Studies report that models use information at the start and end of long inputs better than the middle. That is an observed tendency, not a law, and its strength varies by model: measure it.
Long-context use is less reliable than short-context use in general. A fact buried in the middle of a huge paste is the likeliest thing to be ignored.
Pasting tens of thousands of tokens of RTL fills the window without filling it reliably. The quadratic cost buys comparisons, not careful reading of every line.
Do not work around this by sending secrets, or proprietary IP to an unapproved service. A hosted service is someone else's computer, so send only what the task needs.
Paste strategy | What happens | Why it disappoints |
|---|---|---|
Whole file dump | Key lines sit mid-window | The middle gets the weakest use. |
Many files at once | Budget spent on unread context | Quadratic work rises while answers stay flat. |
Extract plus pointer | Relevant lines up front, file named | Short input with clear provenance. |
Shrink the input before sending it. Retrieve the relevant slice, put the question and the key lines first or last, and verify the answer against the source.
Extract the failing module or log window instead of the whole tree.
State the file and line numbers so the answer can be checked.
Ask for one thing per turn and confirm each claim against the source.
Long inputs are used least reliably in the middle. Move decisive content to the start or the end, or shorten the input.
Cost and capacity are not comprehension. Basic attention work grows quadratically, reliability falls with length, and middle content is used worst, so most of the paste contributes noise rather than answers.