Language Proficiency as a Recoverable Dimension in Multilingual LLM Embeddings
Proceedings of the Joint Workshop on Readability and Text Simplification (READIxTSAR) @ LREC 2026
Abstract
Understanding whether proficiency is encoded as structured knowledge rather than inferred from surface correlates is critical for interpreting and applying LLMs in educational contexts. We investigate whether multilingual large language model (LLM) embeddings encode language proficiency as a structured recoverable dimension rather than merely supporting predictive classification. Using the UniversalCEFR benchmark, which spans 13 languages and the full proficiency range from A1 to C2, we evaluate the frozen LLM embedding space in two complementary ways. First, we test whether proficiency levels can be predicted directly from frozen embeddings across languages and model variants. The results show that embeddings without task-specific fine-tuning consistently support CEFR classification. Variation in results is strongly associated with the amount of annotated data and language family, suggesting that data availability and cross-linguistic structure matter more than architectural differences. Second, we examine how CEFR levels are organized inside embedding space. We find that texts from lower to higher proficiency levels align along a consistent ordered direction, with higher levels systematically positioned further along this gradient. Distances between levels increase proportionally to their ordinal gap (e.g., A1 vs. C2 is farther apart than B1 vs. B2), indicating a continuous proficiency continuum rather than arbitrary clusters. Together, these findings show that CEFR is not only predictable from multilingual LLM embeddings but is also internally structured as an ordered representational dimension.