Large language models (LLMs) like ChatGPT have taken the world by storm with their impressive text generation capabilities. But beneath the surface lies a fascinating question: what knowledge do these models truly possess? A new study from Northeastern University digs into this by exploring the "implicit vocabulary" of LLMs. It turns out that the way these models break down and represent words is more complex than it seems. LLMs process text using tokens, which are often fragments of words. For example, "northeastern" might get split into meaningless chunks like "_n," "ort," "he," and "astern." So how do LLMs manage to understand the actual meaning of words when they're working with such fragmented pieces? The researchers discovered a curious phenomenon called "token erasure." They found that in the early processing stages of an LLM, information about individual tokens within a word gets "erased" or forgotten. This erasure, they believe, is a footprint of the model forming an internal representation of the *actual* word, not just the individual tokens. Think of it like the model piecing together puzzle fragments to construct a complete picture of a word's meaning. The researchers developed a method to probe this implicit vocabulary, scoring token sequences based on how strongly they exhibited this erasure effect. When applied to models like Llama-2 and Llama-3, the method uncovered a hidden lexicon of multi-word expressions, named entities, and even LaTeX commands. Interestingly, the vocabulary of each model reflects its training data. Llama-3, trained on a larger, more code-heavy dataset, has a richer implicit vocabulary of code-related terms. This research sheds light on the intricate mechanics of how LLMs process and represent language. It also provides a promising new tool for understanding what words and concepts a given LLM truly understands. This deeper knowledge can help refine future models, improving their accuracy and enabling them to tackle more complex tasks. While this research offers a glimpse into the black box of LLMs, the quest to fully understand their knowledge representation continues. Future research could explore how these implicit vocabularies evolve during training and whether they contribute to the model's ability to generalize to new, unseen words and concepts.