Imagine a mountain of digital paperwork, pages shuffled like a deck of cards after a rough game. That's the challenge businesses face daily when dealing with streams of documents. Researchers call this the Page Stream Segmentation (PSS) problem, and it's a crucial first step in automating document processing. But how do you teach an AI to sort these digital stacks efficiently? A new research paper explores the potential of Large Language Models (LLMs) for PSS, using an improved benchmark dataset called TABME++. Historically, PSS systems relied on handcrafted rules, struggling to handle the variety of documents found in real-world scenarios. Researchers then turned to smaller, specialized AI models, but these still had limitations. The paper investigates whether powerful LLMs, already making waves in other areas of AI, can tackle PSS more effectively. They found that fine-tuned decoder-based LLMs, like Mistral-7B, outshine earlier models, correctly segmenting up to 80% of document streams without any human correction. This success is largely thanks to TABME++, which uses commercial-grade OCR (Optical Character Recognition) to accurately extract text from scanned documents. The quality of the OCR proved vital, showing the importance of clear text for AI understanding. While LLMs show great promise, challenges remain, particularly in handling complex, real-world document streams where not every stream needs segmentation. Future research will focus on these complexities, exploring how to best combine different data modalities (like images and text layout) and making LLMs even more efficient and accurate for document processing.