Nearly Optimal Attention Coresets

Edo Liberty, Alexandr Andoni, Eldar Kleiner · 2026

Read Nearly Optimal Attention Coresets by Edo Liberty, Alexandr Andoni, Eldar Kleiner (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-05602.txt.

Description

Attention, the core operation of transformer models, is expensive to compute over long inputs. Edo Liberty, Alexandr Andoni, and Eldar Kleiner prove that a small subset of the keys and values can stand in for the whole set. Such a subset is called a coreset, and theirs is nearly as small as theory allows.

For queries of bounded size, the coreset reproduces attention outputs within a set error. Its size beats the best previously known results. They also prove a matching-style lower bound, showing that noticeably smaller coresets cannot exist in general.

Catalog record

CREATOREdo Liberty, Alexandr Andoni, Eldar Kleiner
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE34 KB
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.05602

Browse the full library · About the Tower of Babel