Reproduce “Words Before Pixels: Selective Modality Routing for Vision-Language Model Unlearning” 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

Laura Yao, Haochen Zhang, Jinhao Duan, Sijia Liu, Tianlong Chen

ModRoute selectively routes some unlearning examples through text while keeping others multimodal. It selects examples using per-example image-gradient saliency and replaces selected images with text renderings. On VLGuard with LLaVA-1.5-7B, it reports a U–U score of 0.847 at γ = 0.5, versus 0.581 for random switching at the same ratio; evidence is strongest for language-expressible targets.

Summary by alphaXiv.

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