Languages are Modalities: Cross-Lingual Alignment via Encoder Injection

A new method called LLINK has been introduced to enhance the performance of instruction-tuned Large Language Models (LLMs) on low-resource, non-Latin scripts. This innovative approach addresses issues like tokenizer fragmentation and weak cross-lingual coupling without the need for retraining the decoder. By aligning sentence embeddings from a frozen multilingual encoder, LLINK promises to improve the accessibility and effectiveness of language models across diverse languages, making it a significant step forward in the field of natural language processing.