Integrity Review Report: MCoT-MVS: Multi-level Vision Selection by Multi-modal Chain-of-Thought Reasoning for Composed Image Retrieval

Academic fraud report · Geng Detector

Summary

Verdict: No clear signs of academic fraud were detected. The paper (DOI: 10.1145/3774904.3792287, arXiv 2603.17360v1), submitted to The Web Conference 2026 (WWW '26), was screened across six dimensions: image reuse, data fabrication, image splicing, statistical anomalies, output/timeline anomalies, and citation/methodology concerns. Reported metrics in Tables 1-5 are internally consistent: for example, MCoT-MVS averages on CIRR yield 82.52, matching the paper's stated Avg; FashionIQ R@10/R@50 subgroup means (63.24, 82.01) are arithmetically correct; ablation claim that removing PVRS drops R@1 by 0.59% matches No.4 (54.50) minus No.6 (53.91). The timeline is plausible: arXiv posting (2026-03-18) precedes the WWW '26 conference (2026-04-13 to 2026-04-17), and cited models (Qwen2.5-VL-32B, Llama-3.2-11B-Vision) predate the work. The authors released code on GitHub. Limitations: this review relied on text only, so no pixel-level image analysis was performed. Final determination should consider independent code reproduction.

Verdict

No clear signs of academic fraud detected (clean).

Key findings

Internal numerical consistency of reported metrics verified by arithmetic cross-checking.
Ablation claims align precisely with tabulated values.
Timeline of model usage (Qwen2.5-VL-32B, Llama-3.2-11B-Vision) is consistent with their public release dates and the 2025-2026 experimental window.
Authors openly released code on GitHub, reducing concerns about reproducibility opacity.
Fixed random seed (124) and Recall-based evaluation protocol leave little room for p-hacking, consistent with norms in the CIR/FashionIQ retrieval literature.
No textual cues of image splicing, caption mismatch, or duplicate panels were found.

Evidence highlights

Table 1 (CIRR) — MCoT-MVS: R@1=55.33, R@5=84.75, R@10=91.45, R@50=98.55; computed mean = (55.33+84.75+91.45+98.55)/4 = 82.52, matching reported Avg=82.52. Baseline means (e.g., TIRG, CASE) also reconcile.
Table 2 (FashionIQ) — MCoT-MVS R@10 subgroup mean = (58.45+63.24+68.02)/3 ≈ 63.236 ≈ 63.24; R@50 subgroup mean = (78.92+81.15+85.97)/3 ≈ 82.013 ≈ 82.01.
Table 3 ablation — Claimed 0.59% drop when PVRS is removed is verified: 54.50 (No.4) − 53.91 (No.6) = 0.59.
Timeline — arXiv submission 2026-03-18 precedes WWW '26 dates 2026-04-13 to 2026-04-17.
Methodology — Algorithm 1 pseudocode (triplet → MLLM reasoning → CLIP feature extraction → Loss) matches the mathematical derivations in Section 3.

Notes

Pixel-level image forensics were not performed because only text inputs were available; Figure 1-3 captions and descriptions appear self-consistent but cannot be conclusively cleared at the pixel level.
DOI: 10.1145/3774904.3792287.
Confidence: moderate-to-high for text-based dimensions; low for image-piracy dimensions.
Recommended follow-up: independent reproduction via the released GitHub repository to confirm the tabulated metrics.