Wow, it's fascinating how you explain the shift! I'm super curious, how exactly do multimodal LLMs bridge that gap of lost image-level information for structured datasets? Such a smart analysis!