Definition
Plain language
The part of an attention unit that copies information from wherever it's currently looking.
As stated in the literature
The output-value circuit of an attention head, governing what gets written into the residual stream once attention is placed; found to be a near-faithful copy operation in persuasion-circuit analysis, localizing misbehavior to the QK side.
Also called: OV
Why it matters: Knowing this part faithfully copies information helps researchers pin a misbehavior on the part that chooses where to look rather than what gets copied.
For example, once an attention unit decides to look at the word 'Paris,' the OV circuit is what actually carries that word's information forward.
Heard on the show
“That's the other circuit, OV.”Episode 038 — How LLMs Get Persuaded: One Attention Head, A Tetrahedron, And A Single Dial