Reproduce “Look Where It Matters: High-Resolution Crops Retrieval for Efficient VLMs” 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
Nimrod Shabtay, Moshe Kimhi, Artem Spector, Sivan Haray, Ehud Rivlin, Chaim Baskin, Raja Giryes, Eli Schwartz
AwaRes lets a vision-language model answer image questions by first viewing a low-resolution overview, then either answering or requesting selected high-resolution crops. Its structured first-turn choice jointly decides whether and where to seek detail. On six benchmarks with Qwen2.5-VL-7B-Instruct, it averages 80.30 versus 80.46 for full resolution at 0.36 retained-token ratio; results cover one base model and a fixed crop menu.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.