) On those conditions, it outperforms all other approaches.

(c) For other conditions, we adapt SMC.

(d) SMC outperforms Harpoon.

  1. Terminology. We view LTM as an ``umbrella'' term that can cover multiple modalities and tasks. We believed we make our contribution very clear since the abstract, but are happy to clarify it further.

Comments on the paper

1 comment
Area_Chair_KvT4OpenReview ↗

Review discussion

Dear reviewers,

the authors have responded to your reviews. If you haven't yet done so, please check their responses and engage in a discussion with them to clarify any open questions or help them address your concerns. I also strongly recommend you check the reviews and answers of your fellow reviewers and take them into consideration during the discussion so that we can preferably reach a consensus at the end of the review period.

Thank you, your AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.