The problem of context growing too large when inserting visual crops or frames alongside text in language models.