{"id":"35cf998c-4a87-4260-999f-bd42be8ef902","arxiv_id":"2506.14206","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"CausalDiffTab adds a causal graph learned from the data as a regularizer to a mixed-type diffusion tabular generator and claims gains on seven datasets, but the evaluation contains a metric direction error.","lead":"A new diffusion model for tabular data, CausalDiffTab, adds a causal structure extracted from the training data as a regularizer to generate more realistic mixed numeric and categorical rows. The paper claims improvements on seven datasets, but its headline \"wins on every metric\" depends on one metric being scored in the wrong direction.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 6's C2ST metric is direction-inverted: higher detection accuracy means easier to distinguish, so the reported 'improvements' are degradations, breaking the 'outperforms across all metrics' claim.","rationale":"The reader's formal weakest_assumption focused on the causal matrix and circularity of the causal evaluation, but the reader's rationale also flagged the C2ST direction error. I agree that the C2ST issue is the most load-bearing because it directly falsifies the paper's headline claim of beating baselines across all metrics, independent of any assumptions about causal discovery. Even if the NOTEARS graph were perfect and the causal loss were correctly implemented, Table 6 as presented would still invert a standard metric and count a degradation as an improvement. The recommended verdict is unchanged: REJECT. A separate, secondary concern is that Eq. 14 minimizes the masked sum of pairwise noise products; if the causal matrix marks allowed edges, minimizing those products on allowed edges suppresses rather than retains aligned correlations, contradicting the textual claim. This would further undermine the mechanism, but the C2ST sign error is sufficient to break the central claim and is the cleanest to test. I chose 'partial' rather than 'agree' because the reader's weakest_assumption is not the same as mine, though the reader did identify the C2ST issue elsewhere in the verdict.","tokens_in":18139,"tokens_out":4328,"duration_ms":46070,"concrete_test":"Run a standard C2ST evaluation using the released code on Adult and Default: train a logistic regression classifier on the real training set versus the generated set (same size, same preprocessing), and report classification accuracy or ROC-AUC. Compare the sign of Ours minus TabDiff. If the reported Table 6 values are accuracy (or AUC where 0.5 is chance), then Ours is worse on most datasets and the central claim fails. As a sanity check, TabDiff's near-1.0 values already imply near-perfect detectability, which contradicts the paper's own low shape/trend errors; if the authors instead intended 1 minus accuracy, Table 6 entries should be near 0.02, not 0.98.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Table 6 the paper reports C2ST (detection score) and states \"Higher scores reflect superior performance.\" Under the standard Classifier Two-Sample Test, the score is the accuracy of a classifier trained to distinguish real from synthetic; higher accuracy means the synthetic data is more distinguishable from real, hence worse fidelity. The paper's own text says the metric \"reflects the difficulty of distinguishing synthetic data from real data,\" but then treats an increase from 0.9832 (TabDiff) to 0.9963 (Ours) on Adult as a 1.31% improvement. If these numbers are classifier accuracies, CausalDiffTab is easier to detect on 5 of 7 datasets (Adult, Default, Magic, Beijing, Diabetes) than TabDiff, and its \"across all metrics\" claim is false on this metric. This is not a subtle modeling assumption; it is a sign error in a headline evaluation table. A correction to the direction of this metric removes one of the seven claimed wins and directly contradicts the abstract's universal claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CausalDiffTab, a continuous-time diffusion model for mixed numerical/categorical tabular data. A causal graph is learned from the training data with nonlinear NOTEARS, thresholded into a binary mask, and used in a causal-consistency loss that penalizes masked outer products of predicted noise. The regularization weight is adapted during training based on loss fluctuation and noise level. The authors evaluate on seven datasets against GAN, VAE, autoregressive, and diffusion baselines using shape/trend errors, alpha-precision, beta-recall, C2ST, downstream AUC/RMSE, and DCR, and claim state-of-the-art performance across all metrics. The code is publicly released.","tokens_in":18338,"tokens_out":10176,"duration_ms":106426,"significance":"The problem of tabular data generation with causal awareness is relevant, and the adaptive-weighting idea is a reasonable response to training instability. The authors provide a public implementation and extensive benchmarking, which is commendable. However, the current manuscript contains a sign error in the C2ST evaluation and an inconsistency between the causal loss formula and its stated purpose; these issues affect the central claim of universal improvement. With corrections and re-evaluation, the method could be a useful contribution, but the evidence as presented does not establish the headline result.","major_comments":[{"comment":"The C2ST metric is interpreted backwards. C2ST is the accuracy of a classifier trained to distinguish real from synthetic samples; accuracy near 1 means the synthetic data is perfectly detectable and thus indicates worse fidelity, not better. The caption 'Higher scores reflect superior performance' is therefore incorrect. For Adult, CausalDiffTab's 0.9963 is worse than TabDiff's 0.9832, not a 1.31% improvement, and this pattern occurs on most datasets. Because the abstract and conclusion claim wins 'across all metrics', this is a load-bearing error.","section":"Section 4.4 / Table 6"},{"comment":"The causal loss as written minimizes the mean of predicted-noise outer products over the causal mask S. If S is the set of allowed causal edges, this suppresses interactions along those edges, which contradicts the text's claim of 'retaining correlations aligned with the allowed causal directions'. If S is meant to be the complement (non-causal pairs), then Eq. (14), Algorithm 1, and the surrounding text must be corrected to say so. As written, the loss does not implement the described mechanism.","section":"Section 3.3, Eq. (14), Algorithm 1"},{"comment":"The causal graph is estimated from the same training data with NOTEARS and then used as both the regularization target and the basis for judging causal plausibility. The paper provides no independent validation of the learned graph (for example, against known edges in the Adult dataset), so the reported causal-consistency gains may be self-confirming. Figure 2 uses the external Husband/Wife constraints, which is a partial check, but this does not validate the full binary thresholded graph or rule out spurious edges.","section":"Section 3.2 and Section 4.4"}],"minor_comments":[{"comment":"The dataset list contains 'Fault' in addition to the seven datasets shown in Table 1; verify the intended dataset names.","section":"Section 4.2"},{"comment":"The captions say 'Higher error rates reflect higher performance', but these are scores, not error rates; the captions should say 'Higher scores'.","section":"Tables 4 and 5"},{"comment":"The text reports a 14.3% average shape improvement across six datasets, but Table 2 shows seven datasets and a 13.7% improvement; the 14.3% figure corresponds to the trend metric in Table 3.","section":"Section 4.4"},{"comment":"The sentence introducing Eq. (13) is incomplete ('increasing the discretization resolution can help approximate a tighter ELBO.'); complete the sentence and integrate it with the equation.","section":"Equation (13)"},{"comment":"The phrase 'We compare MoR with nine popular...' should read 'We compare our method with nine popular...'.","section":"Section 4.3"},{"comment":"Algorithm 1 returns L_base multiplied by an adaptive weight w, while Eq. (14) and Eq. (15) use a fixed lambda; clarify the relationship between lambda and w_hybrid.","section":"Algorithm 1 and Eq. (15)"}],"recommendation":"major_revision","confidential_remarks":"The C2ST direction error and the discrepancy between Eq. (14) and its stated purpose are serious enough that I would not accept the paper in its current form. I would want to see a revised manuscript with corrected metric interpretation, a clarified or corrected causal loss, and an independent check of the learned causal graph before judging the contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick read. CausalDiffTab essentially wraps TabDiff with a NOTEARS-derived binary causal mask and an adaptive regularization weight; the new piece is that the mask is applied to predicted-noise outer products inside a masked categorical diffusion loss. That combination is not in the earlier work, and the paper is honest about building on TabDiff. It also ships code, runs seven datasets, and reports shape/trend/MLE/privacy numbers with error bars on most tables. The shape and trend improvements over TabDiff are small but consistent, and the downstream MLE table shows a modest edge. So the empirical core is real, not fabricated.\n\nSoft spots are load-bearing. Table 6's C2ST direction is wrong. They report logistic-regression detection accuracy and call higher \"superior,\" but under the standard Classifier Two-Sample Test higher accuracy means easier to distinguish, i.e. worse fidelity. Corrected, CausalDiffTab loses to TabDiff on 5 of 7 datasets on this metric, which breaks the abstract's \"outperforms across all metrics\" claim. This is a sign error in a headline table, not a nuance.\n\nSecond, Eq. 14 defines the causal loss as lambda * mean_{(i,j) in S} (eps_i * eps_j) with no absolute value or positive-semidefinite structure. Minimizing that product does not \"retain correlations aligned with allowed causal directions\"; it pushes predicted noise components toward zero or opposite signs. The prose describes a mechanism the equation does not implement, or at least one the authors never specify (masked absolute values? squared deviations from causal signs?). That is a real mechanism gap, not an aesthetic complaint.\n\nThird, circularity: the causal graph is estimated from the same training data with NOTEARS, thresholded at 0.3, then used as both training target and implicit evaluation lens. Figure 2 uses semantic sanity checks rather than the same graph, which is good, but the core \"causal\" claim is only as good as the estimated graph, and no ground-truth or external causal validation is offered. Ablations cover only two datasets, and Table 6 has no error bars.\n\nWho is this for? People working on synthetic tabular data and on evaluation pitfalls in generative models. It is not a field-reorganizing paper; it is a plausible incremental heuristic with a buggy headline metric. The code and experimental scaffold are worth a serious referee, with the clear expectation that the C2ST direction be fixed, the causal loss be rewritten or explained, and the circularity acknowledged. I would not desk-reject; I would send it out, but the authors should expect major revision before acceptance.","headline":"A reasonable incremental heuristic for tabular diffusion, but the C2ST metric is direction-inverted and Eq. 14 doesn't do what the text claims; should go to review with major-revision expectations, not be accepted as is.","tokens_in":18905,"tokens_out":3565,"would_cite":false,"duration_ms":41817,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"CausalDiffTab claims its diffusion model, regularized by a causal graph learned from the real table, outperforms every baseline on all fidelity, privacy, and downstream metrics — roughly 14% lower error than the previous best.","keywords":["tabular data generation","diffusion models","causal discovery","directed acyclic graph","causal regularization","mixed-type data","synthetic data","privacy-preserving generation"],"falsifier":"Take a synthetic table generated from a known directed acyclic graph with known nonlinear mechanisms and run the full pipeline: check whether NOTEARS recovers the true edges and directions, whether swapping in the true graph changes the reported shape, trend, and privacy numbers, and whether a randomly permuted causal mask still delivers the same gains. If a permuted mask keeps the gains, the causal directions are not the driver and the adaptive weighting schedule alone explains the improvement.","tokens_in":17863,"feed_emoji":"📊","tokens_out":23714,"duration_ms":205517,"temperature":0.7,"pith_summary":"CausalDiffTab tries to settle a practical problem: synthetic tabular data that is faithful enough to train on and safe enough to share. The paper claims that a diffusion model for mixed numerical-categorical data improves on every measured axis if it is first told the causal structure of the real table and penalized whenever its predicted noise contradicts that structure. On seven real-world datasets and seven metrics — marginal fidelity, column-correlation fidelity, distinguishability, precision, recall, privacy, and downstream task performance — it reports the best average scores, roughly 14 percent better than the prior state of the art on headline fidelity. A sympathetic reader would care because real tabular data is everywhere but hard to share; if the claim holds, privacy-preserving synthetic copies become closer to drop-in substitutes for the original.","feed_headline":"Causal-aware diffusion cuts tabular-data error by 14 percent","feed_subtitle":"A diffusion model that learns the table's causal links beats predecessors on fidelity, privacy, and downstream tasks.","key_machinery":"The central object is the causal matrix $G \\in \\{0,1\\}^{d \\times d}$: a binary directed acyclic graph, one entry per ordered pair of feature columns, learned before training by nonlinear NOTEARS (an MLP-based regression with a DAG constraint, Eqs. 4–5) on the one-hot-encoded training table and binarized with threshold $\\tau=0.3$ (Eq. 6). It carries the argument because it defines which pairwise predictions the generative model is allowed to align. The matching step is causal pair matching: the denoising network's predicted noise for each column is formed into the outer-product matrix $\\hat{\\epsilon}_i \\hat{\\epsilon}_j$ (numerical predictions use the normalized residual $(x_t - \\hat{\\mu})/\\sigma$; categorical predictions use softmax probabilities), the matrix is masked so only entries $(i,j)$ inside $G$ survive, and the mean of the masked entries is minimized as the causal loss (Eq. 14). The second mechanism is the hybrid adaptive weight (Eq. 16), $w_{\\text{hybrid}} = \\frac{w_{\\max}}{2}(e^{-|\\Delta L|} + \\frac{1}{1+\\sigma_{\\text{mean}}})$, which couples the penalty's strength to training stability (loss fluctuation $\\Delta L$) and to the noise level $\\sigma_{\\text{mean}}$, implementing the hierarchical prior fusion principle of learning low-level features first and injecting high-level causal semantics later.","core_discovery":"The paper's central claim, stated in the abstract and conclusion, is that CausalDiffTab \"outperforms baseline methods across all metrics.\" Concretely, it reports the lowest average shape error (1.17% vs. 1.33% for the previous state of the art, TabDiff, a 13.7% improvement) and trend error (2.10% vs. 2.40%, a 14.3% improvement), the best average $\\alpha$-precision (98.24) and $\\beta$-recall (50.32), the best average detection score (0.9748), the best privacy score (DCR 53.83%, closer to the ideal 50% than TabDiff's 54.11%), and downstream task results that approach real-data training (Adult AUC 0.914 vs. real 0.927; Diabetes AUC 0.692 vs. real 0.704). The mechanism is a causal matrix extracted from the raw training table by nonlinear NOTEARS and binarized at threshold $\\tau=0.3$ (Eq. 6). During denoising, the outer product of the model's per-feature predictions — numerical: the normalized residual $(x_t - \\hat{\\mu})/\\sigma$; categorical: softmax probabilities — is masked by this matrix and minimized as a causal loss (Eq. 14), suppressing predictions that violate the learned directions. The loss is scaled by a hybrid adaptive weight $w_{\\text{hybrid}} = \\frac{w_{\\max}}{2}(e^{-|\\Delta L|} + \\frac{1}{1+\\sigma_{\\text{mean}}})$ (Eq. 16) that weakens the causal penalty during high-noise, unstable training and strengthens it as the model stabilizes, an application of the hierarchical prior fusion idea; the ablation shows that a fixed weight degrades shape and trend, and that nonlinear causal extraction outperforms linear extraction.","pith_inferences":["A permuted-mask control experiment would isolate what causes the gains: if a randomly shuffled causal matrix preserves the reported improvements, the improvement comes from the adaptive weighting schedule itself, not from causal direction information.","The causal checks are measured against the same learned matrix the model was trained to satisfy; auditing the generated data against an independent ground-truth graph (for example, synthetic data from a known DAG) would be a stricter test the paper does not run.","The NOTEARS causal-discovery routine is load-bearing but no NOTEARS citation appears in the reference list, and the hierarchical prior fusion principle is invoked without a citation, so the causal component cannot be verified from the paper's bibliography alone.","If the learned causal graph is trustworthy, the same masked-diffusion mechanism extends naturally to counterfactual generation and missing-value imputation in mixed-type tables, since the causal matrix supplies a conditional sampling target; the paper cites adjacent work in that direction but implements neither extension."],"forward_implications":["Synthetic data from this model can substitute for real data in downstream training: reported MLE scores land close to real-data performance (Adult AUC 0.914 vs. real 0.927; Diabetes AUC 0.692 vs. real 0.704).","Privacy protection improves: DCR scores are closer to the ideal 50% than the previous state of the art on four of five datasets, meaning generated records resemble the training set less closely.","Causally implausible records drop sharply: in the Adult dataset, Husband-Female combinations fall from 398 (TabDiff) to 244 and Wife-Male from 271 to 147.","Fixed-weight causal regularization hurts fidelity, so any future method that injects causal constraints into generative models should follow the adaptive weighting scheme rather than a constant penalty.","The ablation's nonlinear-versus-linear comparison supports the paper's premise that tabular causal relations are too complex for a linear regularizer."],"supporting_citations":[{"why":"Supplies the strongest baseline (TabDiff), the evaluation methodology, and the hyperparameters adopted in Section 4.1; the comparison the paper must beat.","marker":"Shi et al. (2025)"},{"why":"The nonlinear causal-discovery routine that produces the causal matrix in Section 3.2; described in text but absent from the reference list.","marker":"notears framework"},{"why":"Foundational denoising-diffusion formulation underlying the numerical forward and reverse processes.","marker":"Ho et al. (2020)"},{"why":"Continuous-time stochastic differential equation framework used for numerical features (Eqs. 7–9).","marker":"Song et al. (2021b)"},{"why":"Variational-diffusion continuous-time ELBO that justifies the categorical masked-diffusion loss (Eq. 13).","marker":"Kingma et al. (2021)"},{"why":"TabSyn, the latent-space score diffusion baseline whose approach CausalDiffTab contrasts with its in-data-space causal modeling.","marker":"Zhang et al. (2024)"},{"why":"TabDDPM, the prior mixed-type diffusion baseline that must be beaten.","marker":"Kotelnikov et al. (2023)"},{"why":"CTGAN and TVAE baselines whose distributional and stability limits motivate the causal diffusion approach.","marker":"Xu et al. (2019)"}],"fun_headline_variants":["CausalDiffTab: causal-aware diffusion for mixed-type tables","Causal diffusion model beats baselines on mixed tabular data","Causal-aware diffusion cuts tabular data error by 14%","Diffusion with causal priors improves tabular generation","CausalDiffTab: learning table causality for better synthesis"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the causal graph learned automatically from the training table — nonlinear NOTEARS binarized at threshold $\\tau=0.3$ — matches the true causal directions, because the paper's causal checks compare generated data against that same learned graph; a wrong graph would hard-wire spurious correlations as causal and the evaluation would certify them.","fun_headline_variants_meta":{"raw":{"variants":["CausalDiffTab: causal-aware diffusion for mixed-type tables","Causal diffusion model beats baselines on mixed tabular data","Causal-aware diffusion cuts tabular data error by 14%","Diffusion with causal priors improves tabular generation","CausalDiffTab: learning table causality for better synthesis"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000298,"raw_usage":{"total_tokens":1825,"prompt_tokens":1148,"completion_tokens":677,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":764,"completion_tokens_details":{"reasoning_tokens":593}},"tokens_in":764,"tokens_out":677,"duration_ms":7188,"temperature":1.0,"reasoning_tokens":593,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T00:18:06.305945+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a synthetic table generated from a known directed acyclic graph with known nonlinear mechanisms and run the full pipeline: check whether NOTEARS recovers the true edges and directions, whether swapping in the true graph changes the reported shape, trend, and privacy numbers, and whether a randomly permuted causal mask still delivers the same gains. If a permuted mask keeps the gains, the causal directions are not the driver and the adaptive weighting schedule alone explains the improvement.","supporting_citations":[],"review_version":1}