Prompting
Zero-shot Learning
Also known as: zero-shot prompting
Asking a model to do a task with only an instruction and no examples, relying on what it already learned in training.
Zero-shot means asking the AI to do something using only a plain instruction, no examples at all. "Translate this into French" or "summarise this email" just works, because the model already learned those skills during training. Most everyday use of AI is zero-shot: you simply ask.
Zero-shot prompting gives the model an instruction with no worked examples, and it performs the task from its pretrained knowledge alone. It is the opposite of few-shot, which supplies examples. Modern instruction-tuned assistants are strong zero-shot performers, which is why "just ask clearly" usually works.
When zero-shot is the right choice
- The task is common and well-understood (translation, summarising, rewriting).
- You want a short prompt and low cost (no examples taking up space).
- You do not have good examples to hand.
When to add examples instead
If the output format is very specific, or results are inconsistent, switch to few-shot and show what "good" looks like.
Zero-shot denotes task execution from an instruction with no in-context demonstrations, leaning entirely on knowledge and skills acquired in pretraining and, crucially, on instruction tuning, which sharply improved zero-shot instruction-following. It relies on the same in-context learning substrate as few-shot, minus exemplars. A notable result is that a simple trigger ("let's think step by step") elicits zero-shot chain-of-thought reasoning, improving multi-step accuracy without examples (Kojima et al. 2022).
The zero-shot vs few-shot boundary has shifted with model capability: as alignment tuning matured, zero-shot closed much of the gap on many benchmarks, making few-shot most useful for strict formatting and specialised tasks. Trade-offs are token economy (zero-shot is cheaper, leaving more context window) versus controllability (examples reduce ambiguity). Evaluation should distinguish genuine zero-shot from cases where task exposure during pretraining inflates apparent generalisation.