Zero-shot Learning
Last Updated: July 29, 2026 | By Mihail Sebastian | AI Dictionary
A model's ability to handle classes or tasks it never saw in training, using descriptions, attributes, or instructions in place of labeled examples.
What is Zero-shot Learning?
Zero-shot learning is a model’s ability to handle classes or tasks it never saw during training, based on descriptions, attributes, or instructions rather than labeled examples. The idea predates large language models, but LLMs made it everyday practice.
Ask a model to “classify this review as positive or negative” and it complies with zero training examples for that task. The instruction in the prompt stands in for the training data.
How Zero-shot Learning Works
Zero-shot systems map inputs and class descriptions into a shared semantic space. During training, the model links what things look like (or how text reads) to how they are described – attributes such as “striped” or “four-legged”, or full natural-language definitions.
At inference, an unseen input lands in that space and is matched to the description of a class the model never trained on. For large language models the mechanism is simpler to state: pretraining on vast text taught the model enough about most tasks that a clear instruction is all it needs.
Zero-shot vs Few-shot Learning
The practical difference is the number of examples: zero-shot gives the model none, few-shot learning gives it a handful. Zero-shot relies entirely on descriptions or instructions; few-shot lets the model calibrate on real samples of the expected input and output.
| Zero-shot learning | Few-shot learning | |
|---|---|---|
| Examples of the new task | None | A handful |
| What the model relies on | Descriptions, attributes, or instructions | The examples given, plus prior knowledge |
| When to use it | No labeled data exists | A few labels are available or cheap to create |
| Typical accuracy | Lower; hinges on how well the description fits | Higher; examples pin down the expected output |
Example of Zero-shot Learning
Train an image classifier on horses, tigers, and pandas, and store an attribute description for each class. Then hand it a zebra, an animal it has never seen.
The model reads the zebra’s features (horse-like shape, black-and-white stripes) and matches them against the stored description “a horse-like animal with stripes”. It labels the image “zebra” without a single zebra in its training set. The description did the work that training examples normally do.
Related AI terms: Few-Shot Learning · Transfer Learning · Pretrained Model · Word Embedding
Did you like the Zero-shot Learning gist?
Learn about 250+ need-to-know artificial intelligence terms in the AI Dictionary.
Mihail Sebastian — Writes about AI governance, regulation, and the technology behind them. Placeholder bio — replace with a real credential line. About