For few-shot, optimize the label space and input distribution — not whether each example label is correct

SurePrompts·medium signal

A counterintuitive but durable result (Min et al.) is that few-shot performance is driven by the label space and input distribution your examples demonstrate, not by whether each individual example is correctly labeled — even randomly-labeled examples can outperform zero-shot. The practical move: spend your effort making 3–5 examples cover the full range of valid labels and representative inputs, rather than agonizing over perfect ground-truth answers.