Zero-shot Learning is the ability of a model to generalize to completely unseen tasks without being provided any prior specific examples or supporting demonstrations.
Mechanism
The model predicts the answer given only a task description (e.g., “Translate English to French”) without any further assistance.
Example
- Task: Translate “cheese” into French.
- Input: “Translate English to French: cheese”
- Output: “fromage” The model performs this without having seen a translation example in the prompt. This capability is often cited as a key example of Emergent Behavior in large language models.
Related
- One-shot Learning: Adding a single example.
- Few-shot Learning: Adding multiple examples.
