Visual Language

On-Policy Self-Distillation for Extreme Visual Token Reduction: Self-Teaching with Fewer Tokens

The paper addresses extreme visual token reduction by using on-policy self-distillation to train multimodal mo…