Skip to content
#

cue-combination

Here is 1 public repository matching this topic...

PyTorch + Hugging Face implementation of GIB-200K, a 200,000-item benchmark testing whether vision–language models trust retrieved text or their own eyes. Accuracy collapses from 92.5–94.2% to 41.3–59.9% under stress; up to 71.4% of correct answers abandoned when the document lies; LLaVA over-weights text up to 1.32×; every number auditable.

  • Updated Nov 20, 2026
  • Python

Add this topic to your repo

To associate your repository with the cue-combination topic, visit your repo's landing page and select "manage topics."

Learn more