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.
benchmark grounding retrieval-augmented-generation multimodal-llm hallucination-detection vision-language-models knowledge-conflicts multimodal-rag robustness-evaluation vlm-evaluation ideal-observer cross-modal-conflict cue-combination
-
Updated
Nov 20, 2026 - Python