DayNews.ai

Vision-language position tricks slow models down

Spatial position encoding applied to text tokens in vision-language models hurts more than it helps, new research finds.