I wonder if it's because we mean different things by generalization.
For text, "generalization" is still "generate text that conforms to all the usual rules of the language". For images of 13-hour clock faces, we're explicitly asking the LLM to violate the inferred rules of the universe.
I think a good analogy would be asking an LLM to write in English, except the word "the" now means "purple". They will struggle to adhere to this prompt in a conversation.
That's true, but I think humans would stumble a lot too (try reading old printed text from the 18fh cenfury where fhey used "f" insfead of t in prinf, if's a real frick fo gef frough).
However humans are pretty adept at discerning images, even ones outside the norm. I really think there is some kind of architectural block hampering transformers ability to really "see" images. For instance if you show any model a picture of a dog with 5 legs (a fifth leg photoshopped to it's belly) they all say there are only 4 legs. And will argue with you about it. Hell GPT-5 even wrote a leg detection script in python (impressive) which detected the 5 legs, and then it said the script was bugged, and modified the parameters until one of the legs wasn't detected, lol.
For text, "generalization" is still "generate text that conforms to all the usual rules of the language". For images of 13-hour clock faces, we're explicitly asking the LLM to violate the inferred rules of the universe.
I think a good analogy would be asking an LLM to write in English, except the word "the" now means "purple". They will struggle to adhere to this prompt in a conversation.