Beyond Many-Shot Translation: Scaling In - Context Demonstrations For Low - Resource Machine Translation

Luis Frentzen Salim, Esteban Carlin, Alexandre Morinvil, Xi Ai, Lun-Wei Ku · 2026

Read Beyond Many-Shot Translation: Scaling In - Context Demonstrations For Low - Resource Machine Translation by Luis Frentzen Salim, Esteban Carlin, Alexandre Morinvil, Xi Ai, Lun-Wei Ku (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-2602-04764.txt.

Description

Luis Frentzen Salim and 4 co-authors test how far example-based learning scales for low-resource translation in a 2026 paper. Languages such as Javanese and Sundanese lack the large parallel datasets that translation systems usually need.

Instead of training, the method conditions a long-context language model on demonstrations at inference time. The team scaled the context budget up to a million tokens and compared three supervision sources. Monolingual data, instruction-style data, and parallel sentence pairs were each tried. Gains from extra context saturated quickly and could degrade near the maximum window. Corpus type strongly shaped the scaling behavior. Some monolingual supervision rivaled parallel data despite offering less direct guidance. Bigger context windows did not guarantee proportional quality gains.

Catalog record

CREATORLuis Frentzen Salim, Esteban Carlin, Alexandre Morinvil, Xi Ai, Lun-Wei Ku
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE56 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2602-04764

Browse the full library · About the Tower of Babel