Geng Report on 'Pathology-Aware Reconstruction with Discriminative Knowledge Boosting Alignment for CheXray Vision-Language Pre-training' (DOI: 10.1145/3746027.3755336)

Academic fraud report · Geng Detector

Summary

The Geng misconduct report flags the above ACM MM '25 paper by Lihong Qiao, Shiyi Gao, Yucheng Shu, Bin Xiao, Weisheng Li, Xinbo Gao as highly suspicious, though no image-based checks could be performed. The key issue is an implausible data anomaly: in Table 1 on the CXP/CheXpert dataset, the authors' method reports an identical AUC of 89.5 under both the 1% and 10% training-data regimes for ViT-based linear probing, a coincidence that is effectively impossible in practice. The second major finding is a direct logical contradiction between Table 4 and the ablation discussion in Section 5.3.1: adding the supposedly core PAR module on top of DKBA lowers the ChestX-ray14 AUC from 81.0 to 80.8, yet the text claims this 'validates the complementary benefits', i.e., a negative transfer is being rebranded as complementary. Additional formatting/editorial red flags (collapsed rows in Table 1 and a truncated, lower-case 't. we compute...' passage in Section 3.2) reinforce concerns of hasty fabrication. Image reuse checks were not possible due to lack of figure data.

Verdict

🟠 Highly suspicious. Two confirmable data/logic issues and editorial red flags, but image forensics not performed due to missing figures.

Key findings

Implausible identical AUC values: In Table 1 (CXP/CheXpert, ViT-based linear probing), the proposed method reports exactly 89.5 for both 1% and 10% training data—identical to two decimal places under highly different data regimes.
Contradictory ablation claim: In Table 4 (ChestX-ray14, AUC), adding PAR on top of DKBA decreases the score from 81.0 to 80.8, while Section 5.3.1 asserts this 'validates the complementary benefits' of the two modules.
Formatting/splicing irregularities in Table 1 (e.g., run-together digits '90.891.993.193.780.389.594.5') suggest careless copy-paste construction.
Section 3.2 contains a truncated sentence ('global text representations t. we compute the cosine similarities...') that reads as spliced-together text.
Image reuse assessment (Figures 1–6, including t-SNE visualization checks) could not be executed because only textual content was available.

Evidence highlights

DOI: 10.1145/3746027.3755336 (ACM International Conference on Multimedia, MM '25, expected October 2025).
Authors: Lihong Qiao, Shiyi Gao, Yucheng Shu, Bin Xiao, Weisheng Li, Xinbo Gao.
Table 1, CXP column, ViT-based row: 89.5 at 1% data = 89.5 at 10% data.
Table 4, ChestX-ray14 AUC column: DKBA-only (× ✓) = 81.0; PAR + DKBA (✓ ✓) = 80.8.
Section 5.3.1 text claims 'complementary benefits' for a configuration that lowers the headline metric.

Notes

Findings 1, 2, and 3 are verified against the provided text (复核状态: ✅ 成立).
Finding 4 (image reuse) is evidence-insufficient (复核状态: ⚠️ 依据不足); results for Figures 1–6 are not adjudicated.
Report is AI-assisted; formal misconduct determination requires institutional investigation. Misclassification is possible.