Reproduce “Seen but Unsaid: Locating and Repairing Cross-Image Comparison in Vision-Language Models” 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

Zhaoning Zhang

The authors distinguish failures in cross-image matching between answer readout and weak visual comparison. They find decodable correspondence that does not reliably reach answers, while selected attention heads causally support synthetic and same-instance matching. Training low-rank updates to these heads raises SPair accuracy from .430 to .712 on Qwen3-VL-8B, but these heads are not a bottleneck for base semantic performance.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.