What is the purpose of special tokens in function calling?

• A. They help structure tool calls and observations in an LLM's response
• B. They increase vocabulary size
• C. They improve numeric precision
• D. They are cosmetic and not processed

Answer: (A)

Special tokens mark tool calls and observations, enabling structured interaction between an LLM and external tools. They tell the model when to invoke a function, what arguments to pass, and how to incorporate the tool’s results back into its reasoning. This separation keeps tool usage explicit and machine-parseable, so the orchestration system can route requests correctly and the model can reason with the actual observed outputs rather than treating them as ordinary text. Without these markers, tool invocations could blur with normal language, leading to misfires or missed tool results.

They’re not about increasing vocabulary, improving numeric precision, or just cosmetic formatting—their purpose is functional: to coordinate and structure how the model calls tools and uses their responses. For example, a token sequence might clearly indicate a tool call, followed by a distinct observation token that introduces the tool’s reply, which the model can then use in subsequent reasoning.