Integrity Concern Report: "Context-aware contrastive learning via structural harmony preservation for generalized category discovery" (Knowledge-Based Systems, DOI 10.1016/j.knosys.2025.114942)

Academic fraud report · Geng Detector

Summary

This report examines a computer vision / machine learning paper published in Knowledge-Based Systems (2026, vol. 333, article 114942; DOI 10.1016/j.knosys.2025.114942) by Wenbo Hu, Yue Lu, Xinchen Ma, and Ching Y. Suen. The overall assessment is highly suspicious, driven primarily by internal consistency problems between text, tables, and figure captions rather than image or statistical forensics. Four verified findings: (1) the text claims "perfect estimation" of class numbers on CIFAR100, but Table 7 records an estimate of 95 against a ground truth of 100 (5% error); (2) the text claims a 7.2% improvement over the runner-up on CUB, but Table 4 shows a 5.6% gap over the second-best method PIS (70.1 vs 64.5); (3) Table 9 contains a likely typo (Devanagari NMI Old = 9.7, inconsistent with the same row's values and the CACL* row's 97.4); (4) the Fig. 7 caption and main text describe different methods being compared. Automated forensic tests (Benford analysis, copy-move detection) were reviewed and excluded as inapplicable or benignly explained. Limitations: these are data-reporting inconsistencies, not proof of fabrication; expert institutional review is required. Composite posterior: 86.8% (95% CI [51.1%, 97%], BF = 124.52), with the caveat that evidence independence assumptions may inflate the estimate.

Verdict

Highly suspicious (高度可疑). No biologically-derived image manipulation or statistical fabrication indicators were substantiated; however, four verified internal inconsistencies cluster in the paper's data-reporting and claims sections, all biased in the direction of overstating the method's performance. Composite Bayesian synthesis: posterior 86.8% (95% CI [51.1%, 97%]), BF = 124.52, 25 candidate signals excluded after review. This does not constitute a formal finding of academic misconduct; institutional review is needed.

Key findings

Finding 1 (🟠, verified): Text claims CACL achieves "perfect estimation" of class numbers on CIFAR100, but Table 7 records ground truth = 100 and estimate = 95 ("5 %" error). The conclusion — and the dependent "Analysis across Tables 1-5" discussion — overstates the method's capability.
Finding 2 (🟡, verified): Text claims a 7.2% improvement over the second-best method on CUB ("All"), but Table 4 shows CACL = 70.1 vs PIS = 64.5, a gap of only 5.6%. The 7.2% figure is only derivable using the third-ranked PromptCAL (62.9), effectively ignoring the stronger baseline PIS.
Finding 3 (🟡, verified): Table 9 (NMI), Devanagari row for CACL reads "91.8 9.7 84.7". An Old-class NMI of 9.7 is inconsistent with All = 91.8 and New = 84.7, and with the parallel CACL* row ("91.8 97.4 84.7"); 9.7 is almost certainly a digit-transposition typo of 97.4, indicating poor proofreading.
Finding 4 (🟡, verified): The Fig. 7 caption states the embeddings were "learned by CACL and CACL* on the Devanagari dataset", while the main text ("Visualization of Embedding Spaces") states the comparison is "SimGCD and CACL". The actual content of the panels cannot be determined from the text.
Excluded signals (not counted): Benford first-digit test (χ²=565.5, p=0.0000) inapplicable to bounded accuracy percentages; last-digit test (χ²=23.5, p=0.0052) weakened by rounding conventions; near-zero column differences between CACL and CACL* are textually supported (0% estimation error on C-MNIST and Devanagari, consistent with Table 7); AI-text score 0.35/1.0 in low-confidence range; all 21 image copy-move/noise anomalies match benign explanations (32-pixel multiples consistent with multi-panel grid tiling in Figs. 2, 4–6). Identical CACL/CACL* values in Table 6 for C-MNIST and Devanagari are self-consistent with Table 7, not data reuse.

Evidence highlights

DOI: 10.1016/j.knosys.2025.114942; Knowledge-Based Systems 333 (2026) 114942; available online 22 November 2025.
Table 7: CIFAR100 GT = 100, CACL estimate = 95 (error 5%) vs. text claim of "perfect estimation".
Table 4: CACL 70.1 vs PIS 64.5 → actual margin 5.6%, not the claimed 7.2% (70.1 − 64.5 = 5.6 ≠ 7.2).
Table 9 (Devanagari, CACL): "91.8 9.7 84.7" vs CACL* row "91.8 97.4 84.7".
Fig. 7 caption vs. main text disagree on which methods are visualized.
Aggregated evidence (automated, conditional-independence assumption may overstate correlated evidence; LR version geng_lr_v3, n=11, shrinkage k=10): posterior 86.8%, CI [51.1%, 97%], BF = 124.52; top contributors: Finding 1 (35.6%, logLR=1.59), Findings 2 and 4 (21.9% each, logLR=1.08), Finding 3 (20.6%, logLR=1.08).

Notes

All four findings were programmatically verified against the source PDF (1-s2.0-S095070512501980X-main.pdf) with quoted values confirmed present in the original text.
Individually, each finding could reflect careless error rather than fraud; however, their concentration in the claims/data-reporting sections — with each error flattering the method — warrants author clarification.
Recommended actions: contact authors regarding the Table 7/"perfect estimation" contradiction, the CUB margin calculation, the Table 9 value 9.7, and the Fig. 7 caption conflict; raise on PubPeer; if unresolved, submit a correction/concern to the journal editor. Institutional reporting is not indicated at the current evidence level.
Disclaimer: AI-assisted report for academic discussion only; final misconduct determinations require professional investigation; false positives and false negatives are possible.