Large language models (LLMs) are getting smarter, but they still struggle with long documents. Think of trying to find specific information in a massive, sprawling document. You might miss crucial details if they're buried deep within the text. LLMs face a similar challenge. New research reveals that it's not just *where* information is located in a long document that trips up LLMs, it's how that information is *spaced out*. Researchers created a benchmark called LongPiBench to test this. They discovered that while LLMs are now better at finding information regardless of its absolute position in the document (e.g., beginning, middle, end), they still struggle when multiple pieces of relevant information are spread far apart. This "relative positional bias" causes a significant drop in performance. Imagine trying to piece together a puzzle where the key pieces are scattered randomly across the table. As the pieces become more spread out, the puzzle gets harder to solve. LLMs experience a similar difficulty when synthesizing information from a long document. This research also hints that simply making AI models bigger doesn’t solve the problem. While increasing model size helps with absolute positional bias, the challenge of relative positioning remains. This suggests we need new strategies to help AI navigate and understand complex, lengthy texts. The implications are significant for tasks like data analysis, document summarization, and information retrieval. Future research will focus on understanding *why* this bias exists and developing techniques to overcome it, paving the way for even more powerful and useful AI.