How does lost-in-the-middle position bias affect where you place key facts in a prompt?

## The shape of the failure If you take a long prompt, hold its total length constant, and slide one required fact from the very beginning to the very end while re-asking the same question, you do not get a flat accuracy line. You get a curve that is high at both ends and lower in between. This is the finding usually referred to as **lost in the middle**, and it is one of the most robust empirical facts about long-context behaviour: presence in the window does not imply usability, and *where* something sits matters as much as *whether* it is there. ## Why the ends win Two effects overlap. **Primacy**: the opening of the prompt is where instructions, system framing and role material conventionally live, so models are heavily trained to weight it — and every subsequent token attends back over it. **Recency**: the tokens nearest the point of generation are the freshest and most strongly attended, which is why an instruction repeated at the very end often overrides one given ten thousand tokens earlier. The middle inherits neither advantage: it is far from the framing at the top and far from the generation point at the bottom, so it competes for attention on the merits alone against a very large field. The exact shape is model-specific and moves with training. Some models are notably more end-heavy than start-heavy; some flatten the curve considerably after long-context mid-training. What has been stable is the direction: the middle is the weakest region, and the sag deepens as the prompt gets longer. ## A concrete case A clinical-trial protocol assistant is answering eligibility questions over a 300-page protocol plus its amendments. The team runs a depth sweep: the same exclusion criterion — a renal-function threshold — is planted at 0%, 25%, 50%, 75% and 100% of the way through, and the same question is asked each time. The criterion is found essentially every time at the top and at the bottom. At around 55% depth, the model answers as though the criterion does not exist and declares the patient eligible. Nothing about the fact changed; only its coordinates did. That single experiment is the most persuasive artifact you can bring to a design review, because it converts a vague worry about "long prompts" into a measured, reproducible bias. ## What to do about it **Order deliberately.** If you are assembling retrieved passages, do not hand them over in raw score order top-to-bottom. Place the highest-scoring evidence at the edges of the block and let the weaker material occupy the middle, where it will be underweighted anyway. Some teams interleave — best first, second-best last — precisely to exploit both ends. **Restate the task at the end.** After a long body of text, repeat the question or the output requirement immediately before generation. This costs a handful of tokens and reliably recovers instruction-following that the body of the prompt had washed out. **Cut, don't only reorder.** The sag is a function of length. Reducing the amount of low-value material shortens the region where the bias bites and often outperforms any reordering scheme, because it moves the useful content closer to an edge as a side effect. **Make position visible.** Section headers, document identifiers and explicit chunk boundaries give the model addressable structure, so an answer can be grounded in "section 7.2" rather than in a wall of text. This does not eliminate the bias but it makes mid-prompt material easier to locate when the model is prompted to look for it. **Verify, do not assume.** Reordering feels like it should work, and it usually helps, but the size of the help is model-specific. The check is the same instrument that revealed the problem: hold length constant, sweep depth, measure. If your reordering strategy is not measurably better on that sweep, it is superstition. ## The limits of the mitigation Ordering tricks address *retrieval of a single fact*. They do much less for tasks that must use material from everywhere at once — counting occurrences across a whole document, reconciling a dozen sections, or summarizing faithfully. In those tasks there is no edge to promote the important content to, because all of it is important. That is the point at which the answer stops being prompt layout and becomes decomposition: process the document in pieces and combine the results, rather than hoping a single pass attends evenly to everything. It is also worth being honest that the effect is not universal or fixed. Newer models flatten the curve, and some report near-flat recall on simple single-fact retrieval while still sagging badly on harder mid-prompt reasoning. Claiming a specific U shape for a specific model without having measured it is exactly the kind of stale prior that long-context work punishes.