The Illusion of Grounding: Do Vision-Language Models Weigh Retrieved Text and Images Like an Ideal Observer?
Misleading documents caused the models to abandon 34.8%–71.4% of answers they had initially got right.
The researchers tested four open vision-language models on 2,000 items with retrieval noise, degraded images and conflicting text and visual evidence. Accuracy fell from 92.5%–94.2% on clean inputs to 41.3%–59.9% under adverse conditions. Two LLaVA models overweighted text relative to their own image evidence, while InternVL3 favoured the image. Comparing answers with and without retrieval improved accuracy under misleading documents at the cost of answering fewer questions; falling back to the image failed when it was heavily degraded.
ArXiv · AI/CL/LG's note
The researchers tested four open vision-language models on 2,000 items with retrieval noise, degraded images and conflicting text and visual evidence. Accuracy fell from 92.5%–94.2% on clean inputs to 41.3%–59.9% under adverse conditions. Two LLaVA models overweighted text relative to their own image evidence, while InternVL3 favoured the image. Comparing answers with and without retrieval improved accuracy under misleading documents at the cost of answering fewer questions; falling back to the image failed when it was heavily degraded.
ArXiv · AI/CL/LG's note
score 5