SAE-V: Interpreting Multimodal Models for Enhanced Alignment