{"id":"8a2ea58a-8cce-4595-ba39-9a43c9f9d522","arxiv_id":"2506.00710","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"RelDiff synthesizes relational databases by generating their foreign-key graph with a stochastic block model and denoising all attributes jointly with a graph-conditioned diffusion model.","lead":"RelDiff creates fake versions of real databases by treating the links between tables as a graph and generating all table contents together, instead of table by table. It reports better fidelity than existing methods on 11 benchmark databases, which matters for sharing sensitive data without exposing the original.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The k=2 GNN receptive field may cut off correlations at schema depth >2, yet 3-hop fidelity is a headline result; a GNN-depth ablation is needed to test this.","rationale":"I selected the k-hop locality assumption as the most load-bearing concern because the paper's headline improvement is specifically about column correlations between connected tables, and the reported results include 3-HOP scores on Berka (max depth 4). The reverse process in Section 4.3 approximates each denoising transition by a GNN with k=2 layers, so a single transition cannot see table pairs at schema distance 3. Iterative sampling over many steps could partially compensate, which is why a k-ablation is the right test rather than an immediate rejection. I agree with the reader's weakest_assumption. Secondary issues such as cardinality being fixed by construction, borrowed baseline numbers, and zero DCR error bars are real reporting weaknesses, but they affect how much the empirical comparisons prove rather than whether the model can represent the dependencies it claims. The recommendation is unchanged CONDITIONAL: the method is promising, but the central fidelity claim needs the GNN-depth ablation before accepting it as general.","tokens_in":24782,"tokens_out":11779,"duration_ms":123981,"concrete_test":"Run the Berka (max depth 4) and F1 (max depth 3) configurations with GNN depth k=3 and k=4, keeping all hyperparameters, training epochs, and evaluation identical, and compare 3-HOP Trend (Berka), 2-HOP Trend (F1), and C2ST-Agg against the reported k=2 numbers. If the deeper models improve beyond the error bars, the k=2 locality assumption is the binding constraint and the multi-table fidelity claim is not robust; if the metrics are unchanged, iterative denoising already propagates the relevant dependencies and the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that RelDiff preserves multi-table correlations across arbitrary schemas rests on the reverse-diffusion approximation in Section 4.3: pθ(x_v^s | x_v^t, N_k(v)_t). With the implementation's k=2 GNN layers (Appendix C), a node's single denoising transition sees only its 2-hop neighborhood. For a leaf table linked through two intermediate tables to a root (schema depth 3, e.g., Berka's reported 3-HOP metric), the leaf and root are distance 3 apart, so neither node's one-step transition can directly condition on the other. Long-range statistical dependencies in the clean data could therefore be missed exactly where the paper reports k-hop correlation gains. Iterative reverse sampling over T=100 steps could in principle propagate information further than one receptive field per step, so the limitation is not a mathematical certainty; it is an unvalidated assumption. Because the paper's strongest empirical claim is about preserving column correlations between connected tables, and because no ablation varies GNN depth, the k=2 locality choice is the most load-bearing unsecured assumption. Eq. (5) also factorizes the reverse posterior over nodes, but the local-receptive-field choice is what makes that approximation concrete and testable. The concern is not that the method is wrong, but that its central fidelity claim may not hold on deeper schemas unless the iterative process compensates.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes RELDIFF, a two-stage generative model for relational databases. Structure is generated by a 2K+SBM random graph model that preserves node degree sequences and hierarchical block structure, with referential integrity enforced by construction. Attributes are synthesized by a joint graph-conditioned diffusion process over the heterogeneous entity graph, parameterized by a GNN with local-neighborhood conditioning, combined with tabular transformer encoders/decoders and per-table MLPs. The method is evaluated on 11 datasets from two benchmarks (SyntheRela and ClavaDDPM) against six baselines using fidelity metrics (Shape, Trend, C2ST, C2ST-Agg, k-hop correlation, cardinality similarity), RDL downstream utility, and a DCR privacy sanity check. The central claims are that RELDIFF consistently outperforms prior methods and achieves up to 80% improvement over the state of the art in preserving column correlations between connected tables.","tokens_in":25081,"tokens_out":5918,"duration_ms":60553,"significance":"If the empirical claims hold, RELDIFF is a strong practical contribution to neural relational database synthesis: it generates arbitrary schemas, guarantees referential integrity by construction, and appears to substantially improve multi-table k-hop correlations over ClavaDDPM and other baselines. The paper ships code and follows two established benchmarks, which aids reproducibility and comparability. The decomposition into structure generation and attribute diffusion is principled and extends the 2K+ framework with a Bayesian SBM. However, the strength of the claims needs to be tightened: the cardinality metric is satisfied by construction rather than learned, the SyntheRela baseline numbers are imported from a prior benchmark rather than rerun, the k=2 GNN receptive field is not shown to support the reported depth-3 correlation gains, and the claim of 'consistently outperforms' is contradicted by several table entries. These issues are fixable without changing the method's core, but they affect how the results should be interpreted.","major_comments":[{"comment":"The reported cardinality similarity scores are circular for this method. Section 4.2 states that structure generation conditions on fixed row counts, exact entity and relationship cardinalities, and the node degree sequence, so the perfect cardinality scores in Tables 1 and 2 are guaranteed by input constraints rather than learned by the generative model. The 'Improv.' columns for Cardinality should not be counted as evidence of learned fidelity; this metric should be either removed from the headline comparisons or explicitly labeled as a constraint-satisfaction check.","section":"§5.2, Tables 1 and 2; §4.2"},{"comment":"The paper approximates the denoising transition as pθ(x_v^s | x_v^t, N_k(v)_t) and sets the number of GNN layers to k=2 in Appendix C, yet it reports k-hop correlation gains at depth 3 on Berka and depth 2 on Instacart. A single denoising step can condition only on a 2-hop neighborhood; the paper provides no ablation on GNN depth and no analysis showing that the T=100 iterative reverse process propagates information along longer paths. Since preserving multi-table correlations is the paper's central empirical claim, an ablation or formal argument is needed to justify that the local-receptive-field assumption does not cap the achievable schema depth.","section":"§4.3, Appendix C, Table 1 (Berka 3-HOP), Table 2 (Instacart 2-HOP)"},{"comment":"The SyntheRela results for all baselines are explicitly imported from Jurkovič et al. (2025) rather than rerun under the same preprocessing, random seeds, and evaluation pipeline as RELDIFF. This makes the head-to-head improvements in Table 1 and the abstract an uncontrolled comparison. Please rerun the baselines in the same environment, or provide a documented justification that the prior numbers are directly comparable to the RELDIFF runs.","section":"Table 1 caption"},{"comment":"The claim that RELDIFF 'consistently outperforms' prior methods is contradicted by several entries in the paper's own tables: Table 1 Cora C2ST-Agg (69.30 for RELDIFF vs. 68.80 for TabularARGN), Table 6 Cora C2ST (54.03 vs. 48.97 for RCTGAN), Table 6 IMDB C2ST (52.07 vs. 49.83 for ClavaDDPM), and Table 6 Walmart C2ST (60.30 vs. 53.50 for ClavaDDPM). The wording should be qualified to reflect that RELDIFF wins on most datasets but not on all, and the specific exceptions should be acknowledged.","section":"Abstract and §5.2"},{"comment":"The preprocessing step that 'convert[s] tables with two parents and no children into many-to-many edges' is asserted to be a one-to-one, lossless transformation, but no formal condition or proof is given. If such a table contains attribute columns beyond its two foreign keys, collapsing it to an edge set can discard information and change the schema semantics. Please state the exact conditions under which the transformation is lossless, or restrict the supported schema class accordingly.","section":"§4.2"}],"minor_comments":[{"comment":"The headers 'A VG 2-W AY' are garbled, and several cells are missing (e.g., PrivLava for Instacart 05); the table should be reformatted for readability.","section":"Table 2"},{"comment":"The 'Improv.' columns should state explicitly that the percentages are relative to the best baseline excluding RELDIFF ORIG, and whether a lower-is-better metric like C2ST-Agg uses a signed improvement.","section":"Tables 1 and 6"},{"comment":"The sentence noting that RELDIFF 'consistently achieves a perfect score' on cardinality because it preserves degree distributions would be better placed in the method section as a property of the design, rather than in the experimental results where it reads as an empirical win.","section":"§5.2"},{"comment":"The parenthetical remark 'we omit defining a common space on which q is defined' makes the notation hard to follow; a concise definition of the joint space and the per-type subspaces would improve readability.","section":"§4.3, Eq. (5)"},{"comment":"The caption states that ClavaDDPM 'retains some hierarchy but not the degrees,' but since RELDIFF takes degrees as an input, the comparison would be clearer if the caption noted that the degree preservation of 2K+SBM is a constraint rather than a learned property.","section":"Figure 2"}],"recommendation":"major_revision","confidential_remarks":"The Table 1 baselines are imported from Jurkovič et al. (2025), a benchmark paper coauthored by the present authors; this should be weighed when evaluating the head-to-head claim. The cardinality circularity and the missing GNN-depth ablation are the most consequential issues; both are addressable within the manuscript's scope. The paper's methodological contribution is solid enough to warrant revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"RelDiff is a real step forward for relational synthetic data, but it overclaims slightly and leaves one load-bearing assumption untested. The core idea is genuinely new: no prior method jointly models arbitrary relational schemas with a hierarchical SBM for structure and a single graph-conditioned diffusion process for attributes across all tables. That decomposition is principled, the implementation is careful, and the evaluation is broad: 11 datasets, 8 metrics, with code released. The k-hop correlation results, especially on deeper schemas like Berka and F1, are strong, and the DCR privacy sanity check is a nice touch.\n\nThe soft spots are real but mostly minor. First, 'consistently outperforms' is not literally true: on Cora C2ST-Agg RelDiff (69.30) is slightly worse than TabARGN (68.80), and on Walmart single-table C2ST it is worse than ClavaDDPM. That is an abstract-level overclaim, not a fatal one. Second, cardinality similarity is guaranteed by construction: Section 4.2 fixes row counts and degree sequences, so perfect scores in Table 1 are circular evidence. The paper openly says this in Section 5.2, but then still presents it as a result in the tables. It should be reframed as a sanity check, not a learned capability. Third, baseline numbers in Table 1 are imported from a prior benchmark rather than rerun. That is transparent and acceptable, but it means the comparison is only as fair as the original benchmark runs. Fourth, the DCR standard errors are all 0.0, which is implausible for three samples and should be corrected.\n\nThe more substantive concern, which the stress-test note puts its finger on, is the k=2 GNN receptive field. The reverse step in Eq. (5) factorizes over nodes and conditions each node on its k-hop neighborhood. With k=2, a leaf and root at schema depth 3 cannot directly influence each other in a single denoising step, yet the paper reports 3-hop correlation gains on Berka. Iterative sampling over T=100 steps could in principle propagate information, so this is not a demonstrated failure; the reported 3-hop results suggest the model is doing something right. But it is an unvalidated assumption, and it is exactly the kind of thing a GNN-depth ablation would settle. Without that ablation, the central fidelity claim on deep schemas is not fully secured.\n\nWho is this for? Researchers working on synthetic data generation, privacy-preserving data sharing, and relational representation learning. It deserves a serious referee. My recommendation: send it to review, and ask the authors to add a GNN-depth ablation, rerun or caveat the borrowed baselines, and fix the cardinality and DCR reporting.","headline":"RelDiff is a genuinely new joint graph-diffusion approach to relational data synthesis with strong results, but it overclaims small wins and needs a GNN-depth ablation to secure its deep-schema fidelity.","tokens_in":25601,"tokens_out":3416,"would_cite":true,"duration_ms":29764,"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":"RelDiff generates complete relational databases by modeling the foreign-key graph with a Bayesian stochastic block model and synthesizing table attributes through a joint graph-conditioned diffusion process, preserving cross-table…","keywords":["relational databases","synthetic data generation","diffusion models","graph neural networks","stochastic block model","referential integrity","foreign keys","mixed-type tabular data"],"falsifier":"On a schema with a foreign-key chain A-B-C-D, engineer an attribute in D that is a deterministic function of an attribute in A, a 3-hop dependency. Train RelDiff and measure the A-D column correlation in the generated data; if the synthetic correlation is near zero while all 1-hop correlations are preserved, the k-hop locality assumption is the limiting factor and the fidelity claim does not extend to deep schemas.","tokens_in":24594,"feed_emoji":"🗄️","tokens_out":6730,"duration_ms":56327,"temperature":0.7,"pith_summary":"The paper tries to establish that a relational database can be synthesized as a whole, not table-by-table, by decomposing the problem into structure and attributes. Structure comes from a 2K+SBM graph generator that samples foreign-key graphs matching the original's degree distribution, cardinalities, and hierarchical organization, so referential integrity holds by construction. Attributes are then produced by a single diffusion model that denoises every table's rows jointly, with a heterogeneous graph neural network letting each row's reverse step see its local neighborhood across the schema. On 11 benchmark databases, this beats six prior methods on multi-table fidelity, with up to 80% improvement in preserving correlations between connected tables, while keeping single-table fidelity nearly intact. If this holds, synthetic relational data can be generated for arbitrary schemas without flattening or pre-specified table orderings.","feed_headline":"Joint graph diffusion beats six baselines on 11 relational databases","feed_subtitle":"Sampling the foreign-key graph with a block model and denoising attributes jointly preserves cross-table correlations.","key_machinery":"The machinery has two coupled parts. 2K+SBM is a nonparametric Bayesian stochastic block model whose partition is learned by nested description-length minimization, constrained so blocks respect table types; it resamples each relationship's edge set under fixed node degree sequences and edge counts, preserving 2K statistics (the joint degree distribution) and hierarchy, which yields referential integrity and correct cardinalities by construction. The attribute model is a hybrid Gaussian-plus-masked diffusion run jointly across all tables: numerical columns use score-based Gaussian diffusion, categorical columns use masked diffusion with a [MASK] state, and dimension tables are handled as fixed lookup embeddings. The denoiser is a heterogeneous GraphSAGE GNN with per-table transformer encoders and decoders, and the key conditional-independence assumption is p_theta(x_v^s | x_v^t, N_k(v)_t), meaning each row's reverse step sees only its noisy k-hop neighborhood, with k=2 GNN layers in the experiments.","core_discovery":"RelDiff's central claim is that explicitly decoupling relational database generation into a microcanonical degree-corrected nested stochastic block model, called 2K+SBM, which resamples the foreign-key entity graph while preserving joint degree distributions and hierarchical structure, and a joint hybrid diffusion process over all tables' numerical and categorical attributes conditioned on the graph through a heterogeneous GNN, yields synthetic databases with both high fidelity and exact referential integrity. The paper argues that prior approaches fail because they reduce relational data to conditionally generated flat tables with fixed orderings and structural assumptions, losing inter-table dependencies. RelDiff instead samples structure first, then denoises the attributes of all entities simultaneously, with each node's denoising step conditioned on its noisy k-hop neighborhood. In experiments across 11 real-world datasets and 8 metrics, the paper reports consistent improvements over ClavaDDPM, RCTGAN, REaLTabFormer, SDV, TabularARGN, and PrivLava, including up to 80% improvement on column correlations between connected tables and a roughly 7x smaller drop from single-table to multi-table fidelity than the closest competitor.","pith_inferences":["The k-hop locality assumption bounds the dependencies RelDiff can represent: schemas where meaningful correlations span more than two foreign-key hops would likely need deeper GNNs or larger k, suggesting a testable trade-off between scalability and fidelity.","Because structure generation is classical and attribute diffusion is Joint, the two components could be swapped independently: a stronger structure generator preserving 3K or motif statistics could upgrade RelDiff without retraining the attribute model.","The DCR scores near 50% are empirical evidence against memorization, but the paper itself notes there is no provable privacy guarantee; adding differential privacy would likely require modifying the diffusion objective rather than the graph sampler.","A natural extension is missing-value imputation: conditioning on partially observed attributes instead of fully masked initial states could turn RelDiff into a relational imputation engine."],"forward_implications":["Referential integrity is guaranteed by the structure sampler rather than by post-hoc repair, so every generated database is schema-consistent.","Multi-table fidelity no longer depends on a fixed table generation order: all tables are denoised jointly, preventing errors from cascading down the hierarchy.","The method handles arbitrary schemas, including multi-child, multi-parent, and deep hierarchies, not just single-parent or many-to-many structures.","Column correlations across connected tables are preserved up to the k-hop neighborhood radius, with measured improvements of up to 80% over the previous state of the art.","Single-table fidelity degrades far less when moving to the relational setting, about seven times less than the closest competitor, indicating that joint generation helps rather than hurts per-table quality."],"supporting_citations":[{"why":"Provides the ClavaDDPM benchmark and evaluation protocol (cardinality, 1-hop/2-hop correlations) that RelDiff must beat, and defines the clustering-guided diffusion baseline.","marker":"Pang et al. (2024)"},{"why":"Supplies the nested stochastic block model inference that finds the hierarchy-preserving partition b* used in 2K+SBM graph generation.","marker":"Peixoto (2014)"},{"why":"Defines the microcanonical degree-corrected SBM whose distribution p(E|b,m,k) RelDiff samples from to preserve degrees and edge counts.","marker":"Peixoto (2017)"},{"why":"Supplies the heterogeneous GraphSAGE architecture and neighborhood sampling that parameterize the k-hop conditional denoiser.","marker":"Hamilton et al. (2017)"},{"why":"Provides the mixed-type tabular diffusion framework (TabDiff) and the stochastic sampler RelDiff adapts for attribute synthesis.","marker":"Shi et al. (2025)"},{"why":"Gives the masked-diffusion reverse transition and loss used for categorical attributes.","marker":"Sahoo et al. (2024)"},{"why":"Establishes the score-based ODE formulation and noise schedule that govern the Gaussian diffusion for numerical attributes.","marker":"Karras et al. (2022)"},{"why":"Introduces the structure/attribute decomposition for many-to-many relational data that RelDiff generalizes to arbitrary hierarchical schemas.","marker":"Xu et al. (2023)"},{"why":"Defines the RDL schema graph and relational entity graph formalism that RelDiff adopts for heterogeneous graph generation.","marker":"Fey et al. (2024)"},{"why":"Supplies the SyntheRela benchmark and the C2ST-Agg metric used to measure multi-table fidelity.","marker":"Jurkovič et al. (2025)"}],"fun_headline_variants":["Decouple graph and attributes: RelDiff wins on 11 databases","Block model for foreign keys + diffusion for attributes = better synthetic DBs","RelDiff: joint graph and attribute diffusion improves synthetic relational data","Graph diffusion outdoes flat-table methods on relational benchmarks","Structure-then-diffusion: RelDiff beats six baselines on 11 datasets"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a row's attributes can be generated from its noisy two-hop neighborhood alone: any inter-table dependency that travels through a longer foreign-key path will not shape the generated values, so on deep schemas the claimed multi-table fidelity would degrade.","fun_headline_variants_meta":{"raw":{"variants":["Decouple graph and attributes: RelDiff wins on 11 databases","Block model for foreign keys + diffusion for attributes = better synthetic DBs","RelDiff: joint graph and attribute diffusion improves synthetic relational data","Graph diffusion outdoes flat-table methods on relational benchmarks","Structure-then-diffusion: RelDiff beats six baselines on 11 datasets"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000533,"raw_usage":{"total_tokens":2564,"prompt_tokens":946,"completion_tokens":1618,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":562,"completion_tokens_details":{"reasoning_tokens":1526}},"tokens_in":562,"tokens_out":1618,"duration_ms":11794,"temperature":1.0,"reasoning_tokens":1526,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T11:59:36.857428+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a schema with a foreign-key chain A-B-C-D, engineer an attribute in D that is a deterministic function of an attribute in A, a 3-hop dependency. Train RelDiff and measure the A-D column correlation in the generated data; if the synthetic correlation is near zero while all 1-hop correlations are preserved, the k-hop locality assumption is the limiting factor and the fidelity claim does not extend to deep schemas.","supporting_citations":[{"cited_title":"Clava DDPM : Multi-relational Data Synthesis with Cluster-guided Diffusion Models // The Thirty-eighth Annual Conference on Neural Information Processing Systems","cited_arxiv_id":null,"evidence_quote":"Provides the ClavaDDPM benchmark and evaluation protocol (cardinality, 1-hop/2-hop correlations) that RelDiff must beat, and defines the clustering-guided diffusion baseline."},{"cited_title":"Hierarchical block structures and high-resolution model selection in large networks // Physical Review X","cited_arxiv_id":null,"evidence_quote":"Supplies the nested stochastic block model inference that finds the hierarchy-preserving partition b* used in 2K+SBM graph generation."},{"cited_title":"Nonparametric Bayesian inference of the microcanonical stochastic block model // Physical Review E","cited_arxiv_id":null,"evidence_quote":"Defines the microcanonical degree-corrected SBM whose distribution p(E|b,m,k) RelDiff samples from to preserve degrees and edge counts."},{"cited_title":"Inductive representation learning on large graphs // Advances in neural information processing systems","cited_arxiv_id":null,"evidence_quote":"Supplies the heterogeneous GraphSAGE architecture and neighborhood sampling that parameterize the k-hop conditional denoiser."},{"cited_title":"TabDiff: a Mixed-type Diffusion Model for Tabular Data Generation // The Thirteenth International Conference on Learning Representations","cited_arxiv_id":null,"evidence_quote":"Provides the mixed-type tabular diffusion framework (TabDiff) and the stochastic sampler RelDiff adapts for attribute synthesis."},{"cited_title":"Simple and Effective Masked Diffusion Language Models // The Thirty-eighth Annual Conference on Neural Information Processing Systems","cited_arxiv_id":null,"evidence_quote":"Gives the masked-diffusion reverse transition and loss used for categorical attributes."},{"cited_title":"Elucidating the Design Space of Diffusion-Based Generative Models // Advances in Neural Information Processing Systems","cited_arxiv_id":null,"evidence_quote":"Establishes the score-based ODE formulation and noise schedule that govern the Gaussian diffusion for numerical attributes."}],"review_version":1}