Reproduce “The Hidden Evolution of Disguised Visual Context inside the VLM” 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
Wish Suharitdamrong, Tony Alex, Xiatian Zhu, Muhammad Awais, Sara Atito
The paper compares how visual tokens enter vision-language models: input-level injection concatenates projected image features with text, while layer-wise methods add visual information within transformer layers. A controlled study finds input-level injection generally performs best, especially on OCR and video, despite similar attention patterns. Low-pass filtering supports high-frequency representations as one contributor, not sole cause, and input-level processing costs more.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.