An "lost in the middle" effect exists in ultra-long contexts, meaning the model has stronger memory for content at the beginning and end of a sequence while attention to middle passages decreases; this phenomenon was named by Stanford researchers.