Reproduce “Encoded but Unused: Bounding the Sources of VLM Errors” 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
Anat Veikherman, Mor Ventura, Nitay Calderon, Lihi Zelnik-Manor, Roi Reichart
VELens estimates how often incorrect answers reflect unused rather than missing visual evidence in two frozen 7B vision-language models. It probes visual representations on VAW images, testing color, material, and shape questions. Among errors, estimated utilization-failure rates are 84.4% for LLaVA-1.6-7B and 70.0% for Qwen2.5-VL-7B; conservative lower bounds are 77.5% and 61.0%. Probe estimates are imperfect.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.