Geng Report: Suspicious References and Implausible Convergence Values in 'Zero-Shot Industrial Anomaly Detection via CLIP-DINOv2 Multimodal Fusion and Stabilized Attention Pooling' (Electronics, MDPI, 2025)

Academic fraud report · Geng Detector

Summary

This report flags a MDPI Electronics 2025 paper (DOI 10.3390/electronics14244785) as highly suspicious based on a Chinese 'Geng' academic-fraud screening. The most serious issue is a citation timeline paradox: although the manuscript was received in October 2025, four references in the bibliography cite journal articles published in 2026 with specific volume and issue numbers (Ref [1], [3], [6], [7] in Expert Systems with Applications and Advanced Engineering Informatics), which is highly atypical and suggests fabricated citations, preprint mislabeling, or coercive citation exchange. A secondary concern is the implausible convergence of learnable weights α=0.5, β=0.5, γ=1.0, which is unrealistic for SGD/Adam training. The report also notes marginal AUROC gains (often 0.1-0.2%) being labeled 'significant improvements', and the absence of any reported standard deviations across all benchmark tables. Limits: no pixel-level image analysis was performed, and the future-dated references require independent bibliographic verification. Overall verdict: highly suspicious, pending confirmation.

Verdict

🟠 Highly Suspicious. Multiple independent red flags were identified, with the citation timeline paradox being the most serious and potentially indicative of academic misconduct.

Key findings

Citation timeline paradox (Red): Four references list journal articles dated 2026 with specific volume and article numbers, while the paper was received on 28 October 2025 and published on 5 December 2025. This pattern is inconsistent with normal 'Online First' practice given the specificity of the cited volume/issue numbers.
Ref [1]: Chen, S.Z. et al., *Expert Syst. Appl.* 2026, *297*, 129489.
Ref [3]: Sheng, F.Q. et al., *Expert Syst. Appl.* 2026, *296*, 128965.
Ref [6]: Wang, E.R. et al., *Expert Syst. Appl.* 2026, *296*, 128995.
Ref [7]: Li, Y. et al., *Adv. Eng. Inform.* 2026, *69* Pt B, 103931.
Implausible convergence of learnable weights (Orange): In Section 3.3, the authors state that after training the model weights converge exactly to α=0.5, β=0.5, γ=1.0. Such clean values are not typical outputs of stochastic gradient-based optimization and raise concerns about whether these values were fabricated post hoc or whether the parameters were actually fixed rather than learnable.
Marginal performance gains mischaracterized as SOTA (Yellow): Across Tables 1–6, the proposed method's reported gains over the next-best baseline (Crane) are frequently in the 0.1–0.2% range (e.g., +0.1, +0.9, +3.1, +0.1, +0.2, +0.1, +1.2 in Table 1 for Image-AUROC). On the KSDD dataset (Table 4), the proposed method reports 99.0 in Pixel-AUROC, which is below Crane's reported 99.2 — a regression despite 'remarkable performance' language.
Absent statistical reporting (Yellow): None of the experimental tables (Tables 1–7) report standard deviations or confidence intervals, which is particularly concerning when the claimed improvements fall within typical noise margins.
Image analysis limited (informational): Visualizations in Figures 9–12 could not be pixel-level inspected for signs of duplication or splicing within this review.

Evidence highlights

Received date: 28 October 2025.
Published date: 5 December 2025 (Electronics, MDPI).
Direct quotation from the paper: "After training on the dataset, the model's weight parameters converge to: α = 0.5, β = 0.5, γ = 1.0."
Table 1 Image-AUROC deltas vs. Crane: +0.1, +0.9, +3.1, +0.1, +0.2, +0.1, +1.2 across seven datasets.
Table 4 Pixel-AUROC on KSDD: proposed = 99.0, Crane = 99.2 (proposed is lower).
DOI of the suspect paper: 10.3390/electronics14244785.

Notes

The future-dated citations are the strongest single signal and should be the priority for any follow-up investigation via Web of Science, Scopus, or direct queries to the cited journals (Elsevier's *Expert Systems with Applications* and *Advanced Engineering Informatics*).
Possible alternative explanations that should be ruled out before a misconduct finding include (a) preprint mislabeling in the typesetting stage, or (b) legitimate advance-access articles whose 2026 volume/issue metadata was assigned retroactively — neither is plausible for four simultaneous, fully-bibliographed entries with specific article IDs.
Convergence values of 0.5/1.0 may alternatively reflect rounding of underlying numbers; the authors should be asked to release training logs and random seeds.
Without variance estimates, all SOTA claims in this paper should be treated as preliminary.
Image-level forensic analysis was not within the scope of this automated screening; recommend manual review if suspicions escalate.