One-Shot

One-Shot describes a situation in which a computer program completes a task after exactly one example. The word comes from English and roughly means “a single attempt.” You show the program a sample solution and expect it to transfer the rest on its own. One example: you write “Berlin belongs to Germany” and then ask about Paris. The correct answer would be “Paris belongs to France.” Alongside this there are the related terms Zero-Shot, meaning no example at all, and Few-Shot, meaning a few examples.

Why a single example changes so much

In the past, a program had to be shown thousands of examples for every new task. Only after that could it solve the task reliably. This elaborate process of learning from many examples is called training. Anyone who has to collect, sort, and label data loses weeks of time and a lot of money.

Modern language models no longer need this for many tasks. During their original training they saw so much text that a single pattern is often enough. The model doesn’t actually learn the one example in any lasting sense. It merely serves as a hint about which format and style are intended.

For companies this is an enormous cost factor. A new task then no longer requires training runs, just a few lines of text. That’s why news articles about AI models often report how well they perform in the one-shot case. This figure is considered a measure of how flexible a model is.

The one example as a blueprint

Technically, nothing mysterious happens here. The example simply sits within the input, that is, within the text sent to the model. The model reads this text from start to finish. Then it predicts, word by word, how it makes sense to continue. Because the example comes directly before it, the prediction takes its cue from it.

You can picture this like a template on a form. Above the empty field sits a filled-in sample. No one explains the rules, but everyone recognizes what is meant. A one-shot example in text works exactly the same way.

A common misconception is the assumption that the model remembers the example. It does not. After the conversation ends, the hint is gone again, and the model’s internal values remain unchanged. Anyone wanting a lasting adaptation needs actual retraining, known as fine-tuning.

One-Shot in chatbots and benchmarks

In everyday life, you encounter this principle as soon as you give a chatbot a pattern to follow. Attaching an example of the desired summary usually yields a more fitting result. The same applies to tables, translations, or fixed text formats. In word-processing programs, such an example is often built in and invisible to the user.

In technical articles and company presentations, One-Shot appears as a test condition. Models are then compared with zero, one, or several examples. The differences show how heavily a model relies on assistance. A strong model loses only a little performance between Zero-Shot and One-Shot.

The term is also used in image recognition, for instance with face unlocking. There, a single reference photo is often enough to recognize a person later on. The same basic idea is meant: deriving a general rule from one example. However, this is only reliable if that one example is truly typical.