analiza aiqbench
This research demonstrates how reinforcement learning can train vision-language models to generate medical checklists that inadvertently introduce bias when validating unmentioned findings in radiology reports. The study highlights a critical safety concern where model performance metrics diverge significantly between training formats and independent evaluations, underscoring the risks of relying on AI-generated labels rather than radiologist-adjudicated data for automated checking systems.
Unmentioned Checklist Findings Change How Reinforcement Learning Appears to Improve Chest Radiograph Report Checking
arXiv:2610.05425v1 Announce Type: cross Abstract: Automated checks of radiology reports may rely on AI-generated checklists that leave findings unmentioned. We used reinforcement learning to train a vision-language model to fill in a 12-finding checklist from a chest radiograph without seeing the sentence under test; a separate checking model judged the sentence from the checklist. On held-out patients, a rule-based check and an independent medical checker, neither used in training, measured discrimination gains (Youden index) of 12.6% and 11.8%; only the rule-based check met the prespecified false-alarm criterion. Switching to the training format, which fixes finding order and enters unmentioned findings as absent, raised the training checker's measured gain and lowered the independent checker's, a prespecified comparison that yielded 6.2% (95% interval 2.0% to 10.5%) and, post hoc on held-out patients, 7.7%. Across 8 checking models, acceptance of a label-consistent negative statement about an unmentioned finding ranged from 1.0% to 97.0%. Labels were report-derived, not radiologist-adjudicated.
Abstract:Automated checks of radiology reports may rely on AI-generated checklists that leave findings unmentioned. We used reinforcement learning to train a vision-language model to fill in a 12-finding checklist from a chest radiograph without seeing the sentence under test; a separate checking model judged the sentence from the checklist. On held-out patients, a rule-based check and an independent medical checker, neither used in training, measured discrimination gains (Youden index) of 12.6% and 11.8%; only the rule-based check met the prespecified false-alarm criterion. Switching to the training format, which fixes finding order and enters unmentioned findings as absent, raised the training checker's measured gain and lowered the independent checker's, a prespecified comparison that yielded 6.2% (95% interval 2.0% to 10.5%) and, post hoc on held-out patients, 7.7%. Across 8 checking models, acceptance of a label-consistent negative statement about an unmentioned finding ranged from 1.0% to 97.0%. Labels were report-derived, not radiologist-adjudicated.
40 pages, 7 figures, 17 tables (main text and references pp. 1-19; Supplementary Information as an appendix, pp. 20-40). Submitted to npj Digital Medicine. Code: this https URL
Computer Vision and Pattern Recognition (cs.CV); Computation and Language (cs.CL); Machine Learning (cs.LG)
arXiv:2610.05425 [cs.CV]
(or arXiv:2610.05425v1 [cs.CV] for this version)
https://doi.org/10.48550/arXiv.2610.05425 arXiv-issued DOI via DataCite (pending registration)
Text integral sub licența deschisă a publicației · 206 cuvinte
Subiecte conexe
Reviewing failure modes in science as we prepare for AI researchers
LessWrong · 8 oct. 2026
Latent reasoning architectures would likely undermine CoT, our strongest oversight tool
AI Alignment Forum · 23 sept. 2026
Confounding Masquerading as Improvement: A Systematic Evaluation of Offline Reinforcement Learning for Stroke Antithrombotic Treatment in a 129,000-Patient Registry
arXiv · 1 sept. 2026
An unexamined cause of the OpenAI Hugging Face hacking incident: its binary performance metric
LessWrong · 23 sept. 2026