Large language models (LLMs) are impressive, but they sometimes struggle to consistently use the information they're given. Imagine an LLM trying to answer a question using several web pages. It might get distracted by the position of a sentence on a page rather than focusing on the actual relevance of the information. This "position bias" can lead to inaccurate or inconsistent answers. Researchers have developed a clever technique called CORD (Balancing COnsistency and Rank Distillation) to help LLMs overcome this challenge. CORD works by training the LLM to be consistent in its answers, even if the order of information it receives is shuffled around. It's like teaching an LLM to recognize that the most important fact can appear anywhere on a page, not just at the top. However, simply shuffling information randomly can sometimes be counterproductive. If the original order of information was actually helpful (like search results ranked by relevance), random shuffling can make the LLM lose valuable clues. CORD addresses this by intelligently controlling the amount of shuffling. It learns to balance the benefits of consistency training with the importance of respecting the original order when it's actually meaningful. The results are impressive. CORD significantly improves the accuracy of LLMs in various question-answering tasks, showing that it can effectively reduce position bias while still leveraging helpful ranking information. This research is a significant step toward making LLMs more robust and reliable for real-world applications like fact-checking and knowledge retrieval. Future work could explore how these findings generalize to different LLM architectures and how to further refine the balance between consistency and rank preservation.