Geng Report: Pathology-Aware Reconstruction with Discriminative Knowledge Boosting Alignment for CheXray Vision-Language Pre-training

Academic fraud report · Geng Detector

Summary

This Geng-style fraud detection report alleges severe data fabrication in a 2025 ACM Multimedia paper (DOI: 10.1145/3746027.3755336) on chest X-ray vision-language pre-training. The verdict is red (substantiated). The most damning finding concerns the CheXpert (CXP) results in Table 1, where the authors' model outperforms MRM by an identical +1.0 AUC across all three training-data ratios (1%, 10%, 100%), with MRM's 1% and 10% AUCs being equal (88.5) — a pattern inconsistent with real experimentation and consistent with manually adding a constant to baseline numbers. A second issue highlights a contradiction between Table 1 and the ablation in Table 4, where combining PAR and DKBA drops AUC from 81.0 to 80.8 on ChestX-ray14, yet the main text claims complementarity. The RSNA comparison also shows suspiciously precise marginal wins (+0.1 to +0.2) over the stronger Med-UniC baseline across all three ratios. Minor issues include an ill-formed Equation 9 and the typo "abesent". Confidence is high for the CheXpert arithmetic finding; the ablation inconsistency is clear but interpretation (intentional vs. reporting bias) remains partially inferential.

Verdict

🔴 Substantiated (实锤). The report identifies a strong quantitative inconsistency in Table 1 (CXP) and a self-contradictory ablation. The CheXpert pattern is highly unlikely under genuine experimental conditions and is consistent with manual data manipulation.

Key findings

Mechanical data fabrication in Table 1 (CheXpert column). The reported delta over MRM is +1.0 AUC at 1%, 10%, and 100% training ratios. MRM's 1% and 10% AUCs are both reported as 88.5, and Ours' are both 89.5. Three independent training budgets yielding bit-identical scores, with a constant offset to the baseline, is consistent with post-hoc arithmetic rather than measured performance.
Ablation vs. main-table contradiction (Table 4 vs. Table 1). On ChestX-ray14, the DKBA-only configuration reaches 81.0 AUC, but adding PAR (yielding the full "Ours" model) drops AUC to 80.8. The main text nevertheless asserts complementarity between the two modules.
Suspiciously precise wins over Med-UniC on RSNA. Reported deltas are +0.2 (1%), +0.1 (10%), and +0.1 (100%) over a stronger model that uses additional data; the margins are uniform across very different training ratios and configurations.
Equation 9 notational error. h_i = FFN(LN(MCA(LN(Q,K,V)))) applies a single LayerNorm over Q, K, V whose dimensions need not match, which is not standard practice and is mathematically ill-posed when Q and K have different shapes.
Typo "abesent" in Section 3.1 ("We remove entities classified as abesent."), indicating careless writing/review.

Evidence highlights

Table 1, CheXpert (CXP) AUC:
1% — MRM 88.5, Ours 89.5 (Δ +1.0)
10% — MRM 88.5, Ours 89.5 (Δ +1.0)
100% — MRM 88.7, Ours 89.7 (Δ +1.0)
Table 4, ChestX-ray14: DKBA-only AUC 81.0; PAR + DKBA ("Ours") AUC 80.8 (−0.2 vs. DKBA alone).
Table 1, RSNA: Med-UniC 91.9 / 93.1 / 93.7 vs. Ours 92.1 / 93.2 / 93.8 (Δ +0.2 / +0.1 / +0.1).
Equation 9: h_i = FFN(LN(MCA(LN(Q,K,V)))); text contains "abesent" instead of "absent".
DOI of the target paper: 10.1145/3746027.3755336.

Notes

The CheXpert +1.0 pattern and the 1%/10% tie constitute the strongest quantitative evidence; under realistic stochasticity these equalities are extremely improbable.
The RSNA comparison is suggestive but not conclusive — small uniform margins can also arise from optimistic cherry-picking of runs/checkpoints.
The ablation drop (80.8 < 81.0) is a genuine contradiction with the text's qualitative claim, but does not by itself prove fabrication; it could reflect reporting/selection bias.
No image-manipulation evidence is presented in this report; concerns are limited to numerical reporting and writing quality.
Authors were not contacted for comment within this report; recommended actions include requesting raw logs/training seeds and an institutional investigation.