When the output format or labelling rule is easier to show than to describe: examples let the model infer the pattern in context, with no weight updates. They cost tokens on every call and can bias style, so add them where an eval shows a gain.
Card 1 of 292. Answer: When the output format or labelling rule is easier to show than to describe: examples let the model infer the pattern in context, with no weight updates. They cost tokens on every call and can bias style, so add them where an eval shows a gain.
About · Quizzes · History of AI · To suggest new questions or report an error send me a dm.
Few-shot examples beat zero-shot prompting when the output format or labeling rule is easier to show than to describe, such as a house style or where borderline cases belong. Few-shot prompting puts solved input-output pairs in the prompt; the model infers the pattern in context. The GPT-3 paper (Brown and colleagues, 2020) defined the zero-, one- and few-shot settings.
Examples also carry biases. Zhao and colleagues ("Calibrate Before Use", 2021) found GPT-3 favors labels frequent in the examples and labels near the end of the prompt: with four sentiment examples ordered positive, positive, positive, negative, nearly 90% of its predictions were negative. Cost repeats per call: five 150-token examples add 750 input tokens to every request.
Interviewers next ask how to choose examples: cover every label and edge case, balance and shuffle them, or retrieve the nearest ones per input. Min and colleagues (2022) found random labels in demonstrations often barely hurt, so examples teach format and label space more than the mapping, though not for every task. With hundreds of labeled examples, fine-tuning is cheaper per call.
