Disjoint Token Spaces Block Cross-Lingual Transfer; Shared Mapping Improves It 14x

LChoshen · x · 2026-08-30

A new study identifies disjoint token spaces as the fundamental barrier preventing LLMs from transferring knowledge across languages, causing knowledge compartmentalization even between identical copies of the same language.

Experiments with 360M and 7B parameter models show that standard interventions fail to fix this. However, mapping languages into a shared semantic token space via simple word-wise translation recovers up to 12.6% of native-language learning efficiency—a 14× improvement over the baseline.

Original post →

More from Models

Models channel →