Reproduce “CRAFT: Causal Responsibility and Failure Tracing 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
Chunzheng Zhu, Jiaqi Zeng, Hongbo Zhao, Yihang Chen, Yijun Wang, Jianxin Lin
The authors introduce CRAFT, an inference-time method selecting and suppressing attention heads linked to medical vision-language model failures, while screening for clean-accuracy damage. It targets misleading text overriding image-supported diagnoses and definite answers to obscured images. In selected Hulu-Med 4B tests, conflict-following falls to 70.26% and unknown responses rise to 65.15%; in a separate test, clean accuracy falls to 90.67%.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.