Reproduce “The in-context inductive biases of vision-language models differ across modalities” 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
Kelsey Allen, Ishita Dasgupta, Eliza Kosoy, Andrew K. Lampinen
The authors compare models’ generalization from pictures and textual descriptions of ambiguous categories defined by shape and color, also varying text descriptor order. Models tend toward shape, usually more from images, while text order shifts choices. In odd-one-out, Claude 3.5 Sonnet instead shows stronger shape bias from text; results vary by task and model and measure choices, not internal representations.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.