Large Language Models (LLMs) are often tested using multiple-choice questions. But what if these tests don't accurately reflect what the AI knows? New research suggests these models possess hidden knowledge, obscured by the limitations of current testing methods. Imagine a student who understands complex concepts but struggles with the rigid format of a multiple-choice exam. They might know the answer but fail to select the correct letter. LLMs face a similar challenge. Researchers have uncovered a fascinating insight: by examining the "attention" mechanism within LLMs—how they focus on different parts of a question—we can get a clearer picture of their true understanding. This new technique, focusing on specific "select-and-copy" attention heads, reveals how LLMs identify and process relevant information. The results are impressive. This method boosts the performance of smaller LLMs by up to 16% on standard tests and a remarkable 60% on a synthetic test designed to bypass formatting issues. Even larger, more sophisticated models show gains, particularly when understanding the nuances of language and context is key. This discovery is more than just a test score booster. It provides a window into the inner workings of LLMs, offering clues about how they reason and learn. It suggests that while outputting the right letter can be tricky, these models grasp the underlying concepts more effectively than we thought. The implications are exciting. This research opens doors to unlock even more of LLMs’ potential, paving the way for more robust and transparent AI systems.