Skip to content
BENIBIX Learn

Context Rot vs Lost in the Middle

People think context rot and lost in the middle are one problem, and they are not. Lost in the middle is about where a fact sits. The model misses facts in the middle of a long input more than ones at the start or the end. Context rot is about how much text there is. The more you add, the worse the model does, wherever the fact sits.

AI💬 LLMs

Context Rot

A model gets worse at its task as the text you give it grows longer.

Context rot is when a language model gets worse as you give it more text to read. The context window may hold far more, but quality still slides as the input grows. A 2025 study tested 18 models and found this in all of them, even on simple tasks. Text that looks like the answer but is not makes it worse. It is why long chats and huge prompts can drift.

Example
Lucy pastes ten long files into one chat and asks about one line. The model mixes up two files. In a fresh chat with only the right file, it answers well.
Why it matters
A bigger context window is not a free pass. Giving the model less, and the right text, often works better.
Full entry
AI💬 LLMs

Lost in the Middle

Models miss facts in the middle of a long text more often.

Lost in the middle is the finding that language models use facts at the start or end of a long input better than facts in the middle. When the key fact sits in the middle of many pages, the model misses it more often. A 2023 study showed the drop clearly, with a U-shaped curve. Newer models handle long inputs better, but the effect still shows up. Putting the key text near the start or end helps.

Example
Jack pastes 20 documents and asks a question whose answer is in document 10. The model misses it, but finds it when that document is moved to the top.
Why it matters
A big context window does not mean every part of it is read just as well.
Full entry

In one line

Context Rot
A model gets worse at its task as the text you give it grows longer.
Lost in the Middle
Models miss facts in the middle of a long text more often.