A developer is delegating a classification task to a foundation model and wants to try doing it without incurring the cost of preparing examples. What is the prompt technique called in which you give only the description of the task you want done, without showing any examples, and have it run?
Select the name of the technique that runs with only a task description and no examples.
Few-shot prompting
Few-shot prompting is a technique that shows a few input-output examples first and then has the task run. For example, it shows a few samples like 'It was the best → positive / It was terrible → negative / I will never buy again → ?' before making the judgment.
This question asks for the name of the case where no examples are shown at all; few-shot, which provides examples, is the opposite, so it is incorrect.
Chain-of-thought prompting
Chain-of-thought is a technique that has the model write out intermediate reasoning steps to raise accuracy on complex problems. For example, it instructs 'Before giving the final answer, write out each step of the calculation in order,' having it think step by step.
It is not a matter of whether examples are present, so it is incorrect.
Zero-shot prompting
This is correct. Zero-shot prompting is a technique that gives no examples at all and has it run with only the task description (instruction). For example, without any samples, it instructs only 'Classify the following review as positive or negative: "I am satisfied the package arrived quickly."'
Negative prompt
A negative prompt is a technique that specifies elements you do not want generated (mainly in image generation). For example, in image generation, 'Draw a picture of a cat (negative prompt: text, watermark, low quality),' specifying unwanted elements separately to exclude them from the result.
It is not the name of the technique of running with only a task description without showing examples, so it is incorrect.
Remember the correct answer, 'zero-shot prompting.'
- A technique that gives no input-output examples at all and has the model run with only the task description (instruction).
- It is easy, but on hard tasks accuracy can be lower than few-shot, which shows examples.
Fine-tuning and continued pre-training (which involve training) and RAG (which searches external knowledge) are separate concepts from how a prompt is written.