Reproduce “Look Before You Reason: Implicit Visual Thinking for Efficient Multimodal Reasoning” 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
Zerui Chen, Changrui Chen, Fei Ni, Jiankang Deng
LoRe helps vision-language models focus on small image details by amplifying already encoded visual features for a selected region before answering. It scales cached visual-token values, rather than creating a new visual observation. On COCO-ObjQA’s small-object split, exact-match accuracy rises from 72.3% to 83.1% for Qwen2.5-VL-7B; LoRe cannot recover detail discarded during image encoding.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.