VLMs Need Words: Vision Language Models Ignore Visual Detail In Favor of Semantic Anchors Paper • 2604.02486 • Published 17 days ago • 10