Integrity Review: D²MoE: Dual Routing and Dynamic Scheduling for Efficient On-Device MoE-based LLM Serving (DOI: 10.1145/3680207.3723493)

Academic fraud report · Geng Detector

Summary

This review flags a paper accepted at ACM MobiCom '25 as highly suspicious due to anomalies in reporting and presentation, though no deliberate fabrication is conclusively proven. The most concerning finding is statistical duplication in Table 3 (Page 8): for LLaMA-MoE-3.5B, EdgeMoE and D²MoE-V1 both report an identical Winogrande score of 59.52; for Mixtral 8x7B, MoQE-DynaIO-INT8 and D²MoE-V2 both report 74.71, and EdgeMoE and D²MoE-V1 both report 75.52. Identical two-decimal scores across different methods are improbable and suggest copy-paste errors or a buggy evaluation pipeline caching stale outputs. Secondary issues include figure-quality problems: Figure 6 (Page 5) and Figure 9 (Page 7) appear to be PPT screenshots retaining presentation metadata ('Zicong Hong PEILabPresentation 2025/4/17 8' and '2025/3/21 9'). The reference list contains duplicated entries (Refs [31]/[32] both cite Mixture-of-Depths; Refs [22]/[23] both cite AWQ), and Table 1 shows a steep INT2→INT3 throughput drop (50.47→45.91 token/s). Verdict: highly suspicious; pending raw evaluation logs and author clarification.

Verdict

🟠 Highly suspicious. Multiple independent inconsistencies in data reporting and figure preparation warrant formal author clarification. No conclusion of intentional fabrication is reached, but the cumulative pattern—identical benchmark scores across different methods, PPT screenshots embedded as figures with leftover presentation metadata, duplicated bibliography entries—indicates insufficient quality control inconsistent with a top-tier venue such as ACM MobiCom.

Key findings

Identical Winogrande scores across distinct methods in Table 3 (Page 8). The likelihood of independent runs producing two-decimal identical results on a commonsense reasoning benchmark is vanishingly small.
LLaMA-MoE-3.5B: EdgeMoE = D²MoE-V1 = 59.52
Mixtral 8x7B: MoQE-DynaIO-INT8 = D²MoE-V2 = 74.71
Mixtral 8x7B: EdgeMoE = D²MoE-V1 = 75.52
Figure 6 (Page 5) and Figure 9 (Page 7) appear to be PowerPoint screenshots, not vector graphics. Embedded metadata strings remain visible:
"Zicong Hong PEILabPresentation 2025/4/17 8" preceding Figure 6
"Zicong Hong PEILabPresentation 2025/3/21 9" appearing four times around Figure 9
Duplicate reference entries in the bibliography (Pages 14–15):
Refs [31] and [32] both list *Mixture-of-Depths* (David Raposo et al., 2024).
Refs [22] and [23] both list *AWQ* (Ji Lin et al., 2024).
Table 1 (Page 3) throughput curve shows a steep drop from INT2 to INT3 (50.47 → 45.91 → 43.82 token/s for INT2/INT3/INT4 respectively). Plausible in isolation, but flagged for cross-verification alongside other anomalies.

Evidence highlights

| Finding | Location | Concrete evidence | |---|---|---| | Duplicated Winogrande scores | Table 3, Page 8 | 59.52, 74.71, 75.52 — identical across different quantization methods/models | | PPT watermark leakage | Figure 6, Page 5 | "Zicong Hong PEILabPresentation 2025/4/17 8" | | PPT watermark leakage | Figure 9, Page 7 | "Zicong Hong PEILabPresentation 2025/3/21 9" (×4 occurrences) | | Duplicate references | References, Pages 14–15 | Refs [31]=[32] (Mixture-of-Depths); Refs [22]=[23] (AWQ) | | Throughput anomaly | Table 1, Page 3 | INT2: 50.47 token/s; INT3: 45.91 token/s; INT4: 43.82 token/s | | DOI | — | 10.1145/3680207.3723493 |

Notes

Confidence. High confidence on the bibliographic duplications and the PPT-watermark text (these are deterministic textual facts). Moderate confidence on the copy-paste hypothesis for Table 3: identical two-decimal scores across distinct methods are statistically implausible under independent evaluation runs, but alternative benign explanations (cached results, post-hoc rounding to a shared baseline, copy-paste of a reference column) cannot be ruled out without raw evaluation logs.
Limits. This review did not access the source code, raw model outputs, or supplementary materials. Verification of the Winogrande anomaly definitively requires the original evaluation scripts and per-sample outputs. The reported throughput values in Table 1 are not, in isolation, indicative of misconduct.
Recommended actions. Authors should publish raw Winogrande per-sample results and evaluation scripts; provide vector-format replacements for Figures 6 and 9; deduplicate the bibliography. Venue reviewers/editors may wish to issue an erratum once underlying causes are confirmed.