<think> and </think> usage during inference

Hello,

I noticed that during inference the reasoning outputs are wrapped with and .
However, from reading the code I don’t see any place where these tags are added as special tokens in the tokenizer.

My questions are:
• Are and actually registered as special tokens in the model/tokenizer?
• Or are they just plain text strings used to structure the model’s output, without being treated differently from normal tokens?

I want to confirm whether these tags are part of the tokenizer’s vocabulary or simply a formatting convention applied to the generated text.

Thanks!

No file, test, or entry point is named. Start by tracing the tokenizer configuration and inference output handling, then determine whether the tags are registered tokens or plain generated text. Done means documenting a clear answer to both questions, with the relevant code locations identified.

Written by the indexing model from the issue text.