Academic Fraud Investigation Report: "Countering Model Collapse in Iterative Self-Training via Dynamic Center-Edge Sampling" (DOI: 10.3390/electronics15040869)

Academic fraud report · Geng Detector

Summary

This investigation report assesses the paper published in Electronics (MDPI, DOI: 10.3390/electronics15040869) by Bingze Zhu and Yubo Xie, and concludes with a high-confidence finding of substantiated data fabrication and methodological irregularities. The central piece of evidence comes from Table 7 (ablation studies on OPT-125M): the full DCES model and its variant without dynamic sampling yield identical Expected Calibration Error (ECE) values of 0.0447, despite showing materially different perplexity and entropy scores. Identical ECE to four decimal places across two statistically distinct models is effectively impossible under continuous-variable calibration metrics, strongly suggesting copy-paste fabrication. Additional concerns include malformed references showing AI-hallucination signatures, an implausibly short 25-day submission-to-acceptance timeline combined with insufficient GPU resources (a single RTX 4090 plus 2× A800) for 10-round iterative self-training on five LLMs including a 6.7B-parameter model, and asymmetric baseline comparisons in Table 5 where two baselines are silently dropped for OPT-6.7B. Confidence in the ECE-based finding is high; the timeline and resource claims are circumstantial. Limits: definitive intent requires raw-data audit and institutional investigation.

Verdict

🔴 Substantiated (实锤) — Strong evidence of data fabrication and multiple methodological irregularities warrant formal institutional review.

Key findings

Table 7 data fabrication (critical): In the OPT-125M ablation, the full DCES model and the variant without dynamic sampling produce identical ECE = 0.0447 despite materially different PPL (29.7028 vs. 32.4577 on Wiki) and Entropy (2.6860 vs. 2.7040). For a continuous statistical expectation metric, exact four-decimal identity across two distinct models is statistically implausible and is characteristic of a copy-paste error where the ECE field was duplicated and not updated.
Reference list corruption (critical): Several references show LLM-hallucination signatures:
Ref [2] begins with "Makes Models Forget. Nature 2024..." — no authors, truncated title (likely refers to Shumailov et al., "AI model collapse...").
Ref [8] similarly lacks authors and begins with "Model Collapse Inevitable?...".
Ref [9] is reduced to arXiv2022, arXiv:2206.05802. [CrossRef] — title, authors, and venue all missing.
Implausible timeline and compute footprint: Submission-to-acceptance in only 25 days (Received 20 January 2026 / Accepted 14 February 2026), while claiming to perform 10 iterative self-training rounds across five LLMs (including OPT-6.7B) with only "a single NVIDIA RTX 4090 GPU" and "2× NVIDIA A800 GPUs". The compute budget is difficult to reconcile with the described pipeline (synthetic data generation, K-means clustering with K=8, MiniLM encoding, full fine-tuning).
Asymmetric baselines in Table 5: OPT-125M, GPT-2, and Qwen3 are compared against four methods, but OPT-6.7B is compared against only three, with PPL Filtering and SemDeDup silently dropped. The in-text justification ("PPL filtering performs poorly") does not justify removing baselines on the most compute-intensive configuration, raising the possibility that those results were not obtained.
Performative qualitative examples: Table 6 showcases baseline degeneration as repeated "1.1.1.1.1.1.1.1.1.1...", which, while plausible under model collapse, appears exaggerated and curated to dramatize the paper's thesis.

Evidence highlights

Table 7, OPT-125M ablation:
Full DCES: PPL (Wiki) = 29.7028, Entropy = 2.6860, ECE = 0.0447
DCES w/o Dynamic Sampling: PPL (Wiki) = 32.4577, Entropy = 2.7040, ECE = 0.0447
Discrepancy: two non-identical distributions with identical ECE to 4 decimals.
Reference anomalies: Refs [2], [8], [9] lack author/title metadata consistent with LLM-hallucinated references.
Publication metadata: Received 20 January 2026, Accepted 14 February 2026 (25 days).
Compute claim: "a single NVIDIA RTX 4090 GPU" and "2× NVIDIA A800 GPUs" supporting OPT-6.7B with 10 iterative rounds plus four other models.
Table 5 (OPT-6.7B): Only 3 baselines reported versus 4 for other model sizes; PPL Filtering and SemDeDup omitted.

Notes

DOI: 10.3390/electronics15040869
Journal: Electronics (MDPI)
Year: 2026
Authors: Bingze Zhu, Yubo Xie (Shanghai Maritime University, per report)
The ECE duplication is the strongest quantitative signal; raw experiment logs would be dispositive.
Reference anomalies are consistent with but not exclusive to LLM-assisted writing.
Timeline and compute concerns are circumstantial and depend on undisclosed runtime details.
Recommended actions include requesting raw data, posting on PubPeer, contacting the journal editorial office, and notifying the authors' institutional research integrity committee.