Each language starts with its own embedding cluster. Cross-lingual training pulls translation-equivalent words from different languages closer together in one shared space — the same concept, in any language, should end up near the same point.
pos_i(t) = lerp(pos_i(0), sharedAnchor(concept_i), alignment)
pair_dist = ||emb(word_en) - emb(word_translated)||
zero-shot accuracy ≈ f(1 / pair_dist) (closer pairs -> classifier trained
on one language works on another)
- Languages — how many separate language clusters populate the space.
- Alignment strength — how strongly cross-lingual training pulls translation pairs toward a shared point; 0 = fully separate per-language spaces, 1 = fully merged.
- Show translation links — draws a line between each concept's realization in every language.
- Zero-shot transfer — simulates applying a classifier trained only on English to every other language, reading off the resulting accuracy from pair distance.
This is the mechanism behind multilingual LLMs like mBERT/XLM-R: once concepts share a geometry across languages, a model fine-tuned in one language can often work in a language it saw little or no labeled data for.