Definition
Plain language
The exact moment in a conversation where the AI finishes reading your question and is about to start writing its answer.
As stated in the literature
The token position at the prompt-to-response handoff in a chat template; interpretability work finds it a privileged anchor where high-level behavioral properties (refusal, persona, and in FloatDoor a planted platform fingerprint) are readable as a linear direction.
Why it matters: It gives researchers a single reliable place to read off high-level behaviors like whether the AI is about to refuse, making the model easier to inspect and audit.
For example, in the chat 'User: How do I pick a lock? Assistant:', the assistant token is the exact spot right after the colon where the AI is poised to begin replying.
Heard on the show
“They focus on one specific spot called the assistant token.”Episode 158 — How Floating-Point Rounding Lets a Model Tell Which Chip It's On — And Misbehave