LLMs show lower translation quality on non-English-centric pairs due to reduced use of language-specific tokens, with a new TAR metric linking token activation to performance across 15 models and 22 pairs.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CL 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
Why do Large Language Models Fail in Low-resource Translation? Unraveling the Token Dynamics of Large Language Models for Machine Translation
LLMs show lower translation quality on non-English-centric pairs due to reduced use of language-specific tokens, with a new TAR metric linking token activation to performance across 15 models and 22 pairs.