Can zero-shot use retrieved information?

Yes, under a protocol that permits it. Retrieved facts are not necessarily completed demonstrations, although retrieved input-output examples would change the prompting condition.

Is zero-shot the same as unsupervised learning?

No. A zero-shot model may have received extensive supervised training on other tasks or classes.

Does zero-shot mean one attempt?

No. It describes the supplied example condition, not the number of generations, retries or evaluation runs.

AI Glossary / Prompting and human-AI interaction / Intermediate

Zero-shot

Definition

Zero-shot describes using a model on a target task or category without supplying task-specific demonstrations or labeled training examples under the stated evaluation protocol. The precise meaning depends on context: zero-shot prompting omits demonstrations in the prompt, while zero-shot classification typically concerns classes absent from labeled training data.

What it means

The zero refers to a particular source of examples, not to all learning. A zero-shot system usually relies on substantial prior training and may receive a task description, class names, attributes or reference material.

In language-model prompting, asking for a translation without showing any completed translations is commonly called zero-shot. In a visual classification study, recognizing a new class from its description without labeled images of that class is another zero-shot setting.

These claims are not interchangeable. A prompt can contain no examples even if the model encountered similar tasks during pretraining. Conversely, a classifier can receive detailed semantic descriptions of unseen classes while remaining zero-shot under its declared class-training protocol.

How it works

A model transfers capabilities or representations learned elsewhere to the target situation. A language model may interpret an instruction through patterns acquired in pretraining and instruction tuning. A class-recognition model may compare an input representation with descriptions or attributes of candidate classes.

The evaluation must specify what information is allowed. Relevant questions include whether task-specific training occurred, whether labeled examples are present in the prompt, whether unlabeled target data is used and whether class descriptions are available.

Performance is measured on independent target cases with an appropriate scoring rule. The term zero-shot does not itself specify the mechanism, output quality or computational cost. It describes an information or adaptation condition that makes an evaluation result interpretable.

Three practical examples

A language model is asked to rewrite a supplied sentence in simpler language without being shown a completed rewrite.
An image-language model compares a photograph with descriptions of candidate categories without additional labeled training images for those categories.
A multilingual system is evaluated on a translation direction that was not included as direct paired training data, under a protocol that documents the other languages and data it used.

Think of it this way

Think of receiving a new assignment with written instructions but no worked example of that assignment. You still rely on what you have learned previously.

The analogy clarifies why zero-shot does not mean zero knowledge. It also highlights a limitation: existing knowledge may transfer well to one assignment and poorly to another, so the absence of demonstrations says nothing by itself about success.

Worked example

Consider an original text-labeling task. The instruction says: "Label each note as request, statement or unclear. Use request when the writer asks someone to do something; use statement for an assertion without a request; use unclear when neither can be established."

The input is: "Please send the attendance list before noon." The intended answer is request. No completed input-output demonstration appears in this prompt, so it is a zero-shot prompting setup. The detailed label definitions do not turn them into demonstrations.

Now add: "Example: Please close the window. Answer: request." The setup becomes one-shot because it includes one completed example. It does not matter that the original instruction is unchanged.

Suppose a hypothetical evaluation uses fifty fresh notes. The zero-shot version gets thirty-eight correct, or 76%. The one-shot version gets forty correct, or 80%. The observed difference is four percentage points on that collection, not a claim that every added demonstration improves performance.

Neither score proves that the model never encountered similar request-classification tasks during training. The demonstrated claim is narrower: one evaluation supplied no completed examples in its inference prompt, and the other supplied one. These invented counts illustrate the importance of defining the zero-shot condition.

Illustrative case study

Fictional case: Elm Documentation assesses whether a general model can classify new internal notes without creating a dedicated training dataset. The team begins with label definitions and a zero-shot prompt.

It records uncertain and mixed-purpose notes separately, rather than treating a confident label as evidence of correctness. After examining development errors, it compares added demonstrations and a simple rule-based baseline on held-out cases.

The team describes its result as zero-shot prompting, not proof of learning a completely unseen concept from nothing. This hypothetical case illustrates precise reporting and claims no measured operational advantage.

Common misconceptions and mistakes

Zero-shot means the model was never trained: prior training usually supplies the transferable capability.
Zero-shot means no instructions or context: descriptions and source material can be allowed.
No examples in the prompt proves no benchmark exposure: training-data overlap is a separate question.
Zero-shot prompting and zero-shot class learning are identical: their definitions of excluded examples differ.
Zero-shot is automatically preferable: task quality, reliability and cost determine whether a setup is useful.

Business use

A zero-shot baseline can help determine whether an existing model already supports a task before investing in labeled data or adaptation. It is also a useful comparison point for measuring the value of demonstrations, retrieval or fine-tuning.

Document exactly what was supplied and what was withheld. Evaluate realistic cases, including ambiguous inputs and categories with little relevant prior knowledge. Avoid treating a zero-shot label as a quality certification. If a more adapted method performs better enough to justify its costs, zero-shot simplicity may not be the decisive consideration.

Check your understanding

A prompt includes a detailed category definition but no completed example. Must it be called one-shot?

Answer: No. A definition is not itself a demonstration. It can remain zero-shot prompting, although the evaluation should document the information supplied.

Related terms

Zero-shot prompting
Zero-shot learning
Few-shot learning
Transfer learning
Data contamination
in-context learning
class description
task adaptation
evaluation protocol

Learn next

Write a precise statement of what zero-shot means for one task. List allowed instructions, descriptions, reference data and training. Establish a baseline on held-out cases, then compare it with an adapted method without changing the evaluation inputs or scoring rules.

Sources

Brown and colleagues, Language Models are Few-Shot Learners, 2020, zero-shot and demonstration-based evaluation settings: https://arxiv.org/abs/2005.14165

Romera-Paredes and Torr, An Embarrassingly Simple Approach to Zero-Shot Learning, ICML 2015, semantic descriptions for new concepts: https://proceedings.mlr.press/v37/romera-paredes15.html

The note-labelling scores and Elm Documentation case are original hypothetical illustrations.

Previous: Zero Redundancy Optimizer Next: Zero-shot learning