Reproduce “Heads That Write, Not Just Point: Image Retrieval Heads 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
Junsung Park, Sangwon Yu, Donghun Kang, Yeongtak Oh, Jaewon Jeong, Gyeongtae Yoo, Han Cheol Moon, Jungbeom Lee and 1 more
The authors examine image selection in multi-image question-answering as in-context image retrieval. They compare heads’ attention to image spans with whether their outputs make target-image information recoverable. The rankings share only 35–47 of their top 100 heads. In output-zero ablations, overlapping heads produce the most consistent accuracy drops, especially for 2 to 20 images, across four open-weight models; this does not establish an exclusive cause.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.