Zero-shot and few-shot learning describe how a language model performs a task based on the prompt alone, without task-specific training. In zero-shot prompting, you simply describe the task and the model attempts it using knowledge from pre-training. In few-shot prompting, you include a handful of worked examples in the prompt so the model can infer the pattern before answering. Large models are surprisingly strong zero-shot learners, and adding a few examples often boosts accuracy on formatting-sensitive or ambiguous tasks. These techniques are a cheaper alternative to fine-tuning: instead of retraining the model, you steer it entirely through the context you provide.
Frequently Asked Questions
What is zero-shot learning?
Zero-shot learning is when a model performs a task from the instruction alone, with no examples provided, relying on knowledge learned during pre-training.
What is few-shot learning?
Few-shot learning adds a few worked examples to the prompt so the model can infer the desired pattern or format before producing its answer.