超越缝合假设：通过语义量化的多模态合成数据评估统一框架

# Beyond the Stitching Assumption: A Unified Framework for Multimodal Synthetic Data Evaluation via Semantic Quantization
Source: [https://arxiv.org/html/2609.22149](https://arxiv.org/html/2609.22149)
Yefeng YuanZhan ShiAffiliation:Santa Clara UniversityAffiliation:Santa Clara, USAEmail:[ashi2@scu\.edu](mailto:)Liang ChengAffiliation:eBay Inc\.Affiliation:San Jose, USAEmail:[liacheng@ebay\.com](mailto:)Yuhong LiuAffiliation:Santa Clara UniversityAffiliation:Santa Clara, USAEmail:[yhliu@scu\.edu](mailto:)

###### Abstract

Multimodal synthetic datasets combine structured attributes with free text, but are often evaluated separately\. Such metrics can remain high after tabular–text pairings are disrupted\. We present a projection\-based evaluator for tabular–text synthetic data\. A fixed sentence encoder maps text to embeddings,kk\-means converts them to cluster states, and tabular variables are represented as categorical or quantile\-binned states\. Real and synthetic contingency tables are compared using Jensen–Shannon divergence \(JSD\), normalized mutual information \(NMI\), conditional JSD \(cJSD\), and joint\-state entropy\. We also report text\-to\-attribute \(T2A\) utility and a holdout\-calibrated proximity flag rate \(PFR\) as a representation\-level diagnostic\. A text\-permutation control preserves both marginal distributions while disrupting their pairing\. Experiments on Amazon Reviews, Kiva Loans, and the Employment Scam Aegean Dataset show that modality\-specific scores remain high under this control\. The projection diagnostics detect disruption when real projected dependence exceeds a permutation baseline, but are less informative for weak or sparse projections\. Some conditioned LLM baselines also exhibit stronger measured dependence than the corresponding real\-data projections\. These results support explicit cross\-modal evaluation with permutation baselines and coverage reporting\.

###### Index Terms:

Multimodal synthetic data, synthetic data evaluation, tabular–text alignment, semantic quantization, record\-proximity diagnostics

## IIntroduction

Recent advances in generative models have made synthetic data a practical tool for data sharing, augmentation, and benchmarking in privacy\-sensitive domains\[[1](https://arxiv.org/html/2609.22149#bib.bib12)\]\. Many modern applications, especially in healthcare, finance, and enterprise analytics, involve multimodal records that pair structured attributes with unstructured text\. Examples include physiological measurements paired with clinical notes\[[2](https://arxiv.org/html/2609.22149#bib.bib34),[3](https://arxiv.org/html/2609.22149#bib.bib13)\], loan metadata paired with borrower narratives\[[4](https://arxiv.org/html/2609.22149#bib.bib35)\], job\-posting metadata paired with job descriptions\[[5](https://arxiv.org/html/2609.22149#bib.bib36)\], and product metadata paired with review text\[[6](https://arxiv.org/html/2609.22149#bib.bib37)\]\. In these settings, synthetic data is useful only if it can be evaluated for both semantic consistency and privacy risk\.

As generative architectures ranging from Generative Adversarial Networks \(GANs\) and diffusion models to Large Language Models \(LLMs\) are applied to multimodal records, evaluation methods have not kept pace with the structure of the generated data\. Existing metrics are often modality\-specific\. Tabular evaluation commonly relies on marginal distribution tests, such as the Kolmogorov–Smirnov test and total variation \(TV\) distance, together with low\-order dependency statistics over structured features, as implemented in widely used synthetic\-data evaluation tools such as SDMetrics and Synthcity\[[7](https://arxiv.org/html/2609.22149#bib.bib23),[8](https://arxiv.org/html/2609.22149#bib.bib28)\]\. Text evaluation, by contrast, uses reference\-similarity metrics such as BERTScore\[[9](https://arxiv.org/html/2609.22149#bib.bib26)\]and distributional metrics such as MAUVE\[[10](https://arxiv.org/html/2609.22149#bib.bib27)\]\. Because these metrics evaluate text without conditioning on the tabular attributes in the same record, they do not by themselves test row\-level cross\-modal consistency\. As a result, an evaluator may validate each modality separately while missing broken row\-level alignment, sensitive cross\-modal associations, or synthetic records that remain too close to sensitive originals\.

This separation leads to what we call the “stitching assumption”: the belief that strong tabular\-only and text\-only scores imply a coherent multimodal record\. For example, a synthetic loan record may pair a sector of “Agriculture” with a fluent loan purpose stating “to purchase bales of clothes for resale” \(as illustrated in Figure[1](https://arxiv.org/html/2609.22149#S1.F1)\)\. The sector distribution may look realistic, and the loan purpose may be fluent, but the paired record is incoherent\. In privacy\-sensitive settings, separate modality\-level checks can also miss full\-record proximity to training examples or repeated tabular–text pairings; more broadly, they can miss spurious cross\-modal associations that affect downstream validity\.

A natural evaluation target is the cross\-modal joint structure between tabular states and text semantics\. One nonparametric way to compare this structure is to quantize the variables and construct contingency tables over their joint states\. However, a full table over all tabular attributes and text\-embedding clusters grows with the Cartesian product of the state spaces and quickly becomes sparse\. Richer probabilistic models can represent more complex joint structure, but they are often too costly, assumption\-dependent, or difficult to use as routine evaluation tools\[[11](https://arxiv.org/html/2609.22149#bib.bib14),[12](https://arxiv.org/html/2609.22149#bib.bib15),[13](https://arxiv.org/html/2609.22149#bib.bib16)\]\. Our goal is therefore not to recover the full continuous multimodal joint distribution\. Instead, we evaluate selected low\-dimensional quantized cross\-modal projections\. Here, a projection denotes a selected view, such as one tabular attribute or a small tabular subset crossed with a text semantic\-state variable\. These projections are designed to be statistically estimable, interpretable, and useful for diagnosing cross\-modal dependence\.

To address this dimensionality bottleneck, this work extends SynEval111Code is available at[https://github\.com/privacy\-enhancing\-technologies/SynEval](https://github.com/privacy-enhancing-technologies/SynEval)\., our earlier framework for multifaceted synthetic\-data evaluation\[[14](https://arxiv.org/html/2609.22149#bib.bib39)\], to multimodal tabular–text synthetic\-data evaluation, as shown in Figure[2](https://arxiv.org/html/2609.22149#S1.F2)\. The original SynEval framework focused on decoupled quality, utility, diversity, and privacy axes for LLM\-generated synthetic data\. In contrast, the present work adds a semantic\-quantization layer for evaluating tabular–text association: text sequences are mapped to dense embeddings and grouped into discrete semantic clusters, while tabular features are discretized into categorical or binned states\. The resulting contingency tables allow direct empirical comparison within selected cross\-modal projections\. This design does not eliminate the curse of dimensionality; instead, it turns joint evaluation into an explicit projection\-selection problem\. The scalable default uses pairwise tabular–text projections, while the same framework supports budgeted higher\-order tabular projections crossed with a text semantic\-state variable when cell budget and support conditions are satisfied\.

Our contributions are threefold:

1. 1\.A controlled test of modality\-isolated evaluation\.We formalize the “stitching assumption” as an evaluation failure mode in which favorable tabular\-only and text\-only scores are treated as evidence of a coherent multimodal dataset\. We introduce a text\-permutation control that preserves the marginal distribution of each modality while disrupting their pairing\.
2. 2\.A projection\-based extension of SynEval for tabular–text data\.We map text into fixed embedding\-cluster states and tabular variables into categorical or quantile\-binned states, then compare selected tabular–text contingency tables\. The method evaluates tractable low\-dimensional projections rather than attempting to estimate the full continuous multimodal joint distribution\.
3. 3\.A multi\-axis evaluation with explicit boundary conditions\.We evaluate projected fidelity, measured dependence, predictive utility, joint\-state entropy, and representation\-level record proximity on three datasets\. The experiments also identify conditions under which the diagnostics become less informative, including weak real\-data dependence, sparse projected support, and higher\-order interactions that are not visible in pairwise projections\.

![Refer to caption](https://arxiv.org/html/2609.22149v1/stitching_assumption.png)Fig\. 1:The “stitching assumption” in multimodal synthetic data evaluation\. Separate tabular and text metrics can assign high scores even when the two components of the same record contradict each other\.![Refer to caption](https://arxiv.org/html/2609.22149v1/framework.png)Fig\. 2:Architectural overview of the proposed framework\. Semantic quantization maps text embeddings and selected tabular features into a joint discrete representation, enabling coarse\-grained joint evaluation of multimodal synthetic data across fidelity, dependence, utility, diversity, and record proximity\.
## IIRelated Work

Synthetic\-data research is increasingly motivated by data\-access constraints, model development needs, and privacy regulations\[[15](https://arxiv.org/html/2609.22149#bib.bib17),[16](https://arxiv.org/html/2609.22149#bib.bib18)\]\. For multimodal records, however, the literature remains fragmented across generation methods, modality\-specific evaluation metrics, and separate privacy or diversity diagnostics\. We therefore review three strands of related work: generative architectures for tabular and textual modalities, limitations of modality\-isolated evaluation, and the dimensionality, diversity, and proximity challenges that arise in joint multimodal evaluation\. Cross\-modal dependence is well established in multimodal learning; our contribution concerns the narrower problem of evaluating whether synthetic tabular–text datasets preserve it under controlled negative tests and selected quantized projections\[[17](https://arxiv.org/html/2609.22149#bib.bib21)\]\.

### II\-AGenerative Architectures for Tabular and Textual Modalities

Structured synthetic data has been generated using statistical models such as Bayesian networks and copulas\[[18](https://arxiv.org/html/2609.22149#bib.bib1),[19](https://arxiv.org/html/2609.22149#bib.bib2)\], as well as neural generators such as Conditional Tabular Generative Adversarial Network \(CTGAN\), Tabular Variational Autoencoder \(TVAE\), and diffusion models\[[20](https://arxiv.org/html/2609.22149#bib.bib3),[21](https://arxiv.org/html/2609.22149#bib.bib6),[22](https://arxiv.org/html/2609.22149#bib.bib7),[23](https://arxiv.org/html/2609.22149#bib.bib8)\]\. These methods improve flexibility but remain sensitive to heterogeneous feature types, preprocessing, and data imbalance\. Text synthesis has been advanced by LLMs\[[24](https://arxiv.org/html/2609.22149#bib.bib9),[25](https://arxiv.org/html/2609.22149#bib.bib4)\], including prompt\-based and fine\-tuned approaches for tabular data\[[26](https://arxiv.org/html/2609.22149#bib.bib10),[27](https://arxiv.org/html/2609.22149#bib.bib5),[28](https://arxiv.org/html/2609.22149#bib.bib11)\]\. However, generating tabular attributes and text as a coherent record remains difficult: some methods flatten tables into text or use late fusion rather than explicitly preserving row\-level cross\-modal dependence\[[29](https://arxiv.org/html/2609.22149#bib.bib19),[30](https://arxiv.org/html/2609.22149#bib.bib20),[17](https://arxiv.org/html/2609.22149#bib.bib21),[31](https://arxiv.org/html/2609.22149#bib.bib22)\]\. This motivates evaluation metrics that measure tabular–text alignment rather than unimodal realism\.

### II\-BLimitations of Modality\-Isolated Quality Metrics

Commonly used synthetic\-data evaluation toolkits mostly focus on unimodal data types, rather than multimodal combinations\. SDMetrics\[[7](https://arxiv.org/html/2609.22149#bib.bib23)\]evaluates univariate column shapes and selected pairwise trends among structured variables, while Synthcity\[[8](https://arxiv.org/html/2609.22149#bib.bib28)\]provides broader fidelity, utility, and privacy evaluations for several tabular\-data settings\. These toolkits do not natively assess whether free\-text fields are consistent with the structured attributes in the same record\.

In natural\-language generation, the Bilingual Evaluation Understudy \(BLEU\)\[[32](https://arxiv.org/html/2609.22149#bib.bib24)\]and Recall\-Oriented Understudy for Gisting Evaluation \(ROUGE\)\[[33](https://arxiv.org/html/2609.22149#bib.bib25)\]measure reference overlap, BERTScore\[[9](https://arxiv.org/html/2609.22149#bib.bib26)\]measures contextual candidate–reference similarity, and MAUVE\[[10](https://arxiv.org/html/2609.22149#bib.bib27)\]compares generated and human\-written text distributions\. MAUVE also quantizes language\-model representations, whereas our method crosses fixed text\-embedding clusters with tabular states to evaluate cross\-modal joint structure\.

Our earlier SynEval framework provided a multifaceted evaluation for LLM\-generated synthetic data, including quality, utility, diversity, and privacy diagnostics\[[14](https://arxiv.org/html/2609.22149#bib.bib39)\]\. That framework was designed primarily around modality\-level and task\-level evaluation axes, and it did not explicitly test whether structured attributes and unstructured text remain mutually consistent within the same row\. The present work extends SynEval by adding semantic quantization, selected cross\-modal contingency tables, the text\-permutation control, and projection\-level diagnostics for tabular–text alignment\.

### II\-CJoint Evaluation Challenges: Scale, Diversity, and Proximity

One common component of synthetic\-data risk evaluation is detecting synthetic records that are unusually close to training records, since such records may indicate memorization or increased disclosure risk\. The Distance to Closest Record \(DCR\) is often used as an empirical nearest\-neighbor proximity diagnostic for this purpose\[[34](https://arxiv.org/html/2609.22149#bib.bib29)\]\. For tabular–text records, however, DCR is not meaningful without an explicit representation and scaling convention\. Naively concatenating encoded tabular features with high\-dimensional text embeddings can make Euclidean distances dominated by one modality rather than by meaningful full\-record similarity\[[35](https://arxiv.org/html/2609.22149#bib.bib30)\]\. This motivates calibrated proximity diagnostics that specify the multimodal representation and interpret metrics like DCR as empirical record\-proximity screens rather than formal privacy guarantees\.

Diversity evaluation raises a different but related joint\-structure problem\. Generative diversity is often assessed with isolated measures such as tabular support coverage, marginal entropy, or textual vocabulary richness\. These measures can miss an analogue of mode collapse in cross\-modal pairings: a generator may produce diverse tabular values and diverse text, but combine them through a narrow set of repetitive tabular–text associations\. Joint entropy provides a standard information\-theoretic measure of spread over joint states\[[36](https://arxiv.org/html/2609.22149#bib.bib33)\]and has also been used in subset\-selection settings such as sensor placement\[[37](https://arxiv.org/html/2609.22149#bib.bib31)\]\. We therefore use joint entropy as a projection\-level diagnostic of diversity over quantized cross\-modal states, while interpreting high entropy together with dependence and fidelity metrics rather than as evidence of semantic validity by itself\.

These limitations motivate a projection\-based evaluator that avoids estimating the full joint distribution\. Evaluations should look beyond isolated modality scores, but they must avoid full joint modeling procedures that are too sparse or costly for routine use\. This motivates our semantic\-quantization approach, which compares selected quantized cross\-modal projections as tractable, interpretable diagnostics of multimodal synthetic records\.

## IIIMethodology

The proposed framework evaluates multimodal synthetic data through selected quantized cross\-modal projections and fixed\-representation diagnostic probes\. Instead of estimating the full continuous multimodal joint distribution, it maps text into discrete embedding\-cluster states and tabular variables into categorical or quantile\-binned states for projection\-level fidelity, dependence, and diversity diagnostics, while also reporting cross\-modal utility and calibrated record\-proximity diagnostics\. All data\-dependent evaluation artifacts—including text\-cluster quantizers, tabular discretizers, categorical vocabularies, normalization statistics, and the record\-proximity scaling parameterλ⋆\\lambda^\{\\star\}—are fitted on the real training partition and applied unchanged to holdout and synthetic records\. The Sentence\-BERT \(SBERT\) encoder itself is pretrained and kept fixed\.

Throughout the paper,DdemoD\_\{\\mathrm\{demo\}\}denotes the fixed 10\-example prompt\-demonstration set,RtrainR\_\{\\mathrm\{train\}\}andRholdoutR\_\{\\mathrm\{holdout\}\}denote the disjoint real training and holdout partitions,R=Rtrain∪RholdoutR=R\_\{\\mathrm\{train\}\}\\cup R\_\{\\mathrm\{holdout\}\}denotes the real evaluation pool, andSSdenotes a synthetic dataset\. No demonstration record appears in eitherRtrainR\_\{\\mathrm\{train\}\}orRholdoutR\_\{\\mathrm\{holdout\}\}\. Unless stated otherwise, projection\-level real reference distributions are computed onRRusing quantizers fitted onRtrainR\_\{\\mathrm\{train\}\}, whereas predictive\-probe evaluation and holdout\-based proximity\-threshold calibration useRholdoutR\_\{\\mathrm\{holdout\}\}\.

### III\-ASemantic Quantization and Projected State Spaces

Let𝒟=\{\(𝐱\(i\),𝐭\(i\)\)\}i=1N\\mathcal\{D\}=\\\{\(\\mathbf\{x\}^\{\(i\)\},\\mathbf\{t\}^\{\(i\)\}\)\\\}\_\{i=1\}^\{N\}denote a generic multimodal dataset, where𝐱\(i\)\\mathbf\{x\}^\{\(i\)\}contains structured tabular attributes and𝐭\(i\)\\mathbf\{t\}^\{\(i\)\}denotes the predefined text field used in the evaluated projection\. Directly estimating the full Cartesian product over all tabular and text states is infeasible: if tabular featurejjhas\|Vj\|\|V\_\{j\}\|states and text fieldllhasKlK\_\{l\}semantic clusters, the full joint state space scales as:

\|𝒞full\|=\(∏j=1m\|Vj\|\)​\(∏l=1pKl\)\.\|\\mathcal\{C\}\_\{\\mathrm\{full\}\}\|=\\left\(\\prod\_\{j=1\}^\{m\}\|V\_\{j\}\|\\right\)\\left\(\\prod\_\{l=1\}^\{p\}K\_\{l\}\\right\)\.\(1\)
Our framework therefore evaluates selected projections rather than the full multimodal joint distribution\. The scalable default is pairwise: one tabular\-state variableCXjC\_\{X\_\{j\}\}is evaluated against one text semantic\-state variableCTC\_\{T\}\. This setting is used for all real\-data projections in the main experiments because it is more tractable and interpretable than higher\-order projections\. Empirical support is assessed separately for each dataset and semantic resolutionKK, and sparse projections are explicitly flagged\.

#### Budgeted multivariate projections\.

The pairwise setting is scalable and interpretable, but it can miss dependencies that only appear when several tabular variables are considered jointly\. The framework therefore supports budgeted multivariate projections over a user\-specified tabular subsetAA, yielding a joint tabular stateCXAC\_\{X\_\{A\}\}\. The multivariate extension is a candidate\-projection evaluator rather than an automatic structure\-discovery algorithm: higher\-order projections must be specified by domain knowledge, a predefined evaluation plan, or a targeted stress test\.

A candidate multivariate projection may be screened using the following state\-space and real\-support conditions:

\|𝒞XA\|​K\\displaystyle\|\\mathcal\{C\}\_\{X\_\{A\}\}\|\\,K≤Bmax,\\displaystyle\\leq B\_\{\\max\},\(2\)cov≥nmin⁡\(A\)\\displaystyle\\operatorname\{cov\}\_\{\\geq n\_\{\\min\}\}\(A\)=∑u∈𝒰ApR​\(u\),\\displaystyle=\\sum\_\{u\\in\\mathcal\{U\}\_\{A\}\}p\_\{R\}\(u\),cov≥nmin⁡\(A\)\\displaystyle\\operatorname\{cov\}\_\{\\geq n\_\{\\min\}\}\(A\)≥ρ,\\displaystyle\\geq\\rho,wherepR​\(u\)=P^R​\(CXA=u\)p\_\{R\}\(u\)=\\hat\{P\}\_\{R\}\(C\_\{X\_\{A\}\}=u\), and𝒰A=\{u:nR​\(CXA=u\)≥nmin\}\\mathcal\{U\}\_\{A\}=\\\{u:n\_\{R\}\(C\_\{X\_\{A\}\}=u\)\\geq n\_\{\\min\}\\\}is the set of real\-data joint tabular states with at leastnminn\_\{\\min\}records\.BmaxB\_\{\\max\},nminn\_\{\\min\}, andρ\\rhoare user\-specified evaluation\-budget parameters rather than learned model parameters\. The reported real\-data experiments use only the pairwise case\(\|A\|=1\)\(\|A\|=1\); the XOR/parity experiment is a prespecified stress test rather than a projection selected by this screening rule\. Here,BmaxB\_\{\\max\},nminn\_\{\\min\}, andρ\\rhoare evaluation\-budget parameters rather than learned model parameters\.BmaxB\_\{\\max\}limits the number of cells in the projected contingency table,nminn\_\{\\min\}defines the minimum real\-data support required for a tabular joint state to be treated as stable, andρ\\rhorequires that such stable states cover a sufficient fraction of the real\-data mass\.

In our implementation, text fields are encoded with a fixed Sentence\-BERT encoder and clustered withkk\-means fitted on real training embeddings, yielding the text semantic\-state variableCTC\_\{T\}\. Continuous tabular features are discretized using training\-split quantile bins, while categorical features retain their observed training\-split categories, yielding tabular\-state variablesCXjC\_\{X\_\{j\}\}\. For any selected projection\(CX,CT\)\(C\_\{X\},C\_\{T\}\), the empirical real projected distribution is

P^\(CX=u,CT=v\)=1\|R\|∑i∈R𝕀\[cX\(i\)=u∧cT\(i\)=v\],\\hat\{P\}\(C\_\{X\}=u,C\_\{T\}=v\)=\\frac\{1\}\{\|R\|\}\\sum\_\{i\\in R\}\\mathbb\{I\}\[c\_\{X\}^\{\(i\)\}=u\\land c\_\{T\}^\{\(i\)\}=v\],\(3\)and the synthetic projected distributionQ^\\hat\{Q\}is defined analogously onSS\.

### III\-BAxis I: Projection Fidelity via Jensen–Shannon Divergence

Projection fidelity measures whether the synthetic quantized table matches the real quantized table\. We use Jensen–Shannon divergence \(JSD\), defined in terms of Kullback–Leibler \(KL\) divergence:

JSD\(P^∥Q^\)\\displaystyle\\mathrm\{JSD\}\(\\hat\{P\}\\parallel\\hat\{Q\}\)=12DKL\(P^∥M\)\+12DKL\(Q^∥M\),\\displaystyle=\\tfrac\{1\}\{2\}D\_\{\\mathrm\{KL\}\}\(\\hat\{P\}\\parallel M\)\+\\tfrac\{1\}\{2\}D\_\{\\mathrm\{KL\}\}\(\\hat\{Q\}\\parallel M\),\(4\)M\\displaystyle M=12​\(P^\+Q^\)\.\\displaystyle=\\tfrac\{1\}\{2\}\(\\hat\{P\}\+\\hat\{Q\}\)\.With base\-2 logarithms, JSD lies in\[0,1\]\[0,1\], and lower values indicate higher projected fidelity\. Importantly, JSD is an overall table\-fidelity metric: it can increase because of tabular marginal mismatch, text\-cluster marginal mismatch, or dependence mismatch\. We therefore complement it with dependence\-preservation diagnostics\. We use the convention0​log⁡0=00\\log 0=0\.

### III\-CAxis II: Cross\-Modal Dependence via Normalized Mutual Information and Conditional Jensen–Shannon Divergence

To more directly measure dependence preservation, we compute normalized mutual information \(NMI\):

NMI⁡\(CX,CT\)=I⁡\(CX,CT\)H⁡\(CX\)​H​\(CT\)\.\\mathrm\{NMI\}\(C\_\{X\},C\_\{T\}\)=\\frac\{I\(C\_\{X\};C\_\{T\}\)\}\{\\sqrt\{H\(C\_\{X\}\)H\(C\_\{T\}\)\}\}\.\(5\)We reportNMIreal\\mathrm\{NMI\}\_\{\\mathrm\{real\}\},NMIsynth\\mathrm\{NMI\}\_\{\\mathrm\{synth\}\}, the signed gapΔNMI=NMIreal−NMIsynth\\Delta\_\{\\mathrm\{NMI\}\}=\\mathrm\{NMI\}\_\{\\mathrm\{real\}\}\-\\mathrm\{NMI\}\_\{\\mathrm\{synth\}\}, and the ratioNMIsynth/NMIreal\\mathrm\{NMI\}\_\{\\mathrm\{synth\}\}/\\mathrm\{NMI\}\_\{\\mathrm\{real\}\}\. A positive signed gap indicates weaker measured dependence in the synthetic projection, whereas a negative signed gap indicates stronger measured dependence than in the corresponding real projection\. The ratio is reported only for non\-degenerate real projections withH⁡\(CX\)\>0H\(C\_\{X\}\)\>0,H⁡\(CT\)\>0H\(C\_\{T\}\)\>0, andNMIreal\>0\\mathrm\{NMI\}\_\{\\mathrm\{real\}\}\>0\. Because this ratio can be unstable when real dependence is close to the finite\-sample shuffle floor, we interpret it together with the signed gap and the shuffle baseline\. A ratio above one is therefore not automatically favorable\.

For conditional diagnostics, we writeP^R\\hat\{P\}\_\{R\}andP^S\\hat\{P\}\_\{S\}for the empirical real and synthetic projected distributions, respectively\. We also compute conditional Jensen–Shannon divergence \(cJSD\) over text\-cluster distributions given tabular states:

cJSD𝒰\\displaystyle\\mathrm\{cJSD\}\_\{\\mathcal\{U\}\}=∑u∈𝒰wu​du,\\displaystyle=\\sum\_\{u\\in\\mathcal\{U\}\}w\_\{u\}d\_\{u\},\(6\)du\\displaystyle d\_\{u\}=JSD\(P^R\(⋅∣u\),P^S\(⋅∣u\)\),\\displaystyle=\\mathrm\{JSD\}\\\!\\left\(\\hat\{P\}\_\{R\}\(\\cdot\\mid u\),\\hat\{P\}\_\{S\}\(\\cdot\\mid u\)\\right\),wu\\displaystyle w\_\{u\}=P^R​\(CX=u\)∑u′∈𝒰P^R​\(CX=u′\)\.\\displaystyle=\\frac\{\\hat\{P\}\_\{R\}\(C\_\{X\}=u\)\}\{\\sum\_\{u^\{\\prime\}\\in\\mathcal\{U\}\}\\hat\{P\}\_\{R\}\(C\_\{X\}=u^\{\\prime\}\)\}\.HereP^R\(⋅∣u\)\\hat\{P\}\_\{R\}\(\\cdot\\mid u\)andP^S\(⋅∣u\)\\hat\{P\}\_\{S\}\(\\cdot\\mid u\)are the empirical real and synthetic text\-cluster distributions conditional onCX=uC\_\{X\}=u\. To reduce small\-cell instability, cJSD is computed on𝒰=\{u:nR\(CX=u\)≥nminR,nS\(CX=u\)≥nminS\}\\mathcal\{U\}=\\\{u:n\_\{R\}\(C\_\{X\}=u\)\\geq n\_\{\\min\}^\{R\},\\,n\_\{S\}\(C\_\{X\}=u\)\\geq n\_\{\\min\}^\{S\}\\\}, withnminR=nminS=5n\_\{\\min\}^\{R\}=n\_\{\\min\}^\{S\}=5\. The threshold of 5 is a small\-cell support rule: states with fewer records yield highly variable empirical conditional distributions and are therefore excluded from cJSD, while their excluded mass is reported throughcovR\\mathrm\{cov\}\_\{R\}andcovS\\mathrm\{cov\}\_\{S\}\. The weightswuw\_\{u\}renormalize real\-data mass over𝒰\\mathcal\{U\}\. We also reportcovR=∑u∈𝒰P^R​\(CX=u\)\\mathrm\{cov\}\_\{R\}=\\sum\_\{u\\in\\mathcal\{U\}\}\\hat\{P\}\_\{R\}\(C\_\{X\}=u\)andcovS=∑u∈𝒰P^S​\(CX=u\)\\mathrm\{cov\}\_\{S\}=\\sum\_\{u\\in\\mathcal\{U\}\}\\hat\{P\}\_\{S\}\(C\_\{X\}=u\), so partial\-support cJSD estimates are not compared as full\-support estimates\.

### III\-DAxis III: Utility via Cross\-Modal Predictability

Utility is evaluated using a train\-on\-synthetic, test\-on\-real \(TSTR\) predictive probe\. In the text\-to\-attribute \(T2A\) direction used in our experiments, a dataset\-specific classifier is trained on synthetic text embeddings to predict the corresponding tabular target and is evaluated on real holdout records\. We report macro\-averaged F1 for multiclass tasks\. For the imbalanced Fake Jobs task, we additionally report balanced accuracy, the area under the receiver operating characteristic curve \(AUROC\), the area under the precision–recall curve \(AUPRC\), the Matthews correlation coefficient \(MCC\), and minority\-class recall\.

### III\-EAxis IV: Relative Joint\-State Entropy

To screen for projected cross\-modal under\-coverage or mode collapse, our framework measures diversity over the projected quantized table rather than over isolated marginals\. For a synthetic projected distributionQ^\\hat\{Q\}, we compute joint entropy and its real\-relative version:

HQ^​\(CX,CT\)\\displaystyle H\_\{\\hat\{Q\}\}\(C\_\{X\},C\_\{T\}\)=−∑u,vq^u,vlog2q^u,v,\\displaystyle=\-\\sum\_\{u,v\}\\hat\{q\}\_\{u,v\}\\log\_\{2\}\\hat\{q\}\_\{u,v\},\(7\)Hrel​\(S\)\\displaystyle H\_\{\\mathrm\{rel\}\}\(S\)=HQ^​\(CX,CT\)HP^​\(CX,CT\)\.\\displaystyle=\\frac\{H\_\{\\hat\{Q\}\}\(C\_\{X\},C\_\{T\}\)\}\{H\_\{\\hat\{P\}\}\(C\_\{X\},C\_\{T\}\)\}\.whereq^u,v=Q^​\(CX=u,CT=v\)\\hat\{q\}\_\{u,v\}=\\hat\{Q\}\(C\_\{X\}=u,C\_\{T\}=v\)and0​log⁡0=00\\log 0=0\. A value substantially below one may indicate under\-coverage, whereas a value above one may reflect either broader support or weakened dependence\. We therefore interpret the entropy ratio jointly with fidelity and dependence metrics\.

### III\-FAxis V: Holdout\-Calibrated Record Proximity

Record\-proximity evaluation uses the distance to closest record \(DCR\) as an empirical nearest\-neighbor diagnostic\. For DCR only, tabular variables are transformed using feature\-wise preprocessing fitted on the real training split, yieldingXscaledX\_\{\\mathrm\{scaled\}\}\. Continuous variables are z\-scored using real\-training means and standard deviations, and continuous holdout and synthetic values outside the observed training range are clipped to\[min,max\]\[\\min,\\max\]before z\-scoring\. Categorical variables are mapped to integer codes using the training\-split category vocabulary, with unseen synthetic categories mapped to an “unknown” code, and the resulting codes are z\-scored using real\-training means and standard deviations\. This coding is used only to obtain a reproducible proximity screen; it is not a semantic ordinal model of categorical distance\. Since nominal\-category distances depend on the chosen encoding, the resulting DCR values and the holdout\-calibrated proximity flag rate \(PFR\) should be interpreted as representation\-level diagnostics rather than as privacy\-risk estimates\. Alternative mixed\-type distances, such as one\-hot or Gower\-style encodings, are left to future work\. Text embeddings are L2\-normalized record\-wise, yieldingE~\\tilde\{E\}\. The multimodal latent representation is

Z=\[Xscaled∥λ⋆E~\],Z=\[X\_\{\\mathrm\{scaled\}\}\\parallel\\lambda^\{\\star\}\\tilde\{E\}\],\(8\)where∥\\paralleldenotes concatenation\. Feature\-wise tabular preprocessing avoids the scalar\-feature degeneration caused by row\-wise tabular normalization\.

The scaling parameterλ⋆\\lambda^\{\\star\}is estimated once per dataset using real training data:

λ⋆=∑rVar⁡\(Xscaled,r\)∑sVar⁡\(E~s\)\+ϵ,ϵ=10−12\.\\lambda^\{\\star\}=\\sqrt\{\\frac\{\\sum\_\{r\}\\mathrm\{Var\}\(X\_\{\\mathrm\{scaled\},r\}\)\}\{\\sum\_\{s\}\\mathrm\{Var\}\(\\tilde\{E\}\_\{s\}\)\+\\epsilon\}\},\\qquad\\epsilon=10^\{\-12\}\.\(9\)The sameλ⋆\\lambda^\{\\star\}is then applied to all synthetic methods for that dataset\. Variances in Eq\.[9](https://arxiv.org/html/2609.22149#S3.E9)are population variances computed over the real training split after the tabular preprocessing described above and after record\-wise L2 normalization of text embeddings\. This scaling equalizes the total empirical variance of the tabular and text blocks on the real training set\. It does not guarantee equal influence on every nearest\-neighbor comparison; we therefore report sensitivity toλ\\lambda\.

For a synthetic recordss, DCR is

DCR⁡\(s,Rtrain\)=minr∈Rtrain⁡‖Zs−Zr‖2\.\\mathrm\{DCR\}\(s,R\_\{\\mathrm\{train\}\}\)=\\min\_\{r\\in R\_\{\\mathrm\{train\}\}\}\\\|Z\_\{s\}\-Z\_\{r\}\\\|\_\{2\}\.\(10\)Rather than interpreting raw minimum DCR directly, we calibrate a proximity threshold using the real holdout\-to\-train DCR distribution\. Letτα\\tau\_\{\\alpha\}denote itsα\\alpha\-quantile:

τα\\displaystyle\\tau\_\{\\alpha\}=Quantileα⁡\{dh:h∈Rholdout\},\\displaystyle=\\operatorname\{Quantile\}\_\{\\alpha\}\\\{d\_\{h\}:h\\in R\_\{\\mathrm\{holdout\}\}\\\},\(11\)dh\\displaystyle d\_\{h\}=DCR⁡\(h,Rtrain\)\.\\displaystyle=\\mathrm\{DCR\}\(h,R\_\{\\mathrm\{train\}\}\)\.Because nearest\-neighbor distances are computed numerically, we use a small numerical toleranceδnum=10−12\\delta\_\{\\mathrm\{num\}\}=10^\{\-12\}and define the effective calibrated threshold as

τ~α=max⁡\(τα,δnum\)\.\\tilde\{\\tau\}\_\{\\alpha\}=\\max\(\\tau\_\{\\alpha\},\\delta\_\{\\mathrm\{num\}\}\)\.\(12\)The holdout\-calibrated proximity flag rate is then

PFRα\(S\)=1\|S\|∑s∈S𝕀\[DCR\(s,Rtrain\)≤τ~α\]\.\\mathrm\{PFR\}\_\{\\alpha\}\(S\)=\\frac\{1\}\{\|S\|\}\\sum\_\{s\\in S\}\\mathbb\{I\}\\left\[\\mathrm\{DCR\}\(s,R\_\{\\mathrm\{train\}\}\)\\leq\\tilde\{\\tau\}\_\{\\alpha\}\\right\]\.\(13\)
We distinguish the dataset\-level proximity flag rate from individual near\-collision flags\. The PFR value is the fraction of synthetic records that fall inside the holdout\-calibrated proximity region under the selected representation and threshold\. Separately, any synthetic record withDCR≤10−6\\mathrm\{DCR\}\\leq 10^\{\-6\}is treated as a near\-collision requiring case\-level inspection\. A low PFR does not by itself imply privacy or high data quality, because unrealistic synthetic records may also lie far from the training data\. Conversely, a high PFR indicates that many records are close under the selected representation, but it is not a formal estimate of disclosure probability\. DCR and PFR are therefore reported as representation\-level record\-proximity diagnostics rather than as privacy guarantees\.

## IVExperiments and Discussion

The proposed framework is a controlled multimodal evaluation rather than a single\-score generator ranking\. The experiments test whether common tabular\-only and text\-only metrics can pass deliberately misaligned records, and whether projection\-level diagnostics reveal the resulting dependence, utility, diversity, and record\-proximity changes\. The real datasetRRserves as the real\-data reference\. The text\-permutation control constructs a negative controlStiltS\_\{\\mathrm\{tilt\}\}by shuffling text against tabular rows, preserving unimodal marginals while randomizing row\-level cross\-modal pairings\.

Our experiments address four questions: \(Q1\) Can conventional tabular\-only and text\-only metrics incorrectly validate misaligned multimodal records? \(Q2\) Under what real\-dependence conditions does the text\-permutation control produce a detectable cross\-modal violation? \(Q3\) How do representative synthesis strategies preserve, destroy, or overstate cross\-modal fidelity, dependence, diversity, utility, and record proximity? \(Q4\) How sensitive are the conclusions to semantic resolutionKK, DCR scalingλ/λ⋆\\lambda/\\lambda^\{\\star\}, class imbalance, sparse projected support, and higher\-order dependence beyond pairwise projections?

### IV\-AExperimental Setup

#### Datasets\.

We evaluate on three real\-world multimodal datasets: \(*i*\)Amazon Reviews\[[6](https://arxiv.org/html/2609.22149#bib.bib37)\], containing review text paired with structured ratings; \(*ii*\)Kiva Loans\[[4](https://arxiv.org/html/2609.22149#bib.bib35)\], containing borrower narratives paired with structured loan attributes such as sector; and \(*iii*\)Fake Jobs, from the Employment Scam Aegean Dataset \(EMSCAD\)\[[5](https://arxiv.org/html/2609.22149#bib.bib36)\], containing job\-posting descriptions paired with fraud\-related structured indicators\. After preprocessing, the retained evaluation subsets contain 6,064 Amazon records, 4,577 Kiva records, and 2,915 Fake Jobs records\.

#### Baselines\.

We constructStiltS\_\{\\mathrm\{tilt\}\}by shuffling the text column against original tabular rows\. The text remains fluent and the tabular marginals remain unchanged, but row\-level pairings are randomized\. Its detectability depends on whether the evaluated projection contains measurable real cross\-modal dependence\. We compareStiltS\_\{\\mathrm\{tilt\}\}against four learned baselines:SindS\_\{\\mathrm\{ind\}\}, which generates tabular fields and text independently;SseqS\_\{\\mathrm\{seq\}\}, which generates tabular fields before conditioning text generation on them;SjointS\_\{\\mathrm\{joint\}\}, which generates both modalities autoregressively in a single prompt; andSlatS\_\{\\mathrm\{lat\}\}, our TabSyn\-SBERT embedding\-level adaptation, which treats SBERT text embeddings as continuous columns alongside the structured variables\[[38](https://arxiv.org/html/2609.22149#bib.bib38)\]\. TabSyn\-SBERT does not generate surface text directly in this setup\.

#### Implementation and data separation\.

The LLM baselines use an internal eBay deployment of GPT\-5\.2 with ten fixed in\-context demonstrations, temperature 0\.8, and up to three retries for malformed outputs\. The released code uses GPT\-4o\-mini as a public substitute and includes the prompt\-construction logic, retry policy, and raw artifacts\.SindS\_\{\\mathrm\{ind\}\}uses CTGAN for the tabular fields and an independent LLM for text;SseqS\_\{\\mathrm\{seq\}\}first generates the tabular fields and then generates text conditioned on them;SjointS\_\{\\mathrm\{joint\}\}generates both modalities in one prompt; andSlatS\_\{\\mathrm\{lat\}\}applies TabSyn\-SBERT to structured variables and SBERT embeddings\.

The fixed demonstration setDdemoD\_\{\\mathrm\{demo\}\}is disjoint from the real evaluation poolRR\. Therefore, no demonstration record appears in eitherRtrainR\_\{\\mathrm\{train\}\}orRholdoutR\_\{\\mathrm\{holdout\}\}\. We apply an 80/20 stratified split toRR\. All data\-dependent evaluation artifacts are fitted onRtrainR\_\{\\mathrm\{train\}\}\. The holdout partition is not used to fit these artifacts and is used for T2A testing and DCR\-threshold calibration, while projection\-level real reference distributions are computed onRR\. Consequently, the reported LLM T2A scores do not involve direct prompt\-example overlap with the real holdout set\.

Because the CTGAN and TabSyn\-SBERT outputs were produced before the post\-hoc evaluation split, their T2A values and holdout\-calibrated PFR values are interpreted as protocol\-specific diagnostics rather than as strict estimates of generator generalization or privacy under a fully untouched holdout design\.

The text\-permutation controlStiltS\_\{\\mathrm\{tilt\}\}is constructed independently of the 10\-shot prompting procedure by permuting text across real tabular rows\. It is a deliberately constructed negative control rather than a learned generator\. Its JSD, NMI, cJSD, and joint\-entropy results test whether the evaluation metrics respond to disrupted pairings\. Because it reuses real components by design, its T2A and PFR values are reported only as control responses, not as estimates of generator generalization or privacy\.

#### Evaluation configuration\.

Text is encoded using Sentence\-BERT \(all\-MiniLM\-L6\-v2, 384 dimensions\)\[[39](https://arxiv.org/html/2609.22149#bib.bib32)\]\. The main setting usesK=20K=20, ten quantile bins for continuous variables, and 1,000 synthetic records per dataset–method pair\. Text\-cluster quantizers, tabular discretizers, categorical vocabularies, normalization statistics, andλ⋆\\lambda^\{\\star\}are fitted onRtrainR\_\{\\mathrm\{train\}\}and applied unchanged to holdout and synthetic records; the SBERT encoder itself is pretrained and kept fixed\.

T2A probes use frozen SBERT embeddings and prespecified classifier families with fixed hyperparameters: logistic regression for Fake Jobs \(class\_weight=balanced,max\_iter=1000,random\_state=42\) and gradient boosting for Amazon and Kiva \(n\_estimators=100,random\_state=42; Kiva additionally usesmax\_depth=5\)\. We report tabular marginal fidelity, text similarity, JSD, NMI ratio and signed gap, cJSD, T2A,HrelH\_\{\\mathrm\{rel\}\}, and PFR; balanced utility metrics are additionally reported for Fake Jobs\.

### IV\-BModality\-Specific Metrics Do Not Test Cross\-Modal Pairing

Table[I](https://arxiv.org/html/2609.22149#S4.T1)reports the original modality\-specific baseline scores\. For methods that generate surface text, the reported BERTScore\-F1 value compares each generated text with the reference text assigned by the evaluation pipeline\. Because independently generated synthetic records do not have a natural one\-to\-one correspondence with real reference records, this index\-matched score is interpreted only as a descriptive candidate–reference similarity score\. It is not interpreted as a corpus\-level text\-distribution metric\. TabSyn\-SBERT does not generate surface text in this setup, so its embedding cosine similarity is reported in a separate column and is not compared directly with BERTScore\-F1\.

TABLE I:Modality\-specific baseline scores\. Higher values indicate greater similarity within each individual metric column\. BERTScore\-F1 and embedding cosine use different inputs and scales and must not be compared across columns\.The tabular value is the originally reported marginal\-fidelity score\. BERTScore\-F1 is reported only for methods that generate surface text\. TabSyn\-SBERT outputs SBERT embeddings rather than surface text, so its cosine similarity is included only as an embedding\-level reference\. No direct comparison is made between the BERTScore\-F1 and embedding\-cosine columns\.

The modality\-specific scores do not test whether the text matches the structured attributes in the same record\. As a corpus\-level text\-marginal control, we compute the JSD between the real and synthetic text\-cluster marginals,JSD\(P^CT∥Q^CT\)\\mathrm\{JSD\}\(\\hat\{P\}\_\{C\_\{T\}\}\\parallel\\hat\{Q\}\_\{C\_\{T\}\}\), while ignoring the tabular variables\. The text\-permutation control obtains 0\.000 on all three datasets because it preserves the text corpus exactly\. Its joint projected JSD is nevertheless nonzero on Amazon, Kiva, and Fake Jobs \(0\.102, 0\.698, and 0\.275, respectively\)\. Thus, the text marginal can remain unchanged while the tabular–text pairing is disrupted\. We therefore retain the index\-matched BERTScore only as a descriptive modality\-specific reference score and rely on the projected joint diagnostics for the cross\-modal analysis\.

### IV\-CMain Projection\-Level Results

Before applying the permutation control, we compare each real projection with its shuffle baseline\. Real/shuffled NMI is0\.0103/0\.00400\.0103/0\.0040for Amazon,0\.6102/0\.00980\.6102/0\.0098for Kiva, and0\.0869/0\.00320\.0869/0\.0032for Fake Jobs\. Amazon is therefore reported separately as a weak\-dependence boundary case; Kiva and Fake Jobs receive the complete multi\-axis comparison\.

#### Amazon: Weak\-Dependence Boundary Case\.

TABLE II:Projected JSD for the Amazon rating–text boundary case\. Lower is closer to the real reference; weak real dependence limits the permutation control’s discriminative power\.Amazon is a weak\-dependence boundary case: its real NMI \(0\.01030\.0103\) is close to the permutation baseline \(0\.00400\.0040\)\. BecauseStiltS\_\{\\mathrm\{tilt\}\}preserves both modality marginals and the selected real dependence is weak, it attains lower projected JSD than the learned baselines\. This does not validate the shuffled records; it shows that the selected projection has limited discriminative power\. Its high PFR instead reflects deliberate reuse of real components and should be interpreted separately from cross\-modal alignment\.

#### Kiva and Fake Jobs: Complete Multi\-Axis Comparison\.

Table[III](https://arxiv.org/html/2609.22149#S4.T3)reports all metrics for Kiva and Fake Jobs\. The metrics are complementary and are not combined into an overall ranking\. The rowRRanchors the distributional columns; T2A and PFR are not applicable to it\.

TABLE III:Projection\-level results for Kiva and Fake Jobs\. Lower JSD/cJSD and smaller\|ΔNMI\|\\lvert\\Delta\_\{\\mathrm\{NMI\}\}\\rvertare preferred; higher T2A is better;HrelH\_\{\\mathrm\{rel\}\}is interpreted relative to 1\. PFR is a proximity flag rate, not a quality or privacy score\. NMI is ratio / signed gap\. Bold marks best directly comparable synthetic values; partial\-coverage cJSD is excluded\.†TabSyn\-SBERT omits 5 of 13 Kiva sectors \(covR=0\.67\\mathrm\{cov\}\_\{R\}=0\.67,covS=1\.00\\mathrm\{cov\}\_\{S\}=1\.00\); its cJSD is partial\-coverage\.‡Best full\-coverage Kiva cJSD\.

Table[III](https://arxiv.org/html/2609.22149#S4.T3)is a diagnostic table rather than a leaderboard\. Its axes describe different properties of the synthetic data and should not be collapsed into a single overall ranking\. In particular, a method can preserve the measured dependence level while failing to cover important tabular states, or it can achieve a low PFR simply by generating records that are far from the real\-data distribution\.

On Kiva,StiltS\_\{\\mathrm\{tilt\}\}andSindS\_\{\\mathrm\{ind\}\}largely remove the real sector–text dependence\. Among the learned methods with full conditional\-state coverage,SseqS\_\{\\mathrm\{seq\}\}is closest to the real NMI and has the lowest full\-coverage cJSD\.SjointS\_\{\\mathrm\{joint\}\}exhibits slightly stronger measured dependence than the real projection\. TabSyn\-SBERT obtains the highest T2A score and a lower numerical cJSD, but its cJSD applies only to the sectors that it generates; several real sectors receive no synthetic records\. Its conditional result must therefore be interpreted together with the reported coverage rather than as uniformly better alignment\.

On Fake Jobs, TabSyn\-SBERT has the lowest projected JSD and cJSD, the smallest absolute NMI gap, and the highest T2A score\. By contrast,SseqS\_\{\\mathrm\{seq\}\}andSjointS\_\{\\mathrm\{joint\}\}exhibit measured fraud–text dependence more than twice the real\-data level, whileStiltS\_\{\\mathrm\{tilt\}\}andSindS\_\{\\mathrm\{ind\}\}retain only a small fraction of the real dependence\. These results show that cross\-modal fidelity requires a measured dependence level close to the real\-data reference, rather than either minimal dependence or maximally strong dependence\. Because TabSyn\-SBERT directly models the same SBERT representation used by the text clusters and predictive probe, its results should be interpreted as strong embedding\-level performance under the present protocol, not as evidence of superior surface\-text generation or a general advantage of diffusion models over LLMs\.

### IV\-DCross\-Modal Utility under Class Imbalance

The T2A column in Table[III](https://arxiv.org/html/2609.22149#S4.T3)reports the main utility metric\. For Fake Jobs, the target is imbalanced, so positive\-class F1 alone can be misleading\. Table[IV](https://arxiv.org/html/2609.22149#S4.T4)reports balanced metrics\.

TABLE IV:Fake Jobs T2A utility under class imbalance\. Higher is better for all reported utility metrics\. Bold values mark the best value in each metric column among non\-reference methods; ties are bolded\. Min\. Recall denotes minority\-class recall\. Values are rounded; “–” denotes a score\-based metric that was not computed for the hard\-label majority baseline\.The balanced metrics confirm that the Fake Jobs result is not an artifact of positive\-class F1 alone\. The text\-permutation control has balanced accuracy below 0\.5, AUROC close to random, low AUPRC, and negative MCC\. Under the fixed SBERT\-based probe, TabSyn\-SBERT achieves the strongest utility scores and the closest measured dependence level among the evaluated methods\. Because TabSyn\-SBERT directly models the same SBERT representation used by the text clusters and the predictive probe, this is a representation\-matched comparison\. It should not be interpreted as evidence of superior surface\-text generation or as a general advantage of diffusion models over LLMs\.

### IV\-EHoldout\-Calibrated Record\-Proximity Analysis

Table[V](https://arxiv.org/html/2609.22149#S4.T5)reports holdout\-calibrated DCR thresholds and PFR\. Thresholds are displayed to three decimals, but PFR is computed before rounding usingτ~α\\tilde\{\\tau\}\_\{\\alpha\}; hence 0\.000 denotes numerical zero or below the display precision in the DCR representation, not byte\-identical raw records\.

TABLE V:Holdout\-calibrated DCR thresholds and PFR values\. Thresholds are shown to three decimals, but PFR is computed before rounding usingτ~α=max⁡\(τα,δnum\)\\tilde\{\\tau\}\_\{\\alpha\}=\\max\(\\tau\_\{\\alpha\},\\delta\_\{\\mathrm\{num\}\}\)\. For Kiva and Fake Jobs, the displayedτ5%\\tau\_\{5\\%\}andτ10%\\tau\_\{10\\%\}remain 0\.000 under the selected DCR representation, so PFR1%, PFR5%, and PFR10%coincide\.Learned PFR is the maximum overSindS\_\{\\mathrm\{ind\}\},SseqS\_\{\\mathrm\{seq\}\},SjointS\_\{\\mathrm\{joint\}\}, and TabSyn\-SBERT, reported as PFR1%/ PFR5%/ PFR10%\. Collision and near\-collision statements refer to the DCR representation after preprocessing and do not necessarily imply byte\-identical raw records\. We useDCR≤10−6\\mathrm\{DCR\}\\leq 10^\{\-6\}as an individual near\-collision inspection threshold; dataset\-level conclusions are based on the calibrated PFR rather than raw minimum DCR alone\.

The calibrated DCR diagnostic flags the learned baselines at low rates, whereasStiltS\_\{\\mathrm\{tilt\}\}has a high PFR because it reuses real tabular and text components and many of its records fall inside the calibrated proximity region\. We distinguish dataset\-level PFR from individual near\-collision flags: records withDCR≤10−6\\mathrm\{DCR\}\\leq 10^\{\-6\}require case\-level inspection, whereas dataset\-level conclusions are based on PFR\. Under this threshold, no learned baseline produces a near\-collision on Amazon\. On Kiva,SseqS\_\{\\mathrm\{seq\}\}produces approximately ten numerical\-zero records \(PFR1%=0\.010\\mathrm\{PFR\}\_\{1\\%\}=0\.010\), indicating representation\-level duplicates rather than broadly elevated record proximity\. For Kiva and Fake Jobs, the displayed calibration threshold remains 0\.000 through the 10th percentile; this reflects redundancy in the evaluated representation and is not evidence of a formal privacy breach\.

### IV\-FSensitivity and Failure Analysis

Table[VI](https://arxiv.org/html/2609.22149#S4.T6)summarizes robustness and boundary\-condition checks\. These checks probe three practical choices that affect interpretation: the semantic resolutionKK, the DCR scaling parameterλ⋆\\lambda^\{\\star\}, and the ability of pairwise projections to detect higher\-order interactions\.

TABLE VI:Robustness and boundary\-condition summary\.For interpretation, we flag a projection as sparse when its occupied\-cell ratio is at most 30% or its minimum synthetic conditional support is belownminS=5n\_\{\\min\}^\{S\}=5\. Under these checks, Fake Jobs is stable atK=20K=20, Amazon is marginal, and Kiva is sparse under the primary sector projection\. The XOR/parity stress test further shows that pairwise projections can miss higher\-order dependence by construction\. Pairwise projections should therefore be treated as scalable diagnostics rather than full joint\-distribution estimators\.

### IV\-GLimitations

The framework is a projection\-based evaluation, not a full estimator of the continuous multimodal joint distribution\. The framework is most informative when the selected projection contains real dependence clearly separated from the empirical shuffle baseline\. Amazon illustrates the opposite boundary case, in which weak measured dependence limits the discriminative power of permutation\-based diagnostics\. JSD can reflect marginal mismatch as well as dependence mismatch; NMI ratios can be unstable near the shuffle floor and can exceed one under stronger measured dependence than in the real\-data projection; and highHrelH\_\{\\mathrm\{rel\}\}does not imply semantic validity\. Pairwise projections can miss higher\-order interactions, sparse cJSD estimates require coverage reporting, and DCR/PFR are empirical proximity diagnostics rather than formal privacy guarantees\. Metric magnitudes should be compared primarily within the same dataset, projection, and semantic resolutionKK\. Preserving the real\-data dependence structure in selected projections is a necessary dataset\-level criterion, but it is not sufficient evidence of instance\-level semantic consistency\.

## VConclusion

This paper shows that favorable modality\-specific scores are not sufficient to validate the pairing between structured attributes and text in a synthetic dataset\. We extend SynEval with fixed text\-embedding quantization, selected tabular–text projections, a text\-permutation control, and complementary diagnostics for projected fidelity, measured dependence, predictive utility, joint\-state entropy, and representation\-level record proximity\.

Across Amazon Reviews, Kiva Loans, and Fake Jobs, the results show that the usefulness of a projection depends on the level of real\-data dependence it contains\. The Kiva and Fake Jobs projections are clearly separated from their shuffle baselines and reveal substantial changes after text permutation or independent generation\. Amazon instead provides a weak\-dependence boundary case in which the selected rating–text projection has limited power to distinguish disrupted pairings\. Some conditioned LLM baselines also exhibit stronger measured dependence than the corresponding real\-data projections, while TabSyn\-SBERT performs strongly under the embedding\-level evaluation protocol\.

These findings support explicit comparison of synthetic and real cross\-modal association rather than reliance on unimodal scores or on high dependence alone\. The framework remains a projection\-based dataset\-level evaluator: preserving the measured dependence structure in selected projections is a necessary criterion for cross\-modal fidelity, but it is not sufficient evidence that every individual record is semantically correct\. Sparse projected support, weak dependence near the shuffle floor, and higher\-order interactions require separate reporting, while DCR and PFR should be interpreted only as representation\-level record\-proximity diagnostics\.

## Acknowledgment

This work was supported in part by a research grant from eBay\. The authors thank eBay for its support of this research\. The views and conclusions expressed in this paper are those of the authors and do not necessarily reflect the official policies or positions of eBay\.

## Artifact Availability

## Declaration of Generative AI Use

Generative AI systems were used in two ways\. First, an internal eBay deployment of GPT\-5\.2 was used as an experimental generative baseline, as described in Section[IV](https://arxiv.org/html/2609.22149#S4)\. Second, generative AI tools were used to assist with scripting, automation, and grammar checking\. All AI\-assisted code, generated artifacts, and experimental analyses were checked by the authors before inclusion in the paper\.

## References

- \[1\]World Economic Forum\(2025\)Synthetic data: the new data frontier\.Briefing PaperWorld Economic Forum\.Note:Accessed: 2026\-07\-03External Links:[Link](https://reports.weforum.org/docs/WEF_Synthetic_Data_2025.pdf)Cited by:[§I](https://arxiv.org/html/2609.22149#S1.p1.1)\.
- \[2\]A\. E\. Johnson, T\. J\. Pollard, L\. Shen, L\. H\. Lehman, M\. Feng, M\. Ghassemi, B\. Moody, P\. Szolovits, L\. Anthony, and R\. G\. Mark\(2016\)MIMIC\-iii, a freely accessible critical care database\.Scientific data3\(1\),pp\. 1–9\.Cited by:[§I](https://arxiv.org/html/2609.22149#S1.p1.1)\.
- \[3\]K\. Huang, J\. Altosaar, and R\. Ranganath\(2019\)ClinicalBERT: modeling clinical notes and predicting hospital readmission\.arXiv preprint arXiv:1904\.05342\.Cited by:[§I](https://arxiv.org/html/2609.22149#S1.p1.1)\.
- \[4\]Kaggle and Kiva\(2018\)Data science for good: kiva crowdfunding\.Note:[https://www\.kaggle\.com/datasets/kiva/data\-science\-for\-good\-kiva\-crowdfunding](https://www.kaggle.com/datasets/kiva/data-science-for-good-kiva-crowdfunding)Accessed: 2024Cited by:[§I](https://arxiv.org/html/2609.22149#S1.p1.1),[§IV\-A](https://arxiv.org/html/2609.22149#S4.SS1.SSS0.Px1.p1.1)\.
- \[5\]S\. Vidros, C\. Kolias, G\. Kambourakis, and L\. Akoglu\(2017\)Automatic detection of online recruitment frauds: characteristics, methods, and a public dataset\.Future Internet9\(1\),pp\. 6\.Cited by:[§I](https://arxiv.org/html/2609.22149#S1.p1.1),[§IV\-A](https://arxiv.org/html/2609.22149#S4.SS1.SSS0.Px1.p1.1)\.
- \[6\]Y\. Hou, J\. Li, Z\. He, A\. Yan, X\. Chen, and J\. McAuley\(2024\)Bridging language and items for retrieval and recommendation\.arXiv preprint arXiv:2403\.03952\.Cited by:[§I](https://arxiv.org/html/2609.22149#S1.p1.1),[§IV\-A](https://arxiv.org/html/2609.22149#S4.SS1.SSS0.Px1.p1.1)\.
- \[7\]N\. Patki, R\. Wedge, and K\. Veeramachaneni\(2016\)The synthetic data vault\.In2016 IEEE international conference on data science and advanced analytics \(DSAA\),pp\. 399–410\.Cited by:[§I](https://arxiv.org/html/2609.22149#S1.p2.1),[§II\-B](https://arxiv.org/html/2609.22149#S2.SS2.p1.1)\.
- \[8\]Z\. Qian, B\. Cebere, and M\. van der Schaar\(2023\)Synthcity: facilitating innovative use cases of synthetic data in different data modalities\.arXiv preprint arXiv:2301\.07573\.Cited by:[§I](https://arxiv.org/html/2609.22149#S1.p2.1),[§II\-B](https://arxiv.org/html/2609.22149#S2.SS2.p1.1)\.
- \[9\]T\. Zhang, V\. Kishore, F\. Wu, K\. Q\. Weinberger, and Y\. Artzi\(2019\)Bertscore: evaluating text generation with bert\.arXiv preprint arXiv:1904\.09675\.Cited by:[§I](https://arxiv.org/html/2609.22149#S1.p2.1),[§II\-B](https://arxiv.org/html/2609.22149#S2.SS2.p2.1)\.
- \[10\]K\. Pillutla, S\. Swayamdipta, R\. Zellers, J\. Thickstun, S\. Welleck, Y\. Choi, and Z\. Harchaoui\(2021\)MAUVE: measuring the gap between neural text and human text using divergence frontiers\.InAdvances in Neural Information Processing Systems \(NeurIPS\),Vol\.34,pp\. 4816–4828\.Cited by:[§I](https://arxiv.org/html/2609.22149#S1.p2.1),[§II\-B](https://arxiv.org/html/2609.22149#S2.SS2.p2.1)\.
- \[11\]B\. W\. Silverman\(2018\)Density estimation for statistics and data analysis\.Routledge\.Cited by:[§I](https://arxiv.org/html/2609.22149#S1.p4.1)\.
- \[12\]D\. Koller and N\. Friedman\(2009\)Probabilistic graphical models: principles and techniques\.MIT press\.Cited by:[§I](https://arxiv.org/html/2609.22149#S1.p4.1)\.
- \[13\]G\. F\. Cooper\(1990\)The computational complexity of probabilistic inference using bayesian belief networks\.Artificial intelligence42\(2\-3\),pp\. 393–405\.Cited by:[§I](https://arxiv.org/html/2609.22149#S1.p4.1)\.
- \[14\]Y\. Yuan, Y\. Liu, and L\. Cheng\(2024\)A multi\-faceted evaluation framework for assessing synthetic data generated by large language models\.arXiv preprint arXiv:2404\.14445\.Cited by:[§I](https://arxiv.org/html/2609.22149#S1.p5.1),[§II\-B](https://arxiv.org/html/2609.22149#S2.SS2.p3.1)\.
- \[15\]European Union\(2016\)Regulation \(EU\) 2016/679 of the european parliament and of the council of 27 april 2016 on the protection of natural persons with regard to the processing of personal data and on the free movement of such data, and repealing directive 95/46/ec \(general data protection regulation\)\.Vol\.L 119\.External Links:[Link](https://eur-lex.europa.eu/eli/reg/2016/679/oj)Cited by:[§II](https://arxiv.org/html/2609.22149#S2.p1.1)\.
- \[16\]California State Legislature\(2018\)California consumer privacy act of 2018 \(CCPA\)\.Note:California Civil Code §§1798\.100 et seq\.External Links:[Link](https://leginfo.legislature.ca.gov/faces/billTextClient.xhtml?bill_id=201720180AB375)Cited by:[§II](https://arxiv.org/html/2609.22149#S2.p1.1)\.
- \[17\]T\. Baltrušaitis, C\. Ahuja, and L\. Morency\(2018\)Multimodal machine learning: a survey and taxonomy\.IEEE transactions on pattern analysis and machine intelligence41\(2\),pp\. 423–443\.Cited by:[§II\-A](https://arxiv.org/html/2609.22149#S2.SS1.p1.1),[§II](https://arxiv.org/html/2609.22149#S2.p1.1)\.
- \[18\]D\. Heckerman\(2008\)A tutorial on learning with bayesian networks\.Innovations in Bayesian networks: Theory and applications,pp\. 33–82\.Cited by:[§II\-A](https://arxiv.org/html/2609.22149#S2.SS1.p1.1)\.
- \[19\]E\. W\. Frees and E\. A\. Valdez\(1998\)Understanding relationships using copulas\.North American actuarial journal2\(1\),pp\. 1–25\.Cited by:[§II\-A](https://arxiv.org/html/2609.22149#S2.SS1.p1.1)\.
- \[20\]L\. Xu, M\. Skoularidou, A\. Cuesta\-Infante, and K\. Veeramachaneni\(2019\)Modeling tabular data using conditional gan\.Advances in neural information processing systems32\.Cited by:[§II\-A](https://arxiv.org/html/2609.22149#S2.SS1.p1.1)\.
- \[21\]A\. Kotelnikov, D\. Baranchuk, I\. Rubachev, and A\. Babenko\(2023\)Tabddpm: modelling tabular data with diffusion models\.InInternational conference on machine learning,pp\. 17564–17579\.Cited by:[§II\-A](https://arxiv.org/html/2609.22149#S2.SS1.p1.1)\.
- \[22\]Y\. Zhou, M\. R\. Bouadjenek, and S\. Aryal\(2025\)MissDDIM: deterministic and efficient conditional diffusion for tabular data imputation\.arXiv preprint arXiv:2508\.03083\.Cited by:[§II\-A](https://arxiv.org/html/2609.22149#S2.SS1.p1.1)\.
- \[23\]M\. Villaizán\-Vallelado, M\. Salvatori, C\. Segura, and I\. Arapakis\(2025\)Diffusion models for tabular data imputation and synthetic data generation\.ACM Transactions on Knowledge Discovery from Data19\(6\),pp\. 1–32\.Cited by:[§II\-A](https://arxiv.org/html/2609.22149#S2.SS1.p1.1)\.
- \[24\]T\. Brown, B\. Mann, N\. Ryder, M\. Subbiah, J\. D\. Kaplan, P\. Dhariwal, A\. Neelakantan, P\. Shyam, G\. Sastry, A\. Askell,et al\.\(2020\)Language models are few\-shot learners\.Advances in neural information processing systems33,pp\. 1877–1901\.Cited by:[§II\-A](https://arxiv.org/html/2609.22149#S2.SS1.p1.1)\.
- \[25\]H\. Touvron, L\. Martin, K\. Stone, P\. Albert, A\. Almahairi, Y\. Babaei, N\. Bashlykov, S\. Batra, P\. Bhargava, S\. Bhosale,et al\.\(2023\)Llama 2: open foundation and fine\-tuned chat models\.arXiv preprint arXiv:2307\.09288\.Cited by:[§II\-A](https://arxiv.org/html/2609.22149#S2.SS1.p1.1)\.
- \[26\]J\. Kim, T\. Kim, and J\. Choo\(2024\)Epic: effective prompting for imbalanced\-class data synthesis in tabular data classification via large language models\.Advances in Neural Information Processing Systems37,pp\. 31504–31542\.Cited by:[§II\-A](https://arxiv.org/html/2609.22149#S2.SS1.p1.1)\.
- \[27\]V\. Borisov, K\. Seßler, T\. Leemann, M\. Pawelczyk, and G\. Kasneci\(2022\)Language models are realistic tabular data generators\.arXiv preprint arXiv:2210\.06280\.Cited by:[§II\-A](https://arxiv.org/html/2609.22149#S2.SS1.p1.1)\.
- \[28\]Z\. Zhao, R\. Birke, and L\. Chen\(2023\)Tabula: harnessing language models for tabular data synthesis\.arXiv preprint arXiv:2310\.12746\.Cited by:[§II\-A](https://arxiv.org/html/2609.22149#S2.SS1.p1.1)\.
- \[29\]S\. Hegselmann, A\. Buendia, H\. Lang, M\. Agrawal, X\. Jiang, and D\. Sontag\(2023\)TabLLM: few\-shot classification of tabular data with large language models\.InInternational Conference on Artificial Intelligence and Statistics,pp\. 5549–5581\.Cited by:[§II\-A](https://arxiv.org/html/2609.22149#S2.SS1.p1.1)\.
- \[30\]X\. Fang, W\. Xu, F\. A\. Tan, J\. Zhang, Z\. Hu, Y\. Qi, S\. Nickleach, D\. Socolinsky, S\. Sengamedu, and C\. Faloutsos\(2024\)Large language models on tabular data: a survey\.arXiv preprint arXiv:2402\.17944\.Cited by:[§II\-A](https://arxiv.org/html/2609.22149#S2.SS1.p1.1)\.
- \[31\]N\. Gaw, S\. Yousefi, and M\. R\. Gahrooei\(2022\)Multimodal data fusion for systems improvement: a review\.IISE Transactions54\(11\),pp\. 1098–1116\.Cited by:[§II\-A](https://arxiv.org/html/2609.22149#S2.SS1.p1.1)\.
- \[32\]K\. Papineni, S\. Roukos, T\. Ward, and W\. Zhu\(2002\)Bleu: a method for automatic evaluation of machine translation\.InProceedings of the 40th annual meeting of the Association for Computational Linguistics,pp\. 311–318\.Cited by:[§II\-B](https://arxiv.org/html/2609.22149#S2.SS2.p2.1)\.
- \[33\]C\. Lin\(2004\)Rouge: a package for automatic evaluation of summaries\.InText summarization branches out,pp\. 74–81\.Cited by:[§II\-B](https://arxiv.org/html/2609.22149#S2.SS2.p2.1)\.
- \[34\]E\. Choi, S\. Biswal, B\. Malin, J\. Duke, W\. F\. Stewart, and J\. Sun\(2017\)Generating multi\-label discrete patient records using generative adversarial networks\.InMachine learning for healthcare conference,pp\. 286–305\.Cited by:[§II\-C](https://arxiv.org/html/2609.22149#S2.SS3.p1.1)\.
- \[35\]C\. C\. Aggarwal, A\. Hinneburg, and D\. A\. Keim\(2001\)On the surprising behavior of distance metrics in high dimensional space\.InInternational conference on database theory,pp\. 420–434\.Cited by:[§II\-C](https://arxiv.org/html/2609.22149#S2.SS3.p1.1)\.
- \[36\]T\. M\. Cover and J\. A\. Thomas\(2006\)Elements of information theory\.John Wiley & Sons\.Cited by:[§II\-C](https://arxiv.org/html/2609.22149#S2.SS3.p2.1)\.
- \[37\]A\. Krause, C\. Guestrin, A\. Gupta, and J\. Kleinberg\(2006\)Near\-optimal sensor placements: maximizing information while minimizing communication cost\.InProceedings of the 5th international conference on Information processing in sensor networks,pp\. 2–10\.Cited by:[§II\-C](https://arxiv.org/html/2609.22149#S2.SS3.p2.1)\.
- \[38\]H\. Zhang, J\. Zhang, Z\. Shen, B\. Srinivasan, X\. Qin, C\. Faloutsos, H\. Rangwala, and G\. Karypis\(2024\)Mixed\-type tabular data synthesis with score\-based diffusion in latent space\.InInternational Conference on Learning Representations,Vol\.2024,pp\. 52829–52857\.Cited by:[§IV\-A](https://arxiv.org/html/2609.22149#S4.SS1.SSS0.Px2.p1.1)\.
- \[39\]N\. Reimers and I\. Gurevych\(2019\)Sentence\-bert: sentence embeddings using siamese bert\-networks\.InProceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing \(EMNLP\-IJCNLP\),pp\. 3982–3992\.Cited by:[§IV\-A](https://arxiv.org/html/2609.22149#S4.SS1.SSS0.Px4.p1.1)\.