Zero-Shot Machine Translation Using Large Language Models for Cross-Lingual Generalization

Abstract

Zero-shot machine translation (MT) has emerged as a critical capability for enabling communication across unseen language pairs without direct supervised training. This study investigates the effectiveness of Large Language Models (LLMs) in improving zero-shot translation performance through their inherent cross-lingual generalization abilities. By leveraging pre-trained multilingual representations and contextual understanding, the proposed approach enables accurate translation across diverse linguistic structures without requiring parallel data for each language pair. The analysis explores how LLMs reduce the performance gap between pre-trained and randomly initialized models by capturing semantic and syntactic relationships across languages. Experimental findings indicate that LLM-based approaches significantly enhance translation quality, robustness, and scalability, offering a promising direction for building universal and data-efficient machine translation systems.