Multimodal Task Vectors Enable Many-Shot Multimodal In-Context Learning Brandon Huang 1* Chancharik Mitra 1* Leonid Karlinsky 3