Reproduce “FOCUS: Closed-Loop Attention Feedback for Efficient Vision-Language Understanding” 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
Akul Santhosh, Aastha Jhunjhunwala
FOCUS adapts streaming video observation, using a small encoder to detect visual change and trigger bursts for a vision-language model. It combines change maps with attention from its response to steer bursts, without deleting tokens or training. On EgoSchema, it used 25.4 calls per clip at 74.8% accuracy, approximately matching baselines; this does not establish an accuracy gain.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.