DictionaryContext Management
Lost in the Middle
AIs pay most attention to the start and end of a long text and tend to miss what is in the middle.
Definition
A finding by Liu et al. (2023) that language models use information at the beginning or end of a long context best, and that performance degrades significantly when the relevant information sits in the middle, even for models built for long contexts. It is one reason to put key instructions and the question near the edges of a prompt.
Example
Twenty documents are pasted and the answer is in document 10; accuracy drops compared with placing that document first or last.