What are special tokens in LLMs?

• A. Tokens used only for punctuation alignment.
• B. Tokens that help models understand and structure text, such as sequence delimiters and control markers.
• C. Tokens representing image metadata.
• D. Tokens used for system time stamps.

Answer: (B)

Special tokens act as markers that guide how the model reads and structures text. They carry extra meaning beyond ordinary words, serving to delimit sequences, separate parts of input, and control how the model processes data. For example, there are tokens that signal the start or end of a sequence, separators between segments in a pair of inputs, or padding tokens to normalize length. These markers help the model understand where one part ends and another begins, manage input lengths, and apply attention or formatting rules correctly. They are essential for tasks like question answering, translation, and classification because they provide consistent structure the model can rely on, rather than treating every token as mere content. Other kinds of tokens may exist in specific contexts, but the general purpose of special tokens is to help understand and structure text through sequence delimiters and control markers.