noun · AI failure
Lost in the middle
also called middle blind spot
When you hand someone a very long document, they remember the start and the end best. The AI does too.
The observed tendency of language models to use information at the start and end of a long input more reliably than information in the middle.
Studies of long inputs have found that models often use information near the beginning and end better than information buried in the middle. The effect varies by model and task, and it is why more context does not always mean better answers.
Real-life example
You paste a hundred-page report and ask about a figure on page fifty. The answer is vague or wrong, but the same question about page two works fine. The page was in the window the whole time. It just sat where models tend to pay the least attention.
Where the analogy breaks
-
It does not happen everywhere. It depends on the model and the task, and newer models are improving, so test your own setup instead of assuming.
-
A bigger window does not cure it. Fitting more text in only gives the middle more room to hide.
-
You can work around it. Put the most important items near the start or end, keep the input short, and repeat key instructions where they will be noticed.