Lost in the middle
The tendency of models to notice info at the start and end of a long prompt but overlook the middle.
Think of it like
Like remembering the first and last speakers at a long meeting and blanking on everyone in between.
Example
A key fact placed on page 20 of a 40-page context gets ignored, while the same fact at the top or bottom gets used.
How it actually works
Studies find a U-shaped attention pattern: relevant information is used far more reliably when it sits near the edges of the context than when it’s buried in the middle. It means a big context window doesn’t guarantee the model actually reads all of it — placement and retrieval quality still matter.
For product teams
Why “it fits in the window” isn’t the same as “the model will use it”; put the crucial stuff at the edges.
For engineers
Positional recall bias where mid-context tokens are underattended relative to the head and tail.
Related
- Context stuffing — Made worse by dumping everything in.
- Needle in a haystack — The stress test for it.
- Effective context — Why usable context is less than the max.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome