Large Language Models (LLMs) have shown remarkable progress in problem-solving, but they often struggle when it comes to non-English languages. This linguistic disparity poses a significant challenge, especially in complex reasoning tasks. However, new research suggests a promising solution: question alignment. Instead of relying on costly and often inaccurate translations of training data, this innovative technique leverages an LLM's existing English proficiency to boost its performance in other languages. The process involves a two-step training framework. First, the model undergoes 'question alignment' training, where it learns to connect non-English questions with their English counterparts. This strengthens the internal language alignment within the LLM. The second stage focuses on 'response alignment,' where the model is fine-tuned on English instruction data, effectively transferring its English reasoning capabilities to other languages. This approach has shown impressive results across diverse reasoning scenarios, including mathematical reasoning with chain-of-thought and program-of-thought, as well as common sense reasoning. Furthermore, the method scales effectively to even the largest LLMs, like the 70B parameter LLaMA2 model, significantly improving their multilingual accuracy. Analysis reveals that question alignment creates a more unified semantic space within the LLM, allowing it to process multilingual queries more consistently and apply its English-learned reasoning skills more effectively. This research opens exciting new avenues for developing truly multilingual LLMs, breaking down language barriers, and making advanced AI accessible to a global audience. While challenges remain, particularly in balancing language consistency with reasoning accuracy, this approach represents a significant step towards more inclusive and powerful AI systems.