Where You Put a Note in the Context Changes the Answer

Position changes the answer. The same three notes, pasted in a different order, can produce a different reply — because models do not read a prompt evenly. Text at the beginning and end gets used; text in the middle gets skimmed. The fix is not a ranking rule. It is a test.

This is one of the best-documented findings in the field — and one of the least acted upon. Nelson F. Liu and six coauthors isolated position as a variable in Lost in the Middle: How Language Models Use Long Contexts, published in Transactions of the Association for Computational Linguistics, Volume 12 (2024).

They reported it plainly: "We find that performance can degrade significantly when changing the position of relevant information, indicating that current language models do not robustly make use of information in long input contexts."¹

Nothing about the content changed. Only where it sat. That is the whole finding, and it is the reason the order of your paste is not a cosmetic decision.

Volume is the other half of this story, and it already has its own post: the context window is working memory, so send less. This post assumes you have already decided the notes are going in. The question here is narrower and stranger. Once the text is in the prompt, does its address matter? It does.

What most people believe about a prompt

The intuitive model is that a prompt is a container. You put things in; the model reads them all; order is presentation, not substance. On this view a bigger window is strictly better, and where a note sits is a matter of tidiness — the same way the order of attachments on an email changes nothing about what the recipient learns.

The belief is reasonable. It is also how the interface behaves: one box, one blinking cursor, no hierarchy. Nothing in the design suggests that the middle of the box is a worse neighborhood than the edges. Anthropic describes the underlying reality differently, framing context as "a finite resource with diminishing marginal returns" and models as drawing on an "attention budget" when parsing large volumes of context.²

A budget is not a container. Budgets get spent unevenly.

Why the middle of your prompt is the weakest place to put anything

Because attention is scarce and not uniformly distributed. Liu and colleagues found "a distinctive U-shaped performance curve" — models are best at using information at the very beginning (primacy bias) or the end (recency bias) of the input, and degrade in the middle.³ The middle is not slightly worse. On their hardest settings, it was worse than sending nothing.

That last sentence is worth sitting with, and it lives in the body of the paper rather than the abstract. Liu's team measured a closed-book baseline: the model answering from memory, with no documents at all. GPT-3.5-Turbo scored 56.1% on it. When the relevant document was buried in the middle of 20- or 30-document inputs, performance fell below that line — "in the worst case, performance in 20- and 30-document settings is lower than performance without any input documents."⁴

Read that as a practical rule. A note in the wrong position can be worse than no note at all, because it does not merely fail to help — it crowds the budget while failing to help.

The effect is not confined to hard reasoning. On a trivial lookup task — find a value next to a key in a list of key-value pairs, a task with no reading comprehension in it whatsoever — Liu's team recorded a worst-case of 45.6% for GPT-3.5-Turbo (16K) at 300 pairs.⁵ Trivial task. Coin-flip accuracy. Position was doing that.

Cheng-Yu Hsieh and colleagues, in Found in the Middle (Findings of the Association for Computational Linguistics 2024), named the mechanism. They report that "LLMs exhibit an U-shaped attention bias where the tokens at the beginning and at the end of its input receive higher attention, regardless of their relevance."⁶

That final clause is the one to keep. Not in proportion to relevance. Regardless of it. The model is not weighing your notes and finding the middle one less convincing; it is allocating attention by address before relevance gets a vote.

The architecture explains the scarcity. Anthropic's engineering write-up puts it in one line: the transformer architecture "enables every token to attend to every other token across the entire context. This results in n² pairwise relationships for n tokens. As its context length increases, a model's ability to capture these pairwise relationships gets stretched thin, creating a natural tension between context size and attention focus."⁷

Quadratic pressure on a fixed budget produces exactly the shape Liu measured.

Test the order — do not rank it

Here is where most writing on this topic overreaches, and where this post will not. The evidence supports edges beat middle. It does not support arrange your notes logically, and it does not support any always-put-X-first rule. The honest instruction is a verb, not a hierarchy: test.

The reason is a result that cuts against the tidy story. Chroma's 2025 technical report by Kelly Hong, Anton Troynikov, and Jeff Huber evaluated 18 models, "including the state-of-the-art GPT-4.1, Claude 4, Gemini 2.5, and Qwen3 models,"⁸ and tested position explicitly: 8 input lengths swept against 11 needle positions for every combination of needle type, haystack topic, and haystack structure.⁹

Then they compared surrounding text that preserved a natural flow of ideas against the same sentences randomly reordered. The result inverted the intuition: "Across all 18 models and needle-haystack configurations, we observe a consistent pattern that models perform better on shuffled haystacks than on logically structured ones."¹⁰

Logical structure made things worse. Anyone selling you a clean ordering doctrine has not read past the abstract.

So the defensible claim is narrow and durable. Position has effects, the edges are stronger than the middle, and the sign of any particular arrangement is an empirical question about your prompt and your model, not a rule you can inherit. Hong, Troynikov, and Huber land in the same place in their conclusion: "Where and how information is presented in a model's context strongly influences task performance, making this a meaningful direction of future work for optimizing model performance."¹¹

Meaningful direction of future work is a researcher's way of saying nobody has the rule yet.

One position lever does have direct evidence behind it, and it costs a single line. Liu's team tried putting the question both before and after the data — a trick they call query-aware contextualization. It "dramatically improves performance on the key-value retrieval task," with all models reaching near-perfect scores at 75, 140, and 300 pairs.⁵

The same paper is candid that it "minimally affects performance trends in the multi-document question answering task."⁵ So it is a real lever with a known range: strong on lookup, weak on reasoning. Repeat your question after your notes.

The practice: five lines and a flip

Do this the next time an answer ignores something you pasted. It takes about a minute, needs no tooling, and converts a vague suspicion that the model missed something into a fact about your own prompt. The five steps below are a diagnostic, not a ranking system: the point is to find out whether arrangement is moving your answers.

1. Label each note. A one-line header (## Note A — Q3 rollout decision) gives you something to point at and the model something to anchor on.
2. Put the load-bearing note at an edge: first or last, not buried between two others.
3. Paste the passage itself, not a link to it. A URL is not context; an unfetched link has no position at all.
4. Repeat the question after the notes, not only before them.⁵
5. Flip the order and re-ask. Same notes, same question, reversed sequence.

Round 1:  [Question]  Note A · Note B · Note C  [Question again]
Round 2:  [Question]  Note C · Note B · Note A  [Question again]

Same notes. Same question. If the two answers disagree,
position is doing work you did not authorize.

Two answers that disagree are the whole diagnostic. Nothing else you can run is this cheap.

If they diverge, you have learned something you cannot learn any other way: the reply is sensitive to arrangement, which means it is not a stable reading of your material.

At that point ordering is a bandage. The durable move is to send less of it — which is a different post, and the reason this one stops here.

What this does not prove

Position is a gradient, not a switch. Anthropic states the limit better than a skeptic would: these factors "create a performance gradient rather than a hard cliff: models remain highly capable at longer contexts but may show reduced precision for information retrieval and long-range reasoning."¹² Nothing here means a middle note goes unread.

It means precision drops.

Every magnitude here is date-stamped, and the dates matter. Liu's numbers come from 2023-era models; the 56.1% closed-book figure belongs to GPT-3.5-Turbo, not to whatever you opened this morning.⁴ What survives is the direction, and the direction has been re-measured. Chroma's 18-model sweep found in 2025 that models "do not use their context uniformly."⁸

Ship the direction; treat every specific number as a fact about a specific model on a specific day. The economics of that gap are their own subject.

It is also worth knowing how thin the practitioner guidance underneath all this is. Writing in The Pragmatic Engineer on 2026-07-15, Gergely Orosz reports Dex Horthy's working heuristic for how much of a million-token window to use: "For a model with a 1M context window, Dex pushes it to around 300-400K when it feels right. For smaller models, he stops at around 100K."¹³

When it feels right. That is one experienced practitioner's gut, honestly labeled as such by the person reporting it. It is not a measured degradation threshold, and no study cited on this page establishes one. If the state of the art among people who do this daily is a feel, your own flip-order test is not a poor substitute for the rule. It is the closest thing to one that exists.

Two further honest limits. This post is about the text you control, and you do not control all of it: tools assemble system prompts, retrieved chunks, and history around your words, which is a bigger context than you typed. And Chroma's shuffled-haystack result concerned the structure of surrounding text, not the ordering of your own notes. It is a warning against tidy doctrine, not an instruction to shuffle your notes on purpose.

Frequently Asked Questions

Does the order of information in a prompt matter?

Yes, measurably. Liu et al. found that "performance can degrade significantly when changing the position of relevant information," with a U-shaped curve favoring the beginning and end over the middle.¹³ The content is identical; only the address changes. How much it matters depends on your model and task, which is why the flip-order test beats any inherited rule.

Why does the AI ignore what I pasted in the middle?

Because attention is allocated partly by position, not purely by relevance. Hsieh et al. found "an U-shaped attention bias where the tokens at the beginning and at the end of its input receive higher attention, regardless of their relevance."⁶ Your middle note is not being judged and dismissed. It is being under-weighted before its content is considered.

Should I put the most important information first or last?

An edge beats the middle — but "always first" is not supported. Chroma found models "perform better on shuffled haystacks than on logically structured ones" across all 18 models tested, which rules out any clean ordering doctrine.¹⁰ Put the load-bearing note at an edge, then flip the order and re-ask to see what your prompt actually does.

Is lost-in-the-middle still a problem with modern long-context models?

The direction survives; the magnitudes are dated. Chroma's 2025 report evaluated 18 current models and found they "do not use their context uniformly."⁸ Anthropic frames it as "a performance gradient rather than a hard cliff" — reduced precision, not failure.¹² Treat 2023-era percentages as history and the shape as live.

Does a bigger context window fix this?

No — capacity and attention are different resources. Anthropic notes that transformers create "n² pairwise relationships for n tokens," so attention gets "stretched thin" as context grows.⁷ Liu's team also found extended-context models were not necessarily better at using their input. The volume side of this argument is covered in its own post.

How do I know if position affected my answer?

Run the flip-order test. Paste the same notes in reverse order, ask the identical question, and compare. If the answers differ in substance, arrangement is influencing the result and the reply is not a stable reading of your material. If they match, position is not your problem — and you have ruled it out in about a minute.

What this changes

Your notes have addresses, and the addresses have consequences. That is an odd fact about a technology sold as comprehension — and it has a plain implication for the material you feed it. The arrangement has to stay yours: plain Markdown files on your own device, handed over in an order you chose, with a sequence you can reverse.

You cannot test what you cannot rearrange.

Nobody has the ordering rule. Everybody can run the test.

If you want the notes you hand to a model to be files you can reorder, label, and hand over on purpose, mnmnote.com keeps them as plain Markdown on your own device.