EFFICIENT MULTIMODAL REASONING WITH FEWER TOKENS

Summary from the TLDR AI newsletter

LLaVA-STF compresses vision token sequences by merging adjacent tokens and adds a multi-block token fusion module, enabling 75% token reduction.