Athulith Paraselli

VLMs show bias favoring text context over visual training data

A new research paper explores how vision-language models (VLMs) manage conflicting information between their training data and provided context. The study found that VLMs exhibit asymmetric biases, favoring text-based c…