Zero-Shot Learning
Zero-shot learning applies a model to target categories without labeled training examples for those categories, using information such as descriptions or attributes. In language-model discussions, zero-shot prompting means giving a task instruction without worked examples in the prompt.
Also known as: ZSL
A recognition model might learn from horse and tiger images while also learning associations with attributes such as stripes and a mane. A description of a zebra can then help it recognize a category with no labeled zebra examples in the training set. The description connects the unseen category to learned visual features.
An LLM asked “Label this review positive or negative” with no demonstrations is instead being used with zero-shot prompting. This describes the prompt, not proof that the model never encountered sentiment labels or similar reviews during training.
The distinction matters for evaluation. For unseen-class recognition, inspect training and pretraining overlap with test categories. For prompting, record which examples and instructions the model received. Neither setup guarantees reliable performance; weak descriptions or an unsuitable model can leave the task unresolved.
Sources
- Xian et al.: Zero-Shot Learning — A Comprehensive Evaluation — Defines unseen-class recognition protocols and highlights leakage from pretraining on test classes.
- Brown et al.: Language Models are Few-Shot Learners — Defines zero-shot language-model evaluation with instructions but no demonstrations.
Go deeper
- Hugging Face: Zero-shot classification pipeline docs
Try classification from candidate label descriptions and inspect the entailment-based scoring setup.