acceptodds
Under review as a conference paper at ICLR 2027

CAST: Conditional Generation and Adversarial Selection for Tabular Data Synthesis with Long Text Columns

Abstract

Synthesizing tabular data with long text columns requires preserving structured attributes, text semantics, and the predictive relationships between them. Existing methods improve individual records through generator fine-tuning or post-generation filtering, but high record-level quality *does not ensure that the resulting table preserves text–attribute relationships at the dataset level*. We propose **CAST**, a conditional generation and adversarial selection framework that keeps the text generator frozen and combines candidate construction under structural constraints with dataset-level selection. Structure-Aware Semantic Generation (SSG) builds multiple text candidates from a shared semantic plan and distinct narrative roles. Conditional Adversarial Selection (CAS) reselects candidates by matching task features to those of real references, with feedback from the current synthetic table and a fixed domain discriminator. Reliability-Guided Conservative Selection (RCS) gates and reweights task feedback using validation performance, then mixes selected candidates with preset-role candidates to reduce reliance on proxy scores. Across three tabular datasets with long text columns, CAST achieves higher downstream utility than the compared synthesis methods, with gains across different generator backbones and modeling strategies.

Primary area:
generative models
Keywords:
Tabular Data Synthesis, Long-Text Tabular Data, Large Language Models, Synthetic Data Generation

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 2 reviews, above 68% of the others.
Reviewer_yuLSOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

CAST generates text candidates through shared semantic plans and narrative roles, then assembles synthetic tables using conditional reference matching, fixed-discriminator feedback, and reliability control while keeping the generative LLM frozen. Experiments across three datasets, multiple generators, and downstream modeling strategies report consistent utility improvements, with fixed-pool comparisons providing further evidence for the selection strategy.

Critical strengths
  1. Combining condition-compatible candidate construction with table-level selection offers a practical approach to text-rich tabular synthesis without generator fine-tuning.
  2. Improvements across tasks and fixed-pool comparisons provide convincing evidence that selection offers benefits beyond simply strengthening target-label cues.
Critical weaknesses
  1. SSG relies on prompts to preserve shared content, while validity checks do not directly detect evidence distortion or cross-field semantic contradictions, so utility gains do not sufficiently establish record-level semantic fidelity.
  2. Direct association-fidelity evaluation covers only sufficiently populated Cloth condition groups and uses the same types of predictive features as selection, leaving rare combinations and broader cross-field relationships insufficiently assessed.
Questions for authors
  1. When synthetic and reference records have different auxiliary-attribute labels, how does the predicted-class agreement term in Equation (6) avoid encouraging text inconsistent with the synthetic structured attributes?
  2. When an attribute’s sole valid predictor barely exceeds its baseline, Equation (4) still assigns it the full feedback budget, so how does RCS prevent weak proxies from disproportionately influencing selection?
  3. Does Table 11’s 102-minute runtime include retrieval, planning, four-candidate generation, and selection, and is its end-to-end scope comparable to AFT-Tab’s reported 157-minute training overhead?
5 comments
AuthorsOpenReview ↗

Official reply to Reviewer yuLS [1/2]

We thank Reviewer yuLS for the positive assessment and the precise questions on Equations (4) and (6). We address each concern below.

W1. Record-level semantic fidelity.

Ans: A direct audit confirms record-level fidelity: synthetic text rarely contradicts its structured fields or reverses the evidence in its shared semantic plan.

Using Claude Opus 5.5 as the judge, we audited 1,254 real records and 1,253 synthetic rows, covering every candidate of each synthetic row (5,012 in total). The judge was blind to the narrative role, selection status, and method. It flagged two types of error: contradictions between text and structured fields, and reversals of the plan's evidence. Table R6 reports the rates.

DatasetContradiction rate (synthetic)Contradiction rate (real)Evidence reversal rate (synthetic)
Cloth0.48%3.57%2.14%
Wine0.12%0.25%2.77%
Kickstarter1.10%0.00%2.13%

Both errors are rare on all three datasets, and on Cloth and Wine synthetic text contradicts its structured fields less often than real text does. Synthetic text is therefore consistent with both its structured fields and the evidence from which it was written. We have added the audit protocol and results to Appendix D.4.

W2. Coverage and independence of the association evaluation.

Ans: Beyond the populated Cloth groups, CAST is closer to the real data than AFT-Tab in every extended evaluation, and both evaluations are independent of selection.

We extended the evaluation to the sparse Cloth groups outside the main support, merged into one bucket, and to Wine and Kickstarter. We also added an association gap that covers every auxiliary attribute without group thresholds. It trains a model on synthetic text, tests it on real data, and compares it with a model trained on real text. Table R7 reports all results, with a split of the real data as the floor (lower is better).

EvaluationReal split (floor)AFT-TabCAST
Conditional W1, Cloth populated groups (31 groups, 85.1% of test records)0.07590.10460.1075
Conditional W1, Cloth sparse groups0.05300.20730.0645
Conditional W1, Wine0.05230.28370.2498
Conditional W1, Kickstarter0.01270.36480.0529
Association gap, Cloth0.00390.03730.0063
Association gap, Wine0.00240.17450.1324
Association gap, Kickstarter0.00120.31690.1223

AFT-Tab is slightly closer only on the populated Cloth groups, which is expected because its discriminator directly targets text–structure consistency [1]. On all other evaluations CAST is clearly closer, most notably on the sparse groups and on Kickstarter.

Both evaluations are independent of selection. The conditional-W1 evaluator (character n-grams, SVD, and ExtraTrees) is fit only on real data. The association gap uses the same model family, trained separately on real and synthetic text. Selection, by contrast, relies on word-level TF-IDF text predictors. We have added Table R7 to Appendix E and described the evaluators, support, and coverage in Appendix C.2.

Q1. The predicted-class term in Eq. (6).

Ans: We thank the reviewer for this careful reading of Eq. (6). The reviewer is correct that, for an attribute outside the condition key, the term can favor a text whose predicted class matches the reference rather than its own label. In practice, however, the term changes few selections. To quantify this, we removed the term for one attribute at a time and re-ran selection on the fixed pool. Table R8 reports the number of changed selections and how many of them occur where the candidate and reference labels differ.

AttributeLabel mismatch (candidate vs. reference)Effective weight (after gating)Selections changedChanges where labels differ
Cloth: Division Name43.40%00 (0%)0
Cloth: Department Name0.13%0.35306 (2.16%)1
Wine: points91.21%0.351,935 (3.06%)1,700
Wine: country1.95%0.352,038 (3.23%)47

On Cloth, the issue does not arise in practice. Division Name, the attribute with the most label mismatch, fails gating. For Department Name, the changes occur almost exclusively where the labels agree. In these rows, the term keeps the own-label accuracy of the selected text close to that of real records.

The issue does arise for ordinal attributes such as Wine points. Even there, it affects only a small fraction of selections, the term carries a small weight (β·W_a ≈ 0.12), and the expected-level residual already compares levels. We have documented this case in Appendix B.3.

AuthorsOpenReview ↗

Official reply to Reviewer yuLS [2/2]

Q2. Budget allocation in Eq. (4).

Ans: RCS excludes weak proxies through gating, while the fixed budget W_a preserves weak but genuine associations in the objective. Scaling the budget by reliability yields no gain beyond seed-to-seed variation. To test this, we compared three budget rules on Cloth with the candidate pool fixed. Table R9 reports the utility of the default rule and the change under each alternative.

Budget ruleUtility (Cloth)
RCS (default): gating, then reweighting within a fixed W_a65.63
W_a scaled by reliability+0.17
w/o RCS + final mixing: no gating, equal split (Table 27)−0.55

Scaling the budget changes utility by +0.17, within the seed-to-seed standard deviation of Full CAST (0.41, Table 21), whereas removing RCS lowers utility in all three seeds (Table 22). It is therefore gating that excludes unreliable feedback.

Keeping W_a fixed is a deliberate choice. A budget that shrinks with reliability would push weakly associated attributes out of selection, whereas faithful synthesis should preserve such associations, in line with the association-fidelity criterion of AFT-Tab [1]. In our runs, the only attribute ever left with a single valid term was Wine price, whose reliability is well above the gating threshold. We have added Table R9 and the gating statistics to Appendix E.

Q3. Scope of the 102-minute GPU time.

Ans: Thank you for pointing out this ambiguity. The 102 minutes are the total GPU time of CAST's full pipeline on Wine Reviews, covering retrieval, semantic-plan extraction, four-role generation, feedback-predictor fitting, and selection. All methods in Table 11 follow the same four-GPU accounting.

Most of this time is spent on plan extraction and generation, and the LLM is never updated. AFT-Tab's 157 minutes cover its training alone, before any generation, so CAST's full pipeline requires less GPU time than AFT-Tab's training. We have revised the corresponding sentence in Appendix D.1 to state both scopes.

Reference

[1] Zhang Y, Yan L, Duan S, et al. AFT-Tab: Adversarial Fine-Tuning for Tabular Data Synthesis with Long Text Columns. ACL 2026: 4581–4594.

AuthorsOpenReview ↗

Follow-up on our response

Dear Reviewer yuLS,

Thank you again for your positive assessment and your precise questions. We understand that this is a busy period, and we truly appreciate your time.

As the discussion period will close soon, we would like to gently check whether our responses have addressed your questions. In particular, Tables R6 and R7 address record-level fidelity and the coverage of the association evaluation, and Tables R8 and R9 analyze Equations (6) and (4).

If anything remains unclear, we would be glad to clarify it.

Best regards,

The Authors

Reviewer_9y7MOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

CAST proposes a candidate-based long text tabular synthesis framework, in which frozen LLMs first generate multiple text candidates under fixed structural properties, and then the conditional selection process uses task specific predictor feedback, real data reference, and real synthetic domain discriminator to select from them. A reliability-guided mechanism further reweights these feedback signals based on validation performance to reduce reliance on unreliable proxies.

Critical strengths

1.CAST explicitly addresses the limitation of conventional tabular generators that mainly model numerical and categorical attributes while neglecting the semantic dependencies between structured attributes and long-form text. The proposed separation of candidate construction and candidate selection provides a coherent framework for maintaining both record-level semantic compatibility and dataset-level predictive structure.

Critical weaknesses

1.A core limitation of CAST is that it relies on feedback generated by the predictor to guide the selection of candidate samples. Although the study used a lightweight predictor and introduced a reliability screening mechanism, the performance of the method under different predictor quality levels, calibration errors, or auxiliary attribute selections was not systematically evaluated. Due to the direct impact of these prediction signals on the sorting and selection of candidate samples, the generated synthetic data may be optimized for specific proxy predictors rather than always maintaining the true correlation between text and attributes. Therefore, the robustness and universality of this selection mechanism have not been fully validated.

2.CAST uses the frozen LLM and generates only a fixed set of candidates under several predefined narrative roles before executing the selection. Therefore, the selection process can only be made from candidates that already contain the required structured text relationships. If the initial candidate pool cannot contain sufficiently diverse or high-quality alternative solutions, subsequent CAS/RCS optimization cannot recover them. This makes the performance of the report potentially sensitive to the number of candidates, role design, reference retrieval strategy, and LLM backbone network, and experiments have not fully determined the behavior of CAST when the candidate pool is severely limited.

Questions for authors
  1. How are the primary target and auxiliary attributes determined in CAST? Are they manually specified for each dataset, selected according to a general data-driven criterion, or determined based on downstream tasks? How sensitive is the method to this choice?

  2. How robust is CAST to errors in the auxiliary task predictors?

  3. How does performance scale with the size and diversity of the candidate pool?

5 comments
AuthorsOpenReview ↗

Official reply to Reviewer 9y7M

We thank Reviewer 9y7M for the thoughtful questions on predictor feedback and the candidate pool. We address each concern below.

W1 & Q2. Robustness to predictor errors.

Ans: CAST is robust to predictor errors because it matches task features to real references rather than maximizing predictor scores. Under the same predictor, Eq. (6) selects the candidate whose task features are closest to those of a real reference. A text that is more predictable than real text is therefore penalized rather than rewarded. Target-margin maximization, which directly maximizes the predictor score, performs worse (Table 22), and utility is always measured by the downstream models rather than the feedback predictors.

To verify this, we fixed the candidate pool and perturbed the feedback predictors in three ways. We removed all auxiliary attributes, corrupted the labels of the auxiliary predictors, and trained all predictors on 10% of the data. Table R4 reports the share of unchanged selections, the utility of full CAST, and the change under each perturbation.

VariantUnchanged selections (Cloth)Unchanged selections (Wine)Utility (Cloth)Utility (Wine)
Full CAST100%100%65.6374.77
Auxiliary attributes: all removed75.7%70.9%−0.36−0.32
Auxiliary predictors: 30% label noise84.2%82.8%−0.37−0.48
All predictors: 10% of training data52.5%64.1%−0.84−0.46
Random selection (reference)~25%25.5%−1.17−1.59

Removing the auxiliary attributes or corrupting their labels costs less than a third of the drop from random selection. Training every predictor on 10% of the data changes many selections, yet utility remains above that of random selection. These results indicate that CAST does not overfit its proxy predictors.

A systematic calibration offset in a predictor's outputs affects a candidate and its reference alike and cancels in their difference, and the feature scale is re-estimated under the same predictor. The Domain Score term uses only within-row ranks and therefore needs no calibration (Section 3.3). We have added Table R4 to Appendix E and stated in Section 3 that CAS matches, rather than maximizes, predictor scores.

Q1. How are the primary target and auxiliary attributes chosen?

Ans: Both are fixed in advance by the dataset configuration rather than tuned on our results, and CAST does not depend on this choice. The primary target is the prediction label defined by the benchmark [1]. The auxiliary attributes are the structured fields configured for each dataset, and the condition field is selected automatically by Eq. (7).

An auxiliary attribute contributes only if its feedback passes gating, that is, if its validation performance exceeds a trivial baseline (Section 3.4). Division Name on Cloth, for example, fails gating and is excluded in this way. Removing all auxiliary attributes costs 0.36 / 0.32 points, less than a third of the drop from random selection (Table R4).

W2 & Q3. Candidate pool size and diversity.

Ans: Utility saturates at three to four candidates. CAST requires a diverse pool rather than a large one, and SSG provides this diversity by design. To measure the effect of pool size, we re-ran selection on Cloth for every role subset that contains Experience First, with the pool and protocol fixed. Table R5 reports the mean utility for each pool size, together with the existing checks on backbones and synthesis realizations.

SettingUtility (Cloth)
1 candidate (Experience First)63.50
2 candidates (mean over 3 role pairs)64.95
3 candidates (mean over 3 role triples)65.60
4 candidates (full CAST)65.63
5 generator backbones (Table 12)65.63–67.90
3 synthesis realizations (Table 23)65.49 ± 0.15

Three candidates already match the utility of four, and the results are stable across backbones and synthesis realizations. Diversity matters more than size: without the shared semantic plan and narrative roles, CAS performs no better than random selection (Table 25), because near-identical candidates offer little room for selection. This is why CAST constructs the pool and selects from it as a single method. We have added Table R5 to Appendix E and discussed this condition in Appendix F.

Reference

[1] Shi X, Mueller J, Erickson N, et al. Benchmarking Multimodal AutoML for Tabular Data with Text Fields. NeurIPS Datasets and Benchmarks Track (Round 2), 2021.

AuthorsOpenReview ↗

Follow-up on our response

Dear Reviewer 9y7M,

Thank you again for your thoughtful review. We understand that this is a busy period, and we truly appreciate your time.

As the discussion period will close soon, we would like to gently check whether our responses have addressed your concerns. In particular, Table R4 tests robustness to predictor errors, and Table R5 studies the size and diversity of the candidate pool.

If any concern remains, we would be glad to clarify it. If our responses have resolved your concerns, we would be grateful if you would kindly take them into account in your final assessment.

Best regards,

The Authors

AuthorsOpenReview ↗

Summary of revisions addressing your concerns

Dear Reviewer 9y7M,

We understand that time is limited near the end of the discussion period. For your convenience, we briefly summarize how the revision addresses each of your concerns.

Robustness to predictor errors (W1, Q2). CAS matches task features to those of real references rather than maximizing predictor scores, and target-margin maximization performs worse (Table 22). With the candidate pool fixed, removing all auxiliary attributes, adding 30% label noise to the auxiliary predictors, or training all predictors on 10% of the data lowers utility less than random selection does (Table R4). A systematic calibration offset affects a candidate and its reference alike and cancels in their difference.

Choice of target and auxiliary attributes (Q1). Both are fixed in advance by the dataset configuration, and the condition field is selected automatically by Eq. (7). Attributes with weak feedback are excluded by gating, and removing all auxiliary attributes costs less than a third of the drop from random selection.

Size and diversity of the candidate pool (W2, Q3). Utility rises from 63.50 with one candidate to 65.60 with three and 65.63 with four (Table R5). Results are also stable across five generator backbones and three synthesis realizations. Without the shared plan and narrative roles, CAS performs no better than random selection (Table 25), so pool diversity matters more than pool size.

We hope this summary makes the changes easy to check, and that the revisions help in your final assessment.

Best regards,

The Authors

Comments on the paper

5 comments
AuthorsOpenReview ↗

Official reply to Reviewer cXRA [1/2]

We thank Reviewer cXRA for the careful review and the constructive suggestions. We address each concern below.

W1 & Q1. Longer text and continuous targets.

Ans: Thank you for the question on scope. Our evaluation follows the standard benchmark of this field: the datasets, tasks, and notion of "long text columns" come from the text–tabular benchmark [1] and AFT-Tab [2]. There, long text columns are free-form text fields, as opposed to numerical, categorical, and short string fields. As Table R1 shows, the main text fields are substantially longer than the short text fields of the same tables.

DatasetMain text fieldMean (words)Median (words)Max (words)Median of short field (words)Full-condition coverage
Kickstarterdesc19.52035name: 599.88%
ClothReview Text60.359115Title: 399.86%
Wine Reviewsdescription40.540135province: 197.79%

For longer text fields, CAST needs no retraining. Its text generator is frozen and can be replaced by a longer-context model (Table 12 already covers five backbones), and selection scores each candidate as a whole. Document-length texts are rarely stored within a table row; generating them is a separate long-context problem [3, 4].

For continuous targets, to our knowledge, no established benchmark for long-text tabular synthesis yet combines a general continuous target with both discrete columns and long text columns. CAST handles such targets within its current design. The condition key is built from equal-sized quantile groups of the target, as CAST already does at initialization, so the groups cannot become sparse. In our data, even exact full-condition groups are well covered (last column of Table R1). We have clarified this scope in Section 4.1, added the short-field lengths to Table 9, and described the quantile key in Appendix B.3.

W2 & Q2. Contribution of each CAS component.

Ans: Each of the three components contributes. The second-round gain comes from re-estimating the feature scales and reliability weights, not from the refit features themselves. To isolate each component, we fixed the candidate pool, structured rows, and final mixing, and ran the three requested comparisons, changing one component at a time. Table R2 reports the utility of full CAST and the change for each variant.

VariantClothWine
Full CAST65.6374.77
Reference assignment: random within condition group−0.87−1.13
Synthetic-data predictors: frozen after round 1−0.79−0.78
Synthetic-data predictors: refit, round-1 scales and weights kept−0.86−0.83
Domain Score: removed−0.31−0.69
w/o CAS: random candidate per row (reference)−1.17−1.59

Among the three components, reference assignment matters most. Eq. (8) assigns synthetic records to real references that span different predictive-feature levels within each condition group, and random assignment loses this coverage. Removing the Domain Score, which measures resemblance to real text, also lowers utility, most clearly on Wine (−0.69).

For refitting, freezing the synthetic-data predictors also freezes the feature scales and reliability weights, so the second round reproduces the first. Refitting the predictors while keeping the round-1 scales and weights yields no gain either. The second-round gain therefore comes from re-estimating the scales and reliability weights on the refit predictors. We have added Table R2 to Appendix E.

References

[1] Shi X, Mueller J, Erickson N, et al. Benchmarking Multimodal AutoML for Tabular Data with Text Fields. NeurIPS Datasets and Benchmarks Track (Round 2), 2021.

[2] Zhang Y, Yan L, Duan S, et al. AFT-Tab: Adversarial Fine-Tuning for Tabular Data Synthesis with Long Text Columns. ACL 2026: 4581–4594.

[3] Bai Y, Zhang J, Lv X, et al. LongWriter: Unleashing 10,000+ Word Generation from Long Context LLMs. ICLR 2025.

[4] Liu H, Zaharia M, Abbeel P. Ring Attention with Blockwise Transformers for Near-Infinite Context. ICLR 2024.

AuthorsOpenReview ↗

Official reply to Reviewer cXRA [2/2]

W3 & Q3. Matched budgets, quoted results, and cost.

Ans: CAST's advantage does not come from a larger budget. With one candidate per row and no selection, CAST already outperforms every baseline on Cloth and Wine, including a CLLM-style generator with four candidates per row. We re-evaluated all quoted baselines under the same downstream protocol (Appendix C.2). For a budget-matched comparison, we also built a CLLM-style baseline that generates four candidates per row and picks one at random. Table R3 reports the candidate budget, GPU time, and utility.

MethodCandidates per rowGPU min (Wine)ClothWineKickstarter
CLLM18454.0143.0758.19
CLLM-style, 4 candidates, random selection410556.6644.36–
TabDDPM+LLM–7519.3823.3357.42
AFT-Tab–157 (training)61.8066.6366.22
CAST, 1 candidate, no selection18963.5071.62–
CAST410265.6374.7771.87

A single CAST candidate outperforms the four-candidate CLLM-style baseline by a large margin. CAST's advantage therefore comes from the shared semantic plan and narrative roles, not from the number of candidates. Selecting among four candidates brings a further improvement while still requiring less GPU time than the CLLM-style baseline.

All frozen-model baselines use the same Qwen2.5-7B backbone as CAST, and both CAST and the CLLM-style generator condition on real training records. Under matched candidate and GPU budgets, CAST retains its advantage, and its full pipeline requires less GPU time than AFT-Tab's training alone. We have added Table R3 to Appendix D.

AuthorsOpenReview ↗

Follow-up on our response

Dear Reviewer cXRA,

Thank you again for your thoughtful review. We understand that this is a busy period, and we truly appreciate your time.

As the discussion period will close soon, we would like to gently check whether our responses have addressed your concerns. In particular, Table R2 isolates each CAS component on Cloth and Wine Reviews, and Table R3 compares the methods under matched budgets.

If any concern remains, we would be glad to clarify it. If our responses have resolved your concerns, we would be grateful if you would kindly take them into account in your final assessment.

Best regards,

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.