Reproduce “AdaLoop: Thinking with the Vision Encoder” 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
Xinran Nicole Han, Yiyang Lu, Katelyn Gan, Todd Zickler
AdaLoop lets a vision-language model revisit visual representations while answering questions about structured images. It reuses a span of frozen vision-encoder transformations on an evolving visual state, with question-conditioned normalization guiding updates. On procedural VisRO-Core6 tasks, Qwen2.5-VL-3B reaches 64.6% in-distribution accuracy versus 60.0% for a single-pass control; these supervised results do not establish gains on arbitrary real-world tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.