Understanding and Improving Length Generalization in Recurrent Models

Ricardo Buitrago Ruiz, Albert Gu · 2025

Read Understanding and Improving Length Generalization in Recurrent Models by Ricardo Buitrago Ruiz, Albert Gu (2025) 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-2507-02782.txt.

Description

Buitrago Ruiz and Gu read in twenty twenty-five why loop models fail on texts far longer than their training sets. Such models run in line time and should take any length in thought. Their guess: training shows just a thin slice of the states long runs reach.

Seeds of noise or of states from other texts widen that slice in training. Five hundred extra steps, a scrap of the first budget, fixed the flaw. Models then held texts a vast span past their training length. Long task scores rose as well, they report.

Catalog record

CREATORRicardo Buitrago Ruiz, Albert Gu
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE54 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2507-02782

Browse the full library · About the Tower of Babel