COMPASS: COntinual Multilingual PEFT with Adaptive Semantic Sampling

Noah Flynn [email protected] · 2026

Read COMPASS: COntinual Multilingual PEFT with Adaptive Semantic Sampling by Noah Flynn [email protected] (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-2604-20720.txt.

Description

Flynn offers COMPASS, a data framework for tuning large language models to languages. COMPASS uses lean fine-tuning by training lightweight, language modules on a picked subset of multilingual data. By favoring extra data from under-seen clusters, COMPASS lifts cross-lingual transfer while cutting clash. He extends this into an ongoing framework, COMPASS-ECDA, which watches for data shifts and updates adapters to prevent model decay. Across three model architectures and multiple multilingual benchmarks, including long-context tasks, COMPASS always beats standard methods guided by language similarity.

Catalog record

CREATORNoah Flynn [email protected]
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE179670
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2604.20720

Browse the full library · About the Tower of Babel