Geng Report: LEAF-Mamba — Local Emphatic and Adaptive Fusion State Space Model for RGB-D Salient Object Detection (DOI: 10.1145/3746027.3754863)

Academic fraud report · Geng Detector

Summary

This Geng academic-integrity review targets the ACM MM '25 paper 'LEAF-Mamba' (Wu, Gao, Fei, Lee, Hsu; DOI: 10.1145/3746027.3754863). The overarching verdict is 'highly suspicious' (orange). The most substantive and verified finding concerns Section 4.3 (Design of AFM, Page 6), where the textual ablation discussion appears to be systematically mismatched with Table 4. Recomputation against Table 4 shows that the 1.41%/4.54% gains attributed in prose to SIM on SSD correspond to CSoP; the 0.82%/2.27% gains attributed to SEM correspond to SIM; and the 1.88%/6.81% joint-attribution figure corresponds to CSoP+SIM rather than SIM+SEM. This indicates either copy-paste errors during drafting or fabricated textual analysis. A second concerning pattern involves implausibly strong results (18.1 GFLOPs, 70.2 FPS at 84.5M parameters) reported alongside SOTA accuracy, but this remains suggestive only, lacking decisive evidence of tampering. Qualitative figures showed no obvious physical manipulation. Confidence is high for the data-descriptor mismatch and low-to-moderate for the performance-implausibility concern.

Verdict

Highly suspicious (orange). The report flags confirmed internal inconsistency between Section 4.3 prose and Table 4 in the LEAF-Mamba paper (DOI: 10.1145/3746027.3754863). One additional claim regarding implausibly strong efficiency-accuracy trade-off is suggestive but unverified.

Key findings

Data misattribution between text and Table 4 (severity: red). Section 4.3 attributes specific SSD-dataset improvements to SIM, SEM, and their combination, but recomputed values map onto different configurations.
Implausibly strong overall efficiency-accuracy profile (severity: orange, unverified). Reported 18.1 GFLOPs and 70.2 FPS at 84.5M parameters substantially exceed comparable Transformer baselines (e.g., VST++ at 12 FPS, DCT at 12 FPS, CAVER at 21.3 FPS), while simultaneously leading accuracy.
No physical image manipulation detected (severity: green). Figure 5 and Figure 6 qualitative comparisons appear naturally rendered with no obvious copy-rotate splicing.

Evidence highlights

Section 4.3 / Page 6 prose claim (verified against Table 4):
Text: "independently applying the SIM (or SEM) achieves performance boost with 1.41% (0.82%) and 4.54% (2.27%) on the SSD dataset in terms of F_β and MAE."
Recomputation from Table 4: the 1.41%/4.54% deltas correspond to CSoP vs. baseline: (0.863−0.851)/0.851 ≈ 1.41%, (0.044−0.042)/0.044 ≈ 4.54%.
The 0.82%/2.27% deltas correspond to SIM vs. baseline: (0.858−0.851)/0.851 ≈ 0.82%, (0.044−0.043)/0.044 ≈ 2.27%.
Text claim of combined SIM+SEM at 1.88%/6.81% matches CSoP+SIM in Table 4, not SIM+SEM.
Efficiency comparison (Table 1 / Abstract): Reported 18.1 GFLOPs, 70.2 FPS, 84.5M parameters; comparison baselines include VST++ (12 FPS), DCT (12 FPS), CAVER (21.3 FPS). Numerical values are reproduced from the paper; the inference of impropriety rests on engineering plausibility, not direct tampering evidence.
Image inspection: Figure 5 and Figure 6 show consistent background textures and methodologically plausible mask morphology; no splicing artifacts observed.

Notes

DOI: 10.1145/3746027.3754863 (ACM International Conference on Multimedia, MM '25).
The data-descriptor mismatch is the strongest evidence in this report and is independently reproducible from the reported Table 4 values; it does not depend on subjective judgment.
The efficiency/accuracy anomaly is flagged as suggestive only; confirmation requires author-supplied training logs, timing scripts (batch size, input resolution, GPU model), and independent code release.
Recommended follow-ups: request raw logs and timing scripts from authors; request clarification of Section 4.3 numerical attributions; post a PubPeer comment documenting the table-text inconsistency; defer citation until community code reproduction.
This report is AI-assisted and intended for academic-discussion purposes; final determinations require institutional investigation.