Radiology VLM Benchmark Audit Reveals Reproducibility Issues.

Mateusz Koz{\l}owski· July 31, 2026 View original

Key takeaways

  • A forensic audit revealed significant reproducibility issues in a radiology VLM benchmark.
  • Errors included incorrect image rendering, data truncation, and inconsistent prompt execution.
  • These issues led to the withdrawal of original performance and clinical claims.
  • Rigorous, machine-verifiable controls are essential for reliable medical AI benchmarks.

Who benefits

HealthcareMedical DevicesPharmaceuticalsRegulatory BodiesAI Research

Summary

A forensic audit of a radiology vision-language model benchmark uncovered significant reproducibility issues across various artifacts, including incorrect image rendering, data truncation, and inconsistent prompt execution. These findings led to the withdrawal of original performance claims and highlight the need for rigorous verification in medical AI benchmarks.

This paper details a retrospective forensic reproducibility audit conducted on a preserved pilot benchmark for radiology vision-language models (VLMs). The audit meticulously traced various components, including prompt bindings, DICOM metadata, output completeness, label extraction, and statistical analyses, to assess the agreement across different artifacts. The findings revealed substantial discrepancies and errors within the benchmark. Specific issues included incorrect polarity inversion for certain images, failure to retain dataset split membership, truncation of reports by the unvalidated extractor, and inconsistent prompt execution (e.g., using the same prompt for different labeled calls). These errors significantly impacted the statistical results, leading to changes in key metrics and McNemar comparisons. Consequently, the original performance, ranking, prompt-effect, and clinical claims derived from the benchmark have been withdrawn. The audit underscores the critical need for machine-verifiable controls and rigorous validation across all stages of medical imaging AI benchmark development and release.

Why it matters

Professionals in medical AI development, research, and regulatory oversight must recognize the critical importance of rigorous reproducibility and verification in benchmarks to ensure the validity and trustworthiness of AI models in healthcare.

How to implement this in your domain

  1. 1Establish strict protocols for data provenance, DICOM rendering, and prompt identity in all medical AI benchmark development.
  2. 2Implement automated, machine-verifiable controls for dataset integrity, model calls, and annotation processes.
  3. 3Conduct independent reproducibility audits of internal and external AI benchmarks before relying on their results for critical decisions.
  4. 4Prioritize transparency and detailed documentation of all artifacts and methodologies in AI research publications.

Original post by Mateusz Koz{\l}owski

"arXiv:2607.25589v1 Announce Type: cross Abstract: Medical-imaging AI benchmarks combine datasets, DICOM rendering, prompts, provider APIs, automated labels, statistical code, manuscripts, and repository releases. Agreement across these artifacts is usually assumed rather than tes…"

View on X

Originally posted by Mateusz Koz{\l}owski on X · view source

Want to go deeper?

Turn these trends into skills with Learnijoy's hands-on AI & tech courses.

Explore courses

More in AI Research

AI ResearchAI Engineering & DevTools

New Framework Improves Partial Multi-View Clustering Performance.

DAS-PMVC is a novel framework for partial multi-view clustering that addresses view asymmetry and irrelevant samples by leveraging dual alignment and structure enhancement. It uses anchor graph structure alignment, structure-enhanced feature learning, and a dual alignment strategy to achieve superior clustering performance on various datasets.

Shubin Ma, Liang Zhao, Chuanye He, Zhenjiao Liu, Liang Zou, Lin Yuanbo Wu, Yu ShaoJul 31, 2026
AI Engineering & DevToolsAI Research

Dual Teachers Improve Adversarial Robustness and Accuracy.

This work extends Information Bottleneck Distillation (IBD) by introducing a "clean teacher" alongside a robust teacher to improve the robustness/accuracy tradeoff against adversarial attacks. The proposed method transfers features from both teachers to a student model, achieving better clean accuracy while maintaining adversarial robustness, outperforming original IBD and competing with state-of-the-art approaches.

Vincent Ryusuke Takahashi, Yoshinari Takeishi, Jun'ichi Takeuchi, Kave SalamatianJul 31, 2026
AI Engineering & DevToolsAI Research

Dynamic Batch Sizes Improve Large Language Model Training Efficiency.

This paper proposes a new approach to deep learning dynamics, deriving joint scaling laws for loss based on both learning rate and batch size schedules. It introduces an optimal dynamic batch size schedule that consistently outperforms static batch size baselines, highlighting its importance for large language model training.

Jiaxiang Li, Zhiqi Bu, Shiyun XuJul 31, 2026