Definition
Plain language
A quick fix that reworks and reshuffles the wording of a question across repeated model calls so its answers stop landing on the same single option.
As stated in the literature
A patch for persona-based silicon sampling that perturbs surface prompt features (rewording, option reordering) across calls to inject output variety, cutting distributional error at near-zero cost.
Also called: Argyle
Why it matters: This trick matters because it cheaply restores variety in a model's answers when it would otherwise collapse onto a single repeated choice.
For example, instead of asking the exact same question ten times, you slightly reword it and shuffle the answer choices each time so the model stops parroting one identical response.
Heard on the show
“They call it prompt-perturbed Argyle, and it cuts the error about twenty-one percent, at basically zero added cost.”Episode 230 — Why AI Survey Panels Break Before the Dice Ever Roll