Definition
Plain language
A meaningless pattern pushed into the model's internals, used to check whether the real pattern is special.
As stated in the literature
Norm-matched control vector injected in place of a semantically identified direction; separates effects of intervention magnitude and generic disruption from effects specific to the target representation.
Also called: random directions
Why it matters: Without this control, almost any internal nudge can be mistaken for evidence about a specific concept when it is really just generic disruption.
For example, if pushing a meaningless pattern of the same strength makes the model ramble just as much as the supposed pain pattern did, the effect was about the shove, not about pain.
Heard on the show
“The experiment includes a random direction of matched overall strength, alongside the pain direction.”Episode 267 — A Pain Axis, a Relief Button, and the Control the Paper Skipped