While large language models (LLMs) continue to grow in scale and gain new zero-shot capabilities, their performance for languages beyond English increasingly lags behind. This gap is due to the "curse of multilinguality," where multilingual language models perform worse on individual languages than a monolingual model trained on that…