Integrity Review Report: Discovering faster matrix multiplication algorithms with reinforcement learning (Fawzi et al., Nature 2022) — DOI 10.1038/s41586-022-05172-4

Academic fraud report · Geng Detector

Summary

Verdict: No integrity concerns identified (cleared). This report reviews the DeepMind paper published in Nature (Vol 610, October 2022) that used AlphaZero-based reinforcement learning to discover faster matrix multiplication algorithms, including a reduction of 4×4 matrix multiplication from 49 to 47 scalar multiplications. Key checks performed: image reuse/duplication, data fabrication, image splicing, statistical anomalies, productivity anomalies, and citation/methodological consistency. No anomalies were found in any category. Figure content is dominated by vector schematics, algorithm flowcharts, and benchmark bar/line plots — not biological images susceptible to splicing. Reported hardware speedups (4.3%–23.9% on Nvidia V100 and Google TPU v2, with std < 0.4 percentage points across >200 runs) are consistent with realistic memory-bandwidth and cache behaviour. Timeline, compute footprint (64 TPU v3 + 1600 TPU v4 over ~1 week), and references to Strassen (1969), Laderman (1976), and Smirnov (2013) are internally consistent. Code was released on GitHub. Confidence is high; limitations include reliance on textual metadata rather than raw image forensic analysis, and the inherent subjectivity of the evaluation criteria.

Verdict

Cleared — no evidence of academic misconduct was identified. The paper meets standards for a high-quality algorithm-discovery study published in a top-tier venue, with open-source code and rigorous mathematical verification.

Key findings

No image reuse or duplication detected — figures are predominantly vector schematics, algorithm flowcharts, and benchmark plots; no biological/photographic images are present.
No data fabrication indicators — hardware speedup figures (4.3%–23.9%) with standard deviation < 0.4 percentage points across >200 runs are consistent with realistic hardware variance.
No image splicing or Photoshop artefacts — figures are vector-based and structurally clean.
No statistical anomalies — core claims are mathematical theorems provable by tensor decomposition; no p-value manipulation is possible.
No productivity anomalies — reported compute footprint (64 TPU v3 training cores, 1600 TPU v4 actor cores, ~1 week to convergence) is plausible for DeepMind in 2022.
No citation or methodological inconsistencies — references to Strassen (1969), Laderman (1976), Smirnov (2013), and tools (JAX 2018, TensorFlow 2016, V100 2017, TPU v2/v3/v4 2017–2021) are internally consistent with the submission timeline (submitted October 2021, accepted August 2022).

Evidence highlights

Central result: reduction of 4×4 matrix multiplication from 49 to 47 scalar multiplications, verified via open-source GitHub code.
Figure 5 speedup range: 4.3% to 23.9% across matrix sizes 8192 to 20480; reported std < 0.4 percentage points over medians of >200 runs.
Methodological innovations documented: Change-of-basis and synthetic demonstrations addressing the 10^12–10^29 action-space scaling challenge.
Hardware and software references all align chronologically with the 2021–2022 study window.

Notes

This assessment is based on textual report metadata rather than direct raw-image forensic analysis of figure files; a small residual possibility of undetected issues remains.
Mathematical correctness claims can be independently verified via the authors' GitHub release.
The reviewer notes that for hard math/AI-cross-disciplinary papers, falsifiability via open code substantially lowers fraud risk compared with biomedical image-heavy papers.
DOI preserved as provided: 10.1038/s41586-022-05172-4.