Definition
Plain language
The part of an attention unit that decides which earlier words to look at.
As stated in the literature
The query-key circuit of an attention head, computing where attention is directed; in persuasion-circuit work it carries roughly 88% of the effect, while the OV value-copying circuit stays innocent.
Also called: QK
Why it matters: Because it controls where a model focuses, locating a behavior here pinpoints the attention step responsible rather than the information being copied.
For example, when reading 'the cat sat on the mat,' the QK circuit is what decides that 'sat' should pay attention back to 'cat.'
Heard on the show
“That's one circuit, usually called QK, for query and keys.”Episode 038 — How LLMs Get Persuaded: One Attention Head, A Tetrahedron, And A Single Dial