Reproduce “Deciding Without Pointing: A Case Study of Whether a Decision Model Localizes the Anomalies It Detects” 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

Hsiang-Hung Lee

Clef-flash uses a decision head to classify defective parts without generating text, and the authors test whether its attention maps locate defects and reflect the decision. On defective MVTec AD images, the head map’s per-image pixel AUROC is 0.391, below chance and a center prior. Image-region interventions nevertheless support defect-region dependence on tested images, not across products.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.