Review of 'Towards Interpretable User Intent Analysis with Deficient Evidence Fusion for Pseudo-Modalities' (ICMR 2025): Identity Mismatch at Open-Source Repository

Academic fraud report · Geng Detector

Summary

This report evaluates the ICMR 2025 paper by Chaochen Wu, Guan Luo, and Meiyun Zuo (DOI: 10.1145/3731715.3733454). The overall verdict is 'questionable,' driven primarily by one unresolved discrepancy rather than confirmed misconduct. All traditional image-manipulation checks are inapplicable because the paper contains only architecture diagrams, bar charts, and text heatmaps; no biomedical micrographs are present, so pixel-level reuse or splicing cannot be assessed. Performance figures (e.g., 94.16% accuracy on BANKING with DET in Table 2) fall within normal NLP benchmarks. The runtime analysis in Table 6 (LIME 11849 s, Shapley 72444 s, DENet 56.81 s) plausibly reflects real perturbation-based explainability costs, supporting authenticity of the experiments. The timeline is internally consistent. The sole red flag is that the open-source repository github.com/yuanxiaoheben/DENet does not match the listed authors' pinyin names. This may indicate a borrowed account, an unlisted contributor, or an unrelated personal handle. Confidence is moderate; no fabrication is confirmed, and conclusions remain provisional pending author clarification and code reproducibility checks.

Verdict

Questionable (存疑). No fabrication is confirmed. The paper is clean across image, timeline, and statistical-experiment dimensions. One identity-related anomaly warrants author clarification.

Key findings

Open-source repository identity mismatch (minor red flag): The linked GitHub account yuanxiaoheben does not correspond to any of the listed authors' pinyin names (Chaochen Wu, Guan Luo, Meiyun Zuo). This could indicate a borrowed account, an unacknowledged contributor (potential ghost authorship), or a non-related personal handle.
Image integrity: not applicable. Figures 2, 5, and 6 are architecture diagrams, bar plots, and text heatmaps; no biomedical micrographs are present, so traditional pixel-level duplication or splicing checks cannot be performed.
Timeline consistency: The paper (ICMR 2025) cites the authors' own 2024 work [3]; no anachronistic citations or impossible temporal claims were found.
Plausible experimental data: Classification accuracies and F1 scores in Table 2 (e.g., BANKING / DET = 94.16%) lie within reasonable NLP ranges; no suspiciously perfect 100% results.
Runtime figures support experimental authenticity: Table 6 reports LIME = 11849 s, Shapley values = 72444 s, and DENet = 56.81 s. The >1000× gap is consistent with real perturbation-based explainability overhead on long text and argues against fabricated runtimes.

Evidence highlights

DOI: 10.1145/3731715.3733454
Affiliations: Chaochen Wu and Meiyun Zuo — Renmin University of China;Guan Luo — Institute of Automation, Chinese Academy of Sciences.
Declared code URL: github.com/yuanxiaoheben/DENet — no pinyin match to any author name.
Table 6 runtime values cited above (preserved exactly as reported).
Self-citation [3]: Cao, Wu, et al., 2024 — temporally consistent.

Notes

This is a low-confidence signal. Recommended follow-ups: (1) contact authors to clarify ownership of the GitHub account; (2) clone the repository and attempt to reproduce Table 1/Table 2 F1 scores, with attention to possible data leakage; (3) optionally raise the explainability claims on PubPeer. Limitations: AI-assisted analysis cannot verify code behavior, authorship contributions, or external reuse of figures absent in the manuscript. Any final determination of misconduct requires institutional investigation.