Reproduce “Vision Transformers Learn Word Meaning from Images Alone” 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
Anders Gjølbye, Sebastian Ray Mason, Stella Frank, Lars Kai Hansen
The authors test whether image-only vision transformers encode relations among rendered words, not just their visual form. They compare embeddings of words on photographs with matched images containing scrambled-letter textures, then measure neighbors in the resulting word directions. DINOv3-H and DINOv3-7B retrieve capitals in the top ten for 60% of country queries, showing word-level association, not sentence understanding.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.