MATH · IN · MODELS

NLLB semantic offsets stay consistent across 135 languages

measured in 1 paper

- In NLLB-200-distilled-600M, the semantic difference vector for a concept pair (e.g. fire-water) is a single linear offset that stays consistent across typologically diverse languages. [mathewson-2026-nllb-multilingual-geometry] - Averaging per-language offsets into a centroid and scoring each language's offset by cosine to it gives a mean consistency of 0.84 (range 0.70-0.94 over 22 concept pairs; 101 Swadesh concepts x 135 languages). [mathewson-2026-nllb-multilingual-geometry] - Consistency correlates weakly with language phylogeny (Spearman rho=0.13, p=0.020); measured with PCA and All-But-The-Top isotropy correction over 12 encoder layers; observational. [mathewson-2026-nllb-multilingual-geometry] - The geometry is a single cross-lingual linear offset (a direction), not a four-point parallelogram/analogy quadruplet. [mathewson-2026-nllb-multilingual-geometry]

Context

semantic offset vectors, word analogy, cross-lingual universality, parallelogram/crystal structure

Papers

Universal Conceptual Structure in Neural Translation: Probing NLLB-200's Multilingual Geometry — Mathewson, Kyle2026 · arXiv:2603.02258