acceptodds
Under review as a conference paper at ICLR 2027

GRAFT: A Lightweight Shared Generative Backbone for Efficient Cross-Schema Tabular Synthesis

Abstract

Tabular data underpins decision-making in finance, healthcare, and business intelligence, yet many real-world tables are small, imbalanced, or access-restricted. Predictive foundation models for tabular data have matured rapidly, and a first generation of generative tabular foundation models has emerged. These generators, however, are built on diffusion, transformer, or language-model backbones that are expensive to adapt to a new schema and slow to sample. We ask whether a far lighter design can play the same role where synthetic data matters most. We propose GRAFT, a cross-schema generative model that decouples universal tabular structure from schema-specific detail. A single sub-million-parameter generator, conditioned on a fixed-length meta-feature vector describing the target table, is shared across all schemas. Lightweight per-schema mapping layers project its output into each feature space, and per-schema local discriminators supply specialized feedback without gradient interference. After pretraining on a corpus of heterogeneous public tables, the generator is frozen and adapts to an unseen schema by training only a new mapping layer and discriminator for ten epochs. Under a strict held-out protocol over 86 unseen schemas and three downstream evaluators, GRAFT improves augmented-classifier accuracy over from-scratch CTGAN and TVAE trained for the same or thirty times larger budgets (Wilcoxon p<10−3 in all cases). It also matches the augmentation utility of TabPFGen, a training-free generator built on a pretrained transformer an order of magnitude larger, while producing rows in a single forward pass rather than an iterative Langevin loop. The advantage is concentrated where it matters most: on the smallest fifth of tables GRAFT leads every baseline, including TabPFGen. We also quantify the price of this design, an augmentation-versus-fidelity trade-off.

Primary area:
generative models
Keywords:
tabular data generation, generative adversarial networks, abular foundation models, cross-schema transfer

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_Lmw7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

GRAFT combines a shared pretrained GAN generator with schema-specific mapping layers and discriminators for rapid adaptation, evaluating augmentation utility and synthetic-data fidelity on 86 classification datasets.

Critical strengths
  1. The compact generator, frozen-backbone adaptation, and single-pass sampling offer a practical design for resource-constrained tabular augmentation.
  2. Evaluation with three downstream classifiers and explicit reporting of fidelity limitations help establish the method’s effectiveness and applicability.
Critical weaknesses
  1. Apparently related datasets occur in different folds in Table 7, potentially conflating transfer within the same data source with generalization to unseen tables and weakening the central claim.
  2. Despite targeting low-cost augmentation, the paper lacks comparisons with resampling and SMOTE/SMOTENC, leaving unclear whether cross-table pretraining provides benefits that justify its additional cost.
Questions for authors
  1. After accounting for preprocessing, adaptation, sampling, and amortized pretraining, what is GRAFT’s actual cost advantage, and how many new tables are needed to realize it?
  2. Given the easily distinguishable synthetic samples and weak synthetic-only utility, which sample properties drive the substantial augmentation gains, and under what conditions do these gains fail?
  3. Does the small-data advantage reflect increasing transfer benefits as training data decreases within the same task, rather than domain and difficulty differences across datasets of different sizes?
Reviewer_NM6COpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes GRAFT, a lightweight cross-schema generative model for tabular data synthesis. GRAFT uses a shared meta-feature-conditioned generator across heterogeneous schemas, together with schema-specific mapping layers and local discriminators, allowing adaptation to unseen schemas while keeping the shared generator frozen. The method is evaluated on 86 held-out tabular datasets and compared with CTGAN, TVAE, and TabPFGen in terms of data augmentation utility, fidelity, and computational cost. The results show that GRAFT provides strong augmentation performance, particularly in small-data settings, while substantially reducing per-schema adaptation and sampling costs, although this efficiency comes with a trade-off in synthetic-data fidelity.

Critical strengths
  • The paper addresses an important and practical problem in tabular data generation. In particular, improving data augmentation for small tabular datasets while reducing adaptation and sampling costs is a meaningful problem with clear practical relevance.

  • The paper is generally well written and clearly motivates the proposed approach. The related work provides a good overview of both conventional tabular generative models and recent tabular foundation models, which helps position GRAFT within the broader literature.

  • The experimental setup is reasonably comprehensive and well designed. The evaluation considers 86 held-out schemas, multiple downstream classifiers, several baselines, and multiple evaluation metrics, providing a broad empirical assessment of the proposed method.

Critical weaknesses
  • The ablation study is limited considering that GRAFT introduces multiple important design components. In particular, it is unclear how much meta-feature conditioning and per-schema local discriminators individually contribute to the reported performance. The paper argues that local discriminators help avoid gradient interference across schemas, but this design choice does not appear to be sufficiently validated against a shared-discriminator alternative.

  • The related work misses some recent studies that are closely aligned with the paper's motivation. For example, ReTabSyn: Realistic Tabular Data Synthesis via Reinforcement Learning (ICML 2026) specifically considers tabular generation under small and imbalanced data settings but is not discussed. Including such recent work would help more clearly position the contribution of GRAFT.

  • Figure 1 is somewhat difficult to follow and could be improved for clarity. Clearly separating the pretraining and adaptation stages and more explicitly indicating which components are shared, schema-specific, frozen, or trainable would make the proposed framework easier to understand.

  • GRAFT shows a notable trade-off between augmentation utility and synthetic-data fidelity. While the method performs well for data augmentation, its C2ST scores are substantially higher and its ML-efficacy is generally lower than those of TabPFGen. Although the paper acknowledges this trade-off, it remains unclear how much the limited fidelity restricts the applicability of GRAFT beyond augmentation-focused settings.

Questions for authors
  • How does the performance change when removing meta-feature conditioning or replacing the per-schema local discriminators with a shared discriminator? Such results would help clarify the individual contributions of the main architectural components.

  • How is GRAFT differentiated from recent approaches such as ReTabSyn that also address tabular synthesis in small and imbalanced data settings? A discussion of this relationship would help clarify the novelty and positioning of the proposed approach.

  • Could the authors revise Figure 1 to more clearly distinguish the pretraining and adaptation stages, particularly which components are shared, schema-specific, frozen, and trainable?

  • Given the relatively high C2ST scores and lower ML-efficacy, how do the authors view the applicability of GRAFT to settings where synthetic-data fidelity, rather than augmentation utility, is important?

Reviewer_5xMHOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes GRAFT, a lightweight CTGAN-based generator shared across tabular schemas. A small generator conditioned on dataset meta-features is pretrained on many tables. For each new table, only a mapping layer and a discriminator are trained for 10 epochs, while the generator stays frozen. On 86 held-out datasets, the authors argue that GRAFT outperforms CTGAN and TVAE on augmentation utility and is on par with TabPFGen, but it lags all baselines on ML-efficacy and C2ST, which the authors frame as an augmentation-versus-fidelity trade-off.

Critical strengths

Motivated by a practical problem, the paper proposes a lightweight cross-schema generator that can be used at low cost on small tables.

Critical weaknesses
  1. The claim that augmentation with synthetic data of very low ML efficacy improves performance is not convincing. The authors attribute the improvement to GRAFT generating “typical, broadly plausible” samples that widen the distribution bulk, but this explanation is hard to accept for the following reasons:


    i) The ML efficacy of GRAFT is too low. With XGB in Table 1, TabPFGen achieves 0.727, whereas GRAFT achieves only 0.442. Even in Table 8, where models are trained for 300 epochs, the ML efficacy of CTGAN and TVAE improves substantially, while that of GRAFT remains nearly unchanged. This raises serious concerns about the fidelity of the generated data.


    ii) Considering the number of classes in the 86 evaluation datasets, the average random-guess accuracy is about 0.34, so an accuracy of 0.44 is only marginally above chance level. It is therefore difficult to verify whether the relationship between the generated features and labels resembles that of the real data, and the bulk-coverage explanation for the improved augmentation accuracy is hard to accept.

    The authors should verify whether the data generated by GRAFT faithfully follows the real distribution, e.g., with α-precision/β-recall, distribution. In addition, reporting the real-only baseline (TRTR) and the majority-class baseline alongside the ML-efficacy results would improve the credibility of the paper.

  2. The baselines are insufficient. SMOTE, which augments data by interpolating between neighboring samples within the same class, is another way of filling the distribution bulk, so I recommend adding it as a baseline. Moreover, since the comparison matches budgets by epochs, recent models such as STaSy, TabDDPM, TabSyn, and TabDiff can also be compared. In addition, some TFMs, e.g., TabPFN, support unsupervised (autoregressive) generation, which is provided in the official TabPFN code (tabpfn-extensions).

  3. The main claim of cost-effectiveness is not adequately supported. TabPFGen is not budget-matched, and matching the number of epochs does not imply matching the amount of computation. The authors should report the actual wall-clock time or FLOPs.

Questions for authors
  1. Can the authors verify fidelity of generated samples, e.g., with α-precision/β-coverage or visualizations?
  2. How does GRAFT compare with SMOTE and recent generators (STaSy, TabDDPM, TabSyn, TabDiff, TabPFN generation)?
  3. What are the actual wall-clock time or FLOPs compared with the baselines?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.