Geng Integrity Report: "Congestion-Aware Floorplanning Optimization for Large-scale HLS Design on Multi-Die FPGA" (TechRxiv preprint, DOI: 10.36227/techrxiv.172565775.51088665/v2)

Academic fraud report · Geng Detector

Summary

This Geng report assesses the TechRxiv preprint by Teng Wang et al. (posted 2 December 2024) on congestion-aware floorplanning for multi-die FPGA high-level synthesis. The overall verdict is highly suspicious (orange). Key issues include severe formatting neglect (raw LaTeX template header retained on every page, untranslated Chinese fragments in figures/captions), heavily malformed BibTeX entries ([4] missing authors/title, [5][14][15] reduced to bare URLs, [6] truncated), and most notably an anomalous Table II where BRAM and DSP utilization percentages for AutoBridge and FrqBooster match to two decimal places across all eight design sizes (e.g., 8.57 vs 8.57 up to 67.81 vs 67.81). The reviewer argues such perfect alignment is implausible for two distinct place-and-route flows. Frequency-speedup figures (73.15%, 81.22%, 18.32%) and tooling references appear chronologically consistent. Image-level forensic analysis was not possible from the supplied text. Confidence is moderate: textual evidence is strong for sloppiness and suspicious table values, but pixel-level image reuse and data-fabrication claims remain unverified without original logs.

Verdict

🟠 Highly suspicious — Multiple indicators of extreme sloppiness and at least one substantive data-integrity concern (Table II). No direct fabrication proven, but the pattern warrants formal follow-up.

Key findings

Reference/BibTeX collapse — Several entries (e.g., [4], [5], [6], [14], [15]) show catastrophic field loss: truncated titles, missing authors, bare URLs. Suggests export errors from EndNote/Zotero never proofread.
Unmodified LaTeX template — Every page header still reads "JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021"; the default placeholder was not replaced.
Mixed-language artefacts — Figures (e.g., Figure 3) contain untranslated Chinese fragments such as 布局/布线 and 硬件核心 embedded in an English manuscript.
Suspicious Table II consistency — In the BRAM and DSP columns comparing AutoBridge vs. FrqBooster across eight configuration points (13×2 through 13×16), values are identical to two decimal places (e.g., DSP 8.57 vs. 8.57, …, 67.81 vs. 67.81; BRAM likewise). Two distinct physical implementation flows would not normally produce byte-identical utilization percentages.
Statistical sanity — Frequency improvement numbers (73.15%, 81.22%, 18.32%) lack Benford-like or last-digit-0/5 artifacts of RNG-generated data; tool (Vivado/Vitis 2019.02) and device (Xilinx U250/U280) choices are chronologically plausible; 2024 citations present in a 2024 preprint are consistent.
Image forensics — Not performed; the provided text extraction lacked full-resolution raster data. No pixel-level duplication, rotation, or splicing claims can be made from the supplied evidence.

Evidence highlights

DOI: 10.36227/techrxiv.172565775.51088665/v2 (preprint, posted 2 Dec 2024)
Table II exact-match pairs (AutoBridge vs. FrqBooster): DSP column identical across all 8 rows (8.57 … 67.81); BRAM column identical across all 8 rows.
Header on every page: "JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021".
Reference anomalies: [4] begins mid-sentence ("for high-level synthesis designs on multi-die fpgas,"); [5][14][15] are bare URLs (e.g., "en-US/ug872largefpga."); [6] truncated to "network and reinforcement learning-based floorplanning,".
Reported Fmax gains: 73.15%, 81.22%, 18.32% (no anomalous digit distribution).

Notes

Limitations: image-level analysis was not feasible from the provided text; final confirmation of Table II requires the authors' raw Vivado implementation logs.
Severity map: Formatting/BibTeX issues = 🟡; Table II consistency = 🟠; Statistics = ✅ clear; Image forensics = untested.
Recommended actions: (1) request original Vivado implementation logs from authors; (2) flag formatting and language-mixing issues on PubPeer; (3) require professional proofreading before any journal submission.
This report is AI-assisted and intended for academic discussion only; final misconduct determinations require institutional investigation.