Geng Investigation Report: Zero-Shot Industrial Anomaly Detection via CLIP-DINOv2 Multimodal Fusion and Stabilized Attention Pooling (Electronics 2025, 14, 4785; DOI: 10.3390/electronics14244785)

Academic fraud report · Geng Detector

Summary

Highly suspicious. The manuscript, authored by Junjie Jiang and colleagues and published in MDPI Electronics on 5 December 2025, demonstrates internally consistent numerical computation and contains one counter-honest metric (Pixel-AUPRO on MVTec: 87.0 vs. Crane's 88.1), which paradoxically lends some credibility to the experimental run. However, the report contains a lethal timeline contradiction: four references are dated 2026 (Expert Syst. Appl. 297, 129489; 296, 128965; 296, 128995; Adv. Eng. Inform. 69 Pt B, 103931), which cannot exist at a December 2025 publication date. The revise-to-accept window is only 2 days (24 Nov to 26 Nov 2025), and all performance tables report single-point values without standard deviations or confidence intervals. No image-level analysis was possible because raw figures were not provided. Overall confidence in the timeline anomaly is high; statistical and process concerns are moderate and typical of low-rigor venues, not definitive fraud markers.

Verdict

🟠 Highly suspicious. The paper's internal numerical logic is consistent and exhibits one non-winning metric, suggesting the experimental runs are likely real. However, the presence of references dated 2026 in a paper published in December 2025 constitutes a hard, objective anomaly that requires explanation. The combination of this timeline error with a 2-day revise-to-accept window and a 2025 grant identifier (LBMHZ25F030002) raises serious concerns about manuscript provenance, even if it does not prove outright fabrication.

Key findings

🔴 Time-traveling citations (refutation-level): Four references in the bibliography carry 2026 volume/issue/page numbers despite a publication date of 5 December 2025.
🟠 Implausibly fast review cycle: Received 28 October 2025, Revised 24 November 2025, Accepted 26 November 2025 (2 days from revision to acceptance).
🟡 Missing statistical reporting: Tables 1–7 report single-point values (e.g., AUROC 94.8 vs. 93.9) with no standard deviations, confidence intervals, or seed counts.
🟢 Counter-honest metric (paradoxical credibility signal): In Table 5, the proposed method scores 87.0 on MVTec Pixel-AUPRO, below the baseline Crane at 88.1 — inconsistent with a fully fabricated "all-wins" table.
🟡 Grant timeline inconsistency: Funding identifier LBMHZ25F030002 appears in a December 2025 paper, implying near-zero gestation from funding award to publication.

Evidence highlights

Reference 1: Expert Syst. Appl. 2026, 297, 129489. — 2026 volume and article number cited in a 2025 paper.
Reference 3: Expert Syst. Appl. 2026, 296, 128965.
Reference 6: Expert Syst. Appl. 2026, 296, 128995.
Reference 7: Adv. Eng. Inform. 2026, 69 Pt B, 103931.
Timeline data: Received 28 Oct 2025 → Revised 24 Nov 2025 → Accepted 26 Nov 2025 → Published 5 Dec 2025.
Performance numbers: Table 5 Pixel-AUPRO MVTec — Crane 88.1, Ours 87.0 (proposed method loses on this metric).
Funding: Project number LBMHZ25F030002 listed in Acknowledgments.
DOI: 10.3390/electronics14244785

Notes

No raw images were available for pixel-level duplication or PS-trace analysis; image-based fraud checks were not performed.
The 2026-dated references are the report's strongest evidence; however, a benign (if unlikely) explanation is that the authors inadvertently copy-pasted "in-press" entries with publisher-issued 2026 volume numbers that were not yet finalized. This still reflects poor editorial vetting.
The 2-day revision-to-acceptance interval is unusual but not unprecedented for MDPI's triage-style review and is here treated as a process concern, not proof of misconduct.
The presence of a single losing metric is a weak positive indicator of genuine experimentation; it is not a defense against the reference-timeline anomaly.
Final determination of academic misconduct requires institutional investigation; this report is an AI-assisted screening summary for academic-integrity discussion only.