Zero shot

Zero shot is a method of letting AI do the work right away without giving any examples.

Zero-shot is a method of asking AI to perform tasks it has never encountered for the first time with only instructions without showing any demonstration examples. This can be likened to a situation where an employee is asked to “divide these reviews into positive and negative” without any samples, and the employee does it right away. The method of showing a few examples and asking an employee to do so is called few-shot.

In the past, AI had to be trained separately with dedicated learning data for each task, but large language models showed the ability to perform tasks they had never learned just by following instructions, thanks to the general knowledge accumulated through prior learning. This zero-shot ability is the secret to being able to use a single model for all kinds of tasks, such as translation, summarization, and classification, and is also used as an evaluation standard to measure the model's versatility.

However, in specialized or difficult tasks, zero-shot performance can vary, so few-shot or fine-tuning that adds examples is often better.

✅ Why it matters

• You can use AI right away without task-specific learning data
• It is the basis for versatility to immediately apply one model to various tasks
• Used as an evaluation index to compare the generalization ability of the model

⚠️ Limits and debates

• Accuracy is unstable in specialized or demanding tasks
• Performance is often lower than few-shot examples provided
• If instructions are ambiguous, it is easy to produce results that are different from the intention.