Reproduce “Visual Grounding First, Multimodal In-context Learning Follows” 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
Minhyuk Seo, Minjae Lee, Chaeeun Lee, Wei Lin, Muhammad Jehanzeb Mirza, Tinne Tuytelaars, Jonghyun Choi
ProCoRe trains multimodal models to use image evidence in in-context captioning without training on in-context examples. It uses visually contrasting image-caption pairs, decomposing captions into unique and shared propositions and rewarding each according to its image support. On MS COCO at two shots, it scores 46.41 ROUGE-L, above two ablations; this supports proposition-level rewards, not a sole causal explanation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.