Investigation Report: Suspicious Duplicate Baseline Data in 'Pathology-Aware Reconstruction with Discriminative Knowledge Boosting Alignment for CheXray Vision-Language Pre-training'

Academic fraud report · Geng Detector

Summary

This report assesses a 2025 ACM MM '25 paper by Lihong Qiao et al. (DOI: 10.1145/3746027.3755336) for potential data fabrication. The primary finding is that in Table 1, the baseline methods GLoRIA and MGCA report identical numerical results across all 12 metric columns (4 datasets × 3 data ratios) under both CNN-based (ResNet-50) and ViT-based (ViT-B/16) backbone categories. Identical values to one decimal place across structurally different architectures on multiple datasets are implausible and strongly suggest data copying rather than independent experimentation. Secondary issues include layout artifacts consistent with bulk copy-pasting (run-together numbers in MRM and Med-UniC* rows) and a logical inconsistency in how Med-UniC is categorized under both architecture groups with different reported numbers. Verdict: strong indicators of academic misconduct (data fabrication/reuse); high confidence in the duplicate-data observation, lower confidence in inferring intent. Final determination requires institutional investigation.

Verdict

🔴 Strong indicators of academic misconduct. The duplicate reporting of GLoRIA and MGCA results under incompatible backbone architectures constitutes a near-conclusive sign of data fabrication or careless reuse. An institutional investigation and Erratum/retraction review are warranted.

Key findings

Identical baseline values across architectures: In Table 1, GLoRIA [14] and MGCA [35] report byte-for-byte identical 12-value result vectors under both the CNN-based (ResNet-50) and ViT-based (ViT-B/16) categories, which is physically implausible.
Layout artifacts indicating copy-paste: Numbers in ViT-based rows for MRM [46] and Med-UniC* are run together (e.g., 82.784.488.5, 90.891.993.780.389.594.5, 68.379.576.077.678.2), suggesting bulk text manipulation without verification.
Architectural classification inconsistency: Med-UniC [34] appears under both CNN-based and ViT-based with differing values (e.g., 88.2 vs 89.4 for 1% CheXpert), raising methodological questions about the comparison framework.
Scope of anomaly: Anomalies span 4 datasets (ChestX-ray14, CheXpert, RSNA, COVIDx) and 3 data ratios (1%, 10%, 100%), totaling 12 metric columns per affected method.

Evidence highlights

GLoRIA (Table 1): Reported vector (77.0, 81.9, 83.8, 86.5, 87.8, 88.2, 87.2, 88.1, 88.9, 75.8, 88.7, 92.1) is identical under both CNN-based and ViT-based rows.
MGCA (Table 1): Reported vector (78.7, 82.7, 84.1, 88.8, 89.1, 89.7, 89.1, 89.9, 90.8, 76.3, 89.0, 92.5) is identical under both CNN-based and ViT-based rows.
Table 1 caption states: "A standard ResNet-50 backbone is denoted as CNN-based, while a standard ViT/B-16 backbone is denoted as ViT-based," making identical results across the two architecturally incompatible backbones implausible.
Med-UniC [34]: Listed separately in both categories with non-identical values (CNN 1% CheXpert = 88.2; ViT 1% CheXpert = 89.4), demonstrating that the authors did run separate numbers for this method—undermining any defense that the duplicates for GLoRIA/MGCA were a reporting convention.
DOI: 10.1145/3746027.3755336 (ACM MM '25, October 27–31, 2025).

Notes

The duplicate-value finding is high-confidence and reproducible directly from the PDF text extraction.
The layout anomalies (run-together digits) are consistent with but not by themselves proof of misconduct; they strengthen the suspicion when combined with Finding 1.
Intent (fabrication vs. clerical error) cannot be determined from the report alone. Possibilities include (a) wholesale copying from a prior paper's Table without re-running experiments on the new backbone, or (b) placeholder values never replaced with actual measurements.
Authors are affiliated with institutions including Chongqing University of Posts and Telecommunications; an institutional inquiry would be the appropriate escalation path.
This automated assessment should be treated as evidentiary support, not a final determination of academic misconduct.