Reproduce “GeoPID: Decomposing and Steering Visual Information in 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
Seulgi Kim, Zhixiong Zhang, Xinwei Zhang, Jie Ling, Ronn Shaw
GeoPID separates shared image-text information from modality-specific directions and selectively amplifies image-specific components in a vision-language model without updating its parameters. It estimates subspaces from visual and text tokens, then uses 100 calibration questions per model-benchmark pair to select settings. Across 22 VLMs and 14 benchmarks, it reports a 7.63% average relative accuracy gain, but results vary and can be negative.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.