Back to the discovery

Source record · 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

Type
Research paper · peer reviewed
Publication
Transactions of the Association for Computational Linguistics, vol. 12, pp. 157–173
Version used
Published TACL article, 2024.tacl-1.9
Source date
February 2024
Recorded in Atlas
22 October 2026
License
CC BY 4.0 (stated on the article’s first page)
DOI
10.1162/tacl_a_00638

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. No separate validity period has been established for the finding.


How Atlas uses this source

Check an AI answer against your documents

Sections 2.2–2.3, pp. 161–162; Section 3.2, p. 163

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.

Open the source passage

A long history is no guarantee of a good memory

Section 2.3, Figure 5, pp. 161–162 (PDF pages 5–6)
“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.

Open the source passage
Section 2.1, pp. 159–160 (PDF pages 3–4)
“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.

Open the source passage
Sections 2.2–2.3, pp. 161–162; Section 3.2, p. 163

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.

Open the source passage