Cross-Lingual Instruction Alignment in Large Language Models via Lightweight Prompt Distillation

With the continued expansion of large language models in multilingual tasks, achieving efficient and robust instruction alignment has become a key technical challenge in the field of natural language processing. This study proposes a lightweight instruction fine-tuning framework that combines cross-lingual transfer learning with a hierarchical prompt distillation strategy. The framework first performs initial optimization on the model using high-quality English instruction data. Then, through a carefully designed hierarchical prompt structure, knowledge is distilled and transferred to models in low-resource languages. The goal is to ensure consistency in instruction responses and accurate semantic alignment in multilingual settings. Experiments on the XGLUE and FLORES-101 benchmarks show that the proposed method achieves an average alignment accuracy of 92.3% across 12 languages, while reducing training costs by 34% compared to reinforcement learning-based methods.