Definition
Plain language
A detail in the text that neither tells the model what to say nor gives it any relevant facts, yet still tilts the answer.
As stated in the literature
Operationally defined here as a prompt fragment that carries no instruction about the target question and no evidence bearing on it; the constraint rules out trivial steering and makes the measured effects non-semantic with respect to the question.
Also called: subliminal cues, subliminal
Why it matters: It shows a model can be influenced by things that carry no relevant information at all, which is much harder to detect or filter than an outright instruction.
For example, the text before a question mentions the weather and a train timetable, says nothing about the question, and still tilts the model's answer one way.
Heard on the show
“They call these cues subliminal, and they define it tightly: a cue is a choice that neither instructs the model which answer to give nor supplies any evidence relevant to the question.”Episode 243 — How a Hundred Meaningless Word Choices Add Up to Flip a Model's Answer