A 2023 Stanford study showed that large language models may focus on the first and last parts of data in a context window, ignoring the middle