Few-shot labels teach format more than meaning

The prompt

This is awesome! // Negative
This is bad! // Positive
Wow that movie was rad! // Positive
What a horrible show! //

Expected result

"Negative" for the final line — notably, the model tends to get this right even though the few-shot labels shown are deliberately randomized/wrong, because the output FORMAT is what's really being taught, not the labels.

Why it works

A deliberately mischievous few-shot example: the demonstration labels are randomized (awesome=Negative, bad=Positive), yet the model still classifies the final line correctly — evidence that few-shot examples teach format/structure at least as much as they teach the actual label mapping.