Multimodal LLMs

Fewer Tokens, More Self-Teaching: On-Policy Self-Distillation for Extreme Visual Token Reduction featured image

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

Recovering multimodal capabilities under extreme visual-token reduction through on-policy self-distillation and a progressive token-budget curriculum.

junxian-li
•