Large language models (LLMs) like those powering ChatGPT are known for their impressive abilities, but they also have some curious blind spots. Researchers discovered a fascinating gap in how LLMs process long texts: they can often "know" where the right information is, but fail to actually "tell" you about it. Imagine having the answer on the tip of your tongue, but not being able to quite articulate it. This is essentially what happens to LLMs with lengthy inputs. This "know-but-don't-tell" phenomenon, as researchers call it, arises from how LLMs prioritize information. LLMs have a kind of positional bias, meaning they pay more attention to information at the beginning and end of a passage, often overlooking key details in the middle. This is like trying to read a long article and only remembering the introduction and conclusion. Through clever probing techniques, researchers peeked inside LLM representations. They discovered that the models do, in fact, encode information about the position of the correct answer, even when they fail to provide it in their output. It's as if the LLM has a map to the information, but struggles to follow it to produce the correct answer. The study also showed that when the correct information is buried deep within a long text, LLMs need to process more layers to access it. This delay in information retrieval can ultimately lead to less accurate responses. Think of it like searching through a cluttered filing cabinet—the deeper the information is buried, the longer it takes to find, and the less likely you are to use it efficiently. These insights are crucial for improving how LLMs process information in the future. By understanding why LLMs struggle with long texts, we can develop new methods to help them overcome this challenge and unlock their full potential. Perhaps one day, we can help these powerful AI systems to find and share the right information every time, regardless of where it’s hidden within a text.