Reproduce “Lost in Read-out: Locating the Image Quality Gap in Vision-Language Models with a Closed-Form Edit” 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
Zhuangzi Li, Shilv Cai, Jian Jin, Zhiyi Pan, Haihong Xiao, Weisi Lin
The authors test whether vision-language models can improve image judgments by changing how they read information in their internal states. They rewrite one last-layer neuron using labels, then compare its scores with human ratings. On Qwen3.5-0.8B, Spearman correlation rises from 0.642 to 0.800 across held-out prompts; gains track probes but do not transfer to all image types or question formats.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.