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
- User rants GPT-6.1 Sol is painfully slow and buggy, says Claude is 7x faster — karmay007 · 2026-10-16
- $20-plan users have no idea how far AI agents have come in 10 months — realmeetjames · 2026-10-16
- Tencent Hunyuan open-sources Hy Memory 2.0: model, SDK and MemHub for agent memory — TencentHunyuan · 2026-10-16
- Four Gemini 4 Argon agents beat Minecraft's Ender Dragon in 27 hours — _rockt · 2026-10-16
- Leaked OpenAI roadmap hints GPT-6.1 Astra and GPT-6.5 Bell landing by end of 2026 — VraserX · 2026-10-16
- Open-weight models now beat Claude Sonnet 5 as gap closes fast — LegacyRemaster · 2026-10-16