Verdict
Highly suspicious. Multiple independent integrity indicators are present, including systematic citation mislabeling, implausible tabular results, and poor typesetting quality inconsistent with a finalized ACM MM '25 paper.
Key findings
- Systematic citation mislabeling (Finding 1, High severity): In Section 2.1, the paper attributes Med-UniC to reference [24], but [24] corresponds to the SAT paper ("Improving medical vision-language contrastive pretraining with semantics-aware triage"); the actual Med-UniC entry is [34]. In Sections 5.2 and 5.4.2, MGCA is attributed to [12], which is He et al.'s MAE; the correct MGCA entry is [35]. The pattern indicates careless citation handling rather than isolated typos.
- Implausible zero-gain in Table 1 (Finding 2, High severity): On CheXpert AUC, MRM shows 88.5 at both 1% and 10% training data, and the proposed "Ours" model shows 89.5 at both 1% and 10% training data, rising only at 100% (88.7 and 89.7 respectively). A 10× increase in labeled medical imaging data yielding a 0.00 absolute AUC change, repeated identically across two distinct models, is highly anomalous and suggestive of copy-paste tabulation.
- Unicode rendering artifacts in Section 3.1 (Finding 3, Medium severity): Strings such as
���and�appear undecoded. Such entities are uncharacteristic of a final accepted ACM publication and indicate either LaTeX-to-HTML/PDF conversion errors or sloppy paste operations during manuscript preparation. - Run-on numeric values in Table 1 (Finding 4, Medium severity): ViT-based rows contain concatenated values (e.g.,
82.784.4,90.891.993.193.780.389.594.5,89.589.792.1), with multiple RSNA and COVIDx entries merged per cell. This is inconsistent with rigorous academic typesetting. - Direct cross-reference: [24] → SAT; [12] → MAE; [35] → MGCA, contradicting in-text attribution.
- Table 1 CheXpert AUC, MRM: 88.5 (1%) vs 88.5 (10%) vs 88.7 (100%).
- Table 1 CheXpert AUC, Ours: 89.5 (1%) vs 89.5 (10%) vs 89.7 (100%).
- Encoded entities
�,�,�,�present in Section 3.1. - Concatenated cell values in Table 1 ViT-based rows as listed above.
- DOI: 10.1145/3746027.3755336.
- Confidence is high for Findings 1, 2, and 4 as they are verifiable from the published text. Finding 3 is also directly observable but may in principle be reproduced by certain rendering pipelines; however, its presence in a camera-ready venue is unusual.
- Authors have not been contacted, and independent raw training logs were not obtained. Statistical verification of the zero-gain claim (e.g., bootstrapped confidence intervals) was not performed.
- Findings 1 and 2 together are the strongest integrity indicators; Findings 3 and 4 are corroborating concerns rather than conclusive evidence of misconduct on their own.
- The report is AI-assisted and intended for academic discussion only; formal adjudication requires institutional investigation.