An exploration of delaying selected visual tokens until they are needed, reducing wasted computation while preserving the model’s useful context.
Full article is a work in progres
Vision-Language Models · Draft
Speeding up VLM inference with late token insertion
