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.
More from Models
- Liquid AI Unveils LFM2.5-2.6B: A Tiny Agent Model for Raspberry Pi — emmanuelvivier · 2026-08-30
- Writer Claims Palmyra X6 Cuts AI Agent Costs by 52% — emmanuelvivier · 2026-08-30
- Claude Refuses to Change 20MB to 19MB — gerardsans · 2026-08-30
- Hands-on with Tencent Hunyuan Hy4: Execution enters top-tier open source — HeyAmit_ · 2026-08-30
- Opus 5 shocks user by silently completing complex full-stack tasks without chain of thought — PawelHuryn · 2026-08-30
- Commentary on GPT-4.5's language quality and compute scarcity — haider1 · 2026-08-30