Large language models (LLMs) sometimes generate incorrect or nonsensical information, often called "hallucinations." But what if these errors aren't truly hallucinations, but rather glimpses into a deeper, more complex internal understanding? New research suggests LLMs might actually "know" the correct answers even when they output something different. Researchers dove into the inner workings of several LLMs, examining their internal representations during the process of generating text. They discovered a fascinating pattern: information about the truthfulness of an answer is highly concentrated within the specific tokens that make up the answer itself. By focusing on these "exact answer tokens," the researchers significantly improved their ability to detect errors in the LLM's output. This suggests that the models have an internal representation of truth that's stronger than we previously realized. But there's a twist. This internal knowledge doesn't always translate to the final answer. The researchers found cases where an LLM encoded the correct answer internally but consistently generated the wrong one. This disconnect between internal representation and external output is a key puzzle. One possibility is that LLMs, trained to generate the most statistically likely text, might sometimes prioritize common phrasing over factual accuracy. This raises the intriguing question: can we tap into this hidden knowledge to help LLMs overcome their tendency to generate inaccurate information? The research also revealed that LLMs don't have a single, universal way of representing truth. Instead, the way they encode truthfulness varies depending on the task at hand. A model might be great at detecting truth in factual questions but struggle when it comes to, say, math problems. This implies that there isn't a one-size-fits-all solution for fixing LLM errors. We need to develop targeted strategies based on the specific type of error. The ability to predict the kinds of errors an LLM is likely to make, based on its internal states, opens exciting possibilities for more nuanced and effective error mitigation techniques. By understanding how LLMs represent truth and error internally, we can begin to develop strategies to help them more reliably generate accurate and truthful information, bringing us closer to more reliable and trustworthy AI.