It is all still there. Yet the important part can get lost.

Imagine the answer you need is somewhere in a set of documents. It moves from the beginning to the middle, but the content stays the same. Should that matter? In Lost in the Middle, it did.

Move the document containing the answer

Same information. Different position.

Start of the contextEnd →
Still there, but harder to use

When the same relevant document moved toward the middle, accuracy declined for the models tested.

A schematic illustration of the experiment’s design, not measured data or a new AI test. Original results: Figure 5, page 161. View the original figure ↗

What supports the answer?

  1. Reported finding: Position affected the answer
    Reported finding

    In the study’s document task, accuracy was highest when the document containing the answer came first or last. It declined when the relevant information was in the middle.

    Scope and conditions: Multi-document question answering with 10, 20 or 30 documents; models and versions listed in Section 2.2. This is not an estimate of today’s model performance.

    Source passage · 2024
    “performance is highest when relevant information occurs at the beginning or end”

    Figure 5 plots accuracy against document position. The discussion continues on page 162 with the decline in the middle.

    Section 2.3, Figure 5, pp. 161–162 (PDF pages 5–6) About the source and its version
  2. Method described: They moved the document, not the answer
    Method described

    The researchers used 2,655 questions from NaturalQuestions-Open. One document contained the annotated answer; other documents served as distractors. They varied the order and number of documents.

    Scope and conditions: Wikipedia passages and a bounded question format. The authors acknowledge that some distractors may still contain reasonable answers and check a less ambiguous subset in Appendix A.

    Source passage · 2024
    “exactly one of the documents contains the answer”

    This section describes the sample of 2,655 questions, how distractors were selected and how the document order was changed.

    Section 2.1, pp. 159–160 (PDF pages 3–4) About the source and its version
  3. Reported finding: More capacity did not automatically mean better use
    Reported finding

    When the input fit into both model variants, the models with longer context windows performed roughly like their shorter-context counterparts.

    Scope and conditions: A comparison within the model pairs studied, using inputs that fit into both windows. The study does not show that longer context windows are generally without value.

    Source passage · 2024

    Page 162 compares model pairs when the same input fits into both windows. Model versions are listed on page 161. Exceptions in the key–value task are discussed on page 163.

    Sections 2.2–2.3, pp. 161–162; Section 3.2, p. 163 About the source and its version

Where this explanation stops.

  • The study tested older models, including GPT-3.5-Turbo (0613), Claude 1.3, MPT-30B-Instruct and LongChat-13B. It does not show how today’s models perform.
  • The main experiment involved answering questions using documents. It did not test personal long-term memory or real conversations lasting months.
  • Results vary by task and model. In the study’s separate key–value retrieval task, for example, the Claude models handled almost all cases correctly. The claim is not that every AI always loses information in the middle.

What to take with you

For people using AI, this raises a useful question: have we simply given the system the information, or have we also checked whether it can find and use it? This is Atlas’s interpretation of the study’s significance, not a test result about today’s chat products.

Follow the answer back to its sources.

Research paper · peer reviewed · Source date: February 2024

Lost in the Middle: How Language Models Use Long Contexts

Nelson F. Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni and Percy Liang

Transactions of the Association for Computational Linguistics, vol. 12, pp. 157–173 · Published TACL article, 2024.tacl-1.9

The publication month is listed as 2/2024 on the PDF’s first page. Atlas recorded the source on 22 October 2026; that is not the date of a new research finding.

DOI: 10.1162/tacl_a_00638

Source passages, license and conditions ↗

How this version was checked

AI-reviewed by a separate editorial agent · 22 October 2026. The approval applies to revision 4, its text, visuals and source references.

Why this piece

Chosen because it separates two things we easily confuse: how much information an AI can take in, and how well it can use it. That distinction remains worth exploring as models change.

AI-reviewed by a separate editorial agent · Ask the editor

What changed, and why.

This history comes from saved content versions. A new revision can be an editorial change; it does not by itself mean that the underlying facts changed.

Latest reviewed revision

Editorial change · revision 4

Adds a schematic visual and its source and file record. The explanation and practical advice are unchanged; no new model test is claimed.

Recorded 22 October 2026 · reviewed 22 October 2026 · AI-reviewed by a separate editorial agent

No earlier reviewed assessment is available to compare. No change in the underlying facts is being claimed.

Available version records (1)
  1. Revision 4 · Editorial change · reviewed

    Adds a schematic visual and its source and file record. The explanation and practical advice are unchanged; no new model test is claimed.

    Recorded 22 October 2026 · reviewed 22 October 2026 · AI-reviewed by a separate editorial agent

    Answer in this version

    MemGPT describes a way to separate stored information from the material in a model’s current context. The system can move information between external memory and a limited context window. It is an architecture worth exploring, not proof that selection always beats a longer history.

    What remained unresolved

    The next investigation would need to compare selection with full history on the same task and model, checking both missed details and incorrectly retrieved information. Atlas has not made that comparison yet.

    Takeaway and scope

    In the older models studied, fitting information into the context did not guarantee that it would be used reliably.

    A study of older models answering questions from documents — not a test of today’s chat products or personal long-term memory.

    Position affected the answer

    In the study’s document task, accuracy was highest when the document containing the answer came first or last. It declined when the relevant information was in the middle.

    Multi-document question answering with 10, 20 or 30 documents; models and versions listed in Section 2.2. This is not an estimate of today’s model performance.

    They moved the document, not the answer

    The researchers used 2,655 questions from NaturalQuestions-Open. One document contained the annotated answer; other documents served as distractors. They varied the order and number of documents.

    Wikipedia passages and a bounded question format. The authors acknowledge that some distractors may still contain reasonable answers and check a less ambiguous subset in Appendix A.

    More capacity did not automatically mean better use

    When the input fit into both model variants, the models with longer context windows performed roughly like their shorter-context counterparts.

    A comparison within the model pairs studied, using inputs that fit into both windows. The study does not show that longer context windows are generally without value.

    Storage and current context can be separate

    MemGPT distinguishes between a model’s current context and external storage. Information outside the context must be retrieved into it to be included in the model’s next inference.

    A description of MemGPT v2’s architecture. This is not a claim that it solves Lost in the Middle or is the best approach for every task.

    A document contains a highlighted relevant passage. A schematic answer may omit that passage even though it was available.
    Schematic · not measured data

    A conceptual illustration of the distinction behind the older context-position study. It shows no measured performance and makes no claim about a current model.

    Sources in this version