Revisiting Cross-Lingual Alignment in the Era of Multilingual LLM Embeddings

Cross-lingual alignment methods are commonly used to improve multilingual embedding spaces by explicitly aligning representations across languages. However, modern multilingual LLM embeddings may already encode substantial universal semantic structure before explicit alignment. In this work, we evaluate nine families of alignment methods across two multilingual embedding backbones using multilingual sentence pairs spanning 88 languages. Our evaluation combines retrieval performance with geometric similarity analysis through Centered Kernel Alignment (CKA) and semantic separation analysis based on cosine similarity distributions and probabilistic ranking behavior. We find that alignment methods can produce substantially different projection geometries while achieving similar cross-lingual retrieval performance, suggesting that successful retrieval does not imply convergence toward a shared universal geometry. Moreover, raw multilingual embeddings already achieve strong cross-lingual retrieval without explicit alignment, while many alignment methods provide only marginal improvements. We further show that highly anisotropic embedding spaces can nevertheless preserve robust semantic ordering across languages. Together, these results reveal a disconnect between representational geometry and retrieval behavior and motivate evaluating multilingual representations beyond retrieval metrics alone.