Context-Aware Large Language Models for Multilingual Understanding

Multilingual large language models (LLMs) have demonstrated strong performance in cross-lingual tasks; however, their ability to incorporate context across diverse languages remains underexplored. This paper proposes a Context-Aware Multilingual Transformer (CAMT) architecture that integrates dynamic context routing, semantic alignment layers, and cultural knowledge embeddings to enhance multilingual understanding. Experiments conducted using the FLORES-200 and XNLI datasets show that CAMT improves context retention by 12.4%, cross-lingual consistency by 9.8%, and cultural disambiguation by 7.1% compared to baseline mT5 and XLM-R models. Results highlight the importance of contextual cues in multilingual communication and underline the potential for building globally robust LLMs.