Reproduce “When VLMs Destroy What They Know: Spatial Representations Fall Below Lexical Priors” 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
Shubh Garg, Chen Shani
The authors introduce EICS to test whether a model’s text representation preserves a changed concept. A probe trained on original captions predicts new labels in captions edited while the image stays fixed. For CLIP ViT-B/32, spatial EICS is +0.070 versus +0.372 for word counts; this finding concerns pooled text representations under caption edits, not all VLM spatial reasoning.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.