Reproduce “Learning to Look: A Foundation for Visual 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

Wenshuai Zhao, Yihao Wang, Benjamin Robson, Juho Kannala, Arno Solin

The authors introduce LOOK to train vision-language models to track where image crops came from. It preserves each crop’s original positional information and uses known coordinates as automatic bounding-box labels. On VisABench, Qwen3-VL-8B-Instruct with LOOK at a 10:10 data ratio has a 0.8% no-image Yes rate, versus 31.2% with ordinary instruction fine-tuning. This VisABench result does not establish faithful grounding on natural images.

Summary by alphaXiv.

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