Generating Logically Consistent Synthetic Supply Chain Data with LLM-Driven Knowledge Graph Reasoning

Alexandra Brintrup; Ge Zheng; Liming Xu; Yunbo Long

arxiv: 2605.26823 · v1 · pith:J2TX7ZNYnew · submitted 2026-05-26 · 💻 cs.CL

Generating Logically Consistent Synthetic Supply Chain Data with LLM-Driven Knowledge Graph Reasoning

Yunbo Long , Ge Zheng , Liming Xu , Alexandra Brintrup This is my paper

Pith reviewed 2026-06-29 18:36 UTC · model grok-4.3

classification 💻 cs.CL

keywords synthetic data generationsupply chain analyticsknowledge graphlogical consistencytabular dataLLM reasoningoperational dependenciesCR-KG

0 comments

The pith

TabKG uses a validated Column Relationship Knowledge Graph to generate synthetic supply chain data that is logically consistent by construction.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Supply chain analytics requires synthetic data that respects operational rules such as column dependencies, temporal orderings, and conditional logic, beyond mere statistical resemblance. TabKG constructs a Column Relationship Knowledge Graph by using a multi-LLM ensemble with majority voting to propose relationships from column metadata, then validates those proposals against real data to discard unsupported edges. Independent columns are generated via a latent diffusion model and dependent columns are reconstructed deterministically according to the validated graph. This ensures logical consistency by construction rather than relying on implicit learning by the generative model. A reader would care because operationally plausible synthetic records enable reliable simulation and decision-making while addressing data scarcity and privacy constraints.

Core claim

The paper presents TabKG, a knowledge-graph-guided framework for generating logically consistent synthetic supply chain tabular data. It builds a Column Relationship Knowledge Graph (CR-KG) to represent operational dependencies, employs a multi-LLM ensemble with majority voting to propose candidate relationships from metadata, validates these against real data to remove hallucinated or unsupported edges, and uses the validated CR-KG to guide generation: independent columns are produced with a latent diffusion model while dependent columns are deterministically reconstructed to satisfy the discovered relationships, enforcing logical consistency by construction.

What carries the argument

The Column Relationship Knowledge Graph (CR-KG), which encodes operational dependencies between columns and directs the deterministic reconstruction of dependent columns after independent-column generation.

If this is right

Synthetic records will satisfy all validated operational dependencies by design.
The method separates statistical generation of independent columns from rule-based reconstruction of dependent columns.
Validation against real data removes hallucinated relationships proposed by the LLM ensemble.
The resulting data supports operational simulation and decision-making without violating process constraints.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

If the validation step misses a dependency, generated records may still violate unrepresented operational rules in practice.
The CR-KG construction approach could transfer to other tabular domains with strong relational constraints, such as financial or healthcare records.
Incremental updates to the CR-KG as new real data arrives might allow the framework to adapt over time without full re-validation.
Pairing TabKG with differential privacy mechanisms could further improve utility for sensitive supply chain datasets.

Load-bearing premise

The multi-LLM ensemble with majority voting plus validation against real data will identify a sufficiently complete and accurate set of operational relationships without missing critical dependencies or retaining hallucinated edges.

What would settle it

Generating synthetic records and finding that a non-negligible fraction violates a known operational dependency that the validated CR-KG failed to capture, or that an edge retained after validation has no supporting evidence in the real data.

Figures

Figures reproduced from arXiv: 2605.26823 by Alexandra Brintrup, Ge Zheng, Liming Xu, Yunbo Long.

**Figure 1.** Figure 1: Illustration of inter-column logical consistency problems in synthetic supply chain data. Real data (left) maintains valid hierarchical, temporal, and mathematical relationships, while synthetic data generated by existing methods (right) frequently violates these constraints, producing records in which cities map to incorrect countries, delivery dates precede order dates, and totals do not equal price mult… view at source ↗

**Figure 2.** Figure 2: Overview of the TabKG framework. Stage 1 serializes column metadata. Stage 2 constructs a candidate knowledge graph via a multi-LLM ensemble with majority voting. Stage 3 validates each edge against real data and prunes hallucinated relationships. Stage 4 uses the validated graph for compression and latent diffusion-based generation. Stage 5 reconstructs the full synthetic dataset via knowledge graph-guide… view at source ↗

**Figure 3.** Figure 3: Example Column Relationship Knowledge Graph (CR-KG) for the Retail dataset. Bold-bordered nodes represent independent columns retained for generation; faded nodes are compressed away and reconstructed during decompression. dependency (whether the source column values xs uniquely determine the target column values xt), mathematical edges by evaluating formula accuracy on real rows, temporal edges by checki… view at source ↗

**Figure 4.** Figure 4: Evaluation framework for synthetic tabular data in supply chain settings. based framework for tabular data synthesis. TabDDPM (Kotelnikov et al. 2023) is included as a foundational diffusion model for tabular data. TabSyn (Zhang et al. 2023) is included as a recent diffusion-based method for mixed-type tabular generation. GReaT (Borisov et al. 2022) is included as an LLM-based method that reformulates tabu… view at source ↗

**Figure 5.** Figure 5: Density plots for the three continuous columns (item profit ratio, product price, and latitude), comparing the distribution of real data and their synthetic counterparts generated by CTGAN, TabDDPM, GReaT, TabSyn, and TabKG. Curves that more closely align with the real data indicate better performance. Both TabKG and TabSyn exhibit distributions that closely match the real data, outperforming other methods… view at source ↗

**Figure 6.** Figure 6: Distribution plots for the three categorical columns (shipping mode, payment type, and order status), comparing synthetic data generated by CTGAN, TabDDPM, GReaT, TabSyn, and TabKG to real data. Distributions that closely match the real data indicate superior performance. Both TabKG and TabSyn exhibit distributions that are significantly closer to the real data compared to other methods. (a) CTGAN (b) TabD… view at source ↗

**Figure 7.** Figure 7: Heatmap illustrating the absolute divergence in pairwise column correlations between the synthetic and real data. Lighter colors indicate smaller differences and better alignment. TabSyn and TabKG exhibit the closest alignment with the real data, outperforming other methods. continuous columns ( [PITH_FULL_IMAGE:figures/full_fig_p013_7.png] view at source ↗

read the original abstract

Synthetic data offers a promising solution to two persistent barriers in supply chain analytics: data scarcity and data privacy. However, for synthetic data to support operational simulation and decision-making, it must do more than reproduce the statistical distributions of real records, and also preserve the \emph{operational logic} that governs supply chain processes, including the temporal orderings, mathematical dependencies, hierarchical taxonomies, and conditional rules that make a record operationally plausible. We consider this logic as the ``physics'' of supply chain data. Existing tabular generative models are primarily optimized for distributional fidelity and downstream predictive utility, and therefore often generate records that appear statistically realistic but violate fundamental operational constraints. This paper introduces \textbf{\textit{TabKG}}, a knowledge-graph-guided framework for logically consistent synthetic supply chain tabular data generation. TabKG constructs a \textbf{\textit{Column Relationship Knowledge Graph (CR-KG)}} to represent data operational dependencies. It uses a multi-LLM ensemble with majority voting to propose candidate relationships from column metadata, validates these relationships against real data to remove hallucinated or unsupported edges, and then uses the validated CR-KG to guide generation. Specifically, TabKG compresses the original table into independent columns, generates these columns using a latent diffusion model, and deterministically reconstructs dependent columns according to the validated relationships, enforcing logical consistency by construction with respect to the discovered operational rules.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

TabKG combines LLM-proposed CR-KG edges, real-data validation, and deterministic reconstruction after diffusion on independent columns, but the completeness of the captured rules remains the open question.

read the letter

The core idea is to let a multi-LLM ensemble suggest column relationships from metadata, drop any that lack support in real data, then generate only the independent columns with a latent diffusion model and fill in the rest exactly according to the surviving rules. That pipeline is the new piece; prior tabular synthesis work has not, to my knowledge, used a validated knowledge graph for deterministic reconstruction in the supply-chain setting.

It is a direct response to the practical complaint that statistically plausible records often break operational constraints such as lead-time ordering or conditional quantity rules. The validation step against real data is a sensible guard against hallucinated edges, and the deterministic reconstruction step does enforce consistency for whatever rules make it into the graph.

The soft spot is exactly the one the stress-test note flags: validation only removes unsupported proposals; it does not search for missing ones. If an important conditional or temporal dependency is never suggested by the ensemble, the diffusion step can still produce values that, once reconstructed, violate it. The abstract gives no held-out constraint checks, expert review, or completeness metric, so we cannot yet tell how large that gap is in practice.

The citation pattern looks standard for the area and there is no visible circularity. The work is aimed at researchers who need synthetic supply-chain tables that can actually drive simulation rather than just match marginals. It is coherent enough on its own terms to warrant referee time; the experiments will decide whether the completeness issue is minor or central.

Referee Report

1 major / 1 minor

Summary. The paper introduces TabKG, a knowledge-graph-guided framework for generating logically consistent synthetic supply chain tabular data. It builds a Column Relationship Knowledge Graph (CR-KG) by using a multi-LLM ensemble with majority voting to propose candidate relationships from column metadata, validates these against real data to remove unsupported edges, generates independent columns using a latent diffusion model, and deterministically reconstructs dependent columns based on the validated CR-KG to enforce operational logic including temporal orderings, mathematical dependencies, and conditional rules by construction.

Significance. If the method successfully captures the operational rules, it would represent a meaningful advance over standard tabular generative models that focus only on distributional fidelity, enabling synthetic data suitable for operational simulation and decision-making in supply chains. The integration of LLM-driven knowledge graph construction with validation and deterministic reconstruction is a promising direction for incorporating domain logic into synthetic data generation.

major comments (1)

[Abstract] Abstract: The central claim that logical consistency is enforced 'by construction' with respect to the discovered operational rules assumes the validated CR-KG contains every relevant dependency (temporal orderings, conditional rules, mathematical dependencies, taxonomies). The pipeline only validates LLM-proposed candidates against real data to remove unsupported edges; no mechanism is described for detecting or adding missing edges (e.g., via exhaustive enumeration, expert review, or held-out constraint checking). This is load-bearing, as an incomplete CR-KG would permit the independent-column diffusion step to generate values that violate unrepresented rules after deterministic reconstruction.

minor comments (1)

[Abstract] The abstract states that the table is 'compressed' into independent columns but does not specify the selection criteria or algorithm for partitioning columns into independent vs. dependent sets.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the thoughtful and detailed review. The comment correctly identifies a key scoping issue in how the method's guarantees are presented. We address it directly below and will revise the manuscript accordingly.

read point-by-point responses

Referee: [Abstract] Abstract: The central claim that logical consistency is enforced 'by construction' with respect to the discovered operational rules assumes the validated CR-KG contains every relevant dependency (temporal orderings, conditional rules, mathematical dependencies, taxonomies). The pipeline only validates LLM-proposed candidates against real data to remove unsupported edges; no mechanism is described for detecting or adding missing edges (e.g., via exhaustive enumeration, expert review, or held-out constraint checking). This is load-bearing, as an incomplete CR-KG would permit the independent-column diffusion step to generate values that violate unrepresented rules after deterministic reconstruction.

Authors: We agree with the observation. The current pipeline proposes candidate edges via the LLM ensemble, then removes those unsupported by real data; it contains no procedure for identifying or incorporating missing edges. Consequently, the CR-KG is not guaranteed to be complete with respect to all operational rules that may exist in the domain. The abstract's phrasing ('with respect to the discovered operational rules') is therefore technically accurate but could be misread as implying broader coverage. We will (1) revise the abstract to emphasize that consistency holds only for relationships present in the validated CR-KG, (2) add an explicit Limitations paragraph discussing the dependence on LLM-proposed candidates and the absence of exhaustive or expert-driven rule discovery, and (3) note this as an avenue for future work. These changes will be made in the next revision. revision: yes

Circularity Check

0 steps flagged

No circularity: consistency enforced from externally validated rules

full rationale

The derivation chain extracts candidate relationships via LLM ensemble on metadata, validates support against real data (removing unsupported edges), then uses the resulting CR-KG only for deterministic reconstruction after independent-column diffusion. Logical consistency is defined relative to this externally validated graph rather than being fitted to or defined by the generation target itself. No self-definitional loops, no fitted parameters renamed as predictions, no load-bearing self-citations, and no ansatz or uniqueness claims imported from prior author work appear in the provided text. The method is self-contained against the real-data validation step.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 2 invented entities

Ledger populated from abstract only; limited visibility into full paper assumptions.

axioms (1)

domain assumption Supply chain data is governed by operational logic consisting of temporal orderings, mathematical dependencies, hierarchical taxonomies, and conditional rules.
Explicitly stated in the abstract as the 'physics' that synthetic data must preserve.

invented entities (2)

Column Relationship Knowledge Graph (CR-KG) no independent evidence
purpose: Represent operational dependencies between table columns to guide consistent data generation.
New construct introduced by the paper; no independent evidence provided in abstract.
TabKG framework no independent evidence
purpose: End-to-end system for logically consistent synthetic supply chain data generation.
Proposed method; no external validation of the full pipeline in abstract.

pith-pipeline@v0.9.1-grok · 5781 in / 1310 out tokens · 27308 ms · 2026-06-29T18:36:11.528727+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

4 extracted references · 1 canonical work pages · 1 internal anchor

[1]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...
[2]

@esa (Ref

\@ifxundefined[1] #1\@undefined \@firstoftwo \@secondoftwo \@ifnum[1] #1 \@firstoftwo \@secondoftwo \@ifx[1] #1 \@firstoftwo \@secondoftwo [2] @ #1 \@temptokena #2 #1 @ \@temptokena \@ifclassloaded agu2001 natbib The agu2001 class already includes natbib coding, so you should not add it explicitly Type <Return> for now, but then later remove the command n...
[3]

\@lbibitem[] @bibitem@first@sw\@secondoftwo \@lbibitem[#1]#2 \@extra@b@citeb \@ifundefined br@#2\@extra@b@citeb \@namedef br@#2 \@nameuse br@#2\@extra@b@citeb \@ifundefined b@#2\@extra@b@citeb @num @parse #2 @tmp #1 NAT@b@open@#2 NAT@b@shut@#2 \@ifnum @merge>\@ne @bibitem@first@sw \@firstoftwo \@ifundefined NAT@b*@#2 \@firstoftwo @num @NAT@ctr \@secondoft...
[4]

Evaluating Inter-Column Logical Relationships in Synthetic Tabular Data Generation

@open @close @open @close and [1] URL: #1 \@ifundefined chapter * \@mkboth \@ifxundefined @sectionbib * \@mkboth * \@mkboth\@gobbletwo \@ifclassloaded amsart * \@ifclassloaded amsbook * \@ifxundefined @heading @heading NAT@ctr thebibliography [1] @ \@biblabel @NAT@ctr \@bibsetup #1 @NAT@ctr @ @openbib .11em \@plus.33em \@minus.07em 4000 4000 `\.\@m @bibit...

work page internal anchor Pith review doi:10.48550/arxiv.2502.04055 2023

[1] [1]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

[2] [2]

@esa (Ref

\@ifxundefined[1] #1\@undefined \@firstoftwo \@secondoftwo \@ifnum[1] #1 \@firstoftwo \@secondoftwo \@ifx[1] #1 \@firstoftwo \@secondoftwo [2] @ #1 \@temptokena #2 #1 @ \@temptokena \@ifclassloaded agu2001 natbib The agu2001 class already includes natbib coding, so you should not add it explicitly Type <Return> for now, but then later remove the command n...

[3] [3]

\@lbibitem[] @bibitem@first@sw\@secondoftwo \@lbibitem[#1]#2 \@extra@b@citeb \@ifundefined br@#2\@extra@b@citeb \@namedef br@#2 \@nameuse br@#2\@extra@b@citeb \@ifundefined b@#2\@extra@b@citeb @num @parse #2 @tmp #1 NAT@b@open@#2 NAT@b@shut@#2 \@ifnum @merge>\@ne @bibitem@first@sw \@firstoftwo \@ifundefined NAT@b*@#2 \@firstoftwo @num @NAT@ctr \@secondoft...

[4] [4]

Evaluating Inter-Column Logical Relationships in Synthetic Tabular Data Generation

@open @close @open @close and [1] URL: #1 \@ifundefined chapter * \@mkboth \@ifxundefined @sectionbib * \@mkboth * \@mkboth\@gobbletwo \@ifclassloaded amsart * \@ifclassloaded amsbook * \@ifxundefined @heading @heading NAT@ctr thebibliography [1] @ \@biblabel @NAT@ctr \@bibsetup #1 @NAT@ctr @ @openbib .11em \@plus.33em \@minus.07em 4000 4000 `\.\@m @bibit...

work page internal anchor Pith review doi:10.48550/arxiv.2502.04055 2023