Reproduce “Visually Irrelevant Hallucinations in Medical Vision-Language Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Ke Zou, Nan Zhou, Tao Yang, Meng Liu, Linchao He, Zehui Liao, Ling Huang, Guangren Huang and 4 more
CLEAR audits phrase-level visual hallucination risk in medical vision-language model outputs. It isolates assessable clinical phrases, proposes an image region for each, alters that region, and re-queries the same model to combine response and image–phrase cues. On MIMIC-CXR report generation with MedGemma-4B-it, it achieves 90.23% AUPRC versus 85.10% for VASE; this is detection, not prevention.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.