Video54%accuracy with the answer buried mid-prompt (56% with no documents at all)
Your AI is skimming your prompt
TL;DR Models read the start and end of long prompts best and lose the middle. Put what matters first and last.
Your AI is skimming your prompt. Stanford hid one answer among 20 documents and moved it around. At the start: 76% right. At the end: 63%. In the middle: 54%. Worse than giving the model nothing at all.
“But that's 2023,” I hear you. So I checked. A 2025 study showed the bias is baked into how the models are built, not a bug. Simple lookups have improved a lot. Hard tasks in a wall of text are still a problem.
Try this
Important stuff first and last. Never bury the key fact in the middle of a long paste.
Sources: Liu et al., TACL 2023; Wu et al., ICML 2025; NoLiMa, ICML 2025.
