Zero-Shot Learning
What is it?
Zero-shot learning is asking a model to do a task with no examples — just an instruction — relying on knowledge it already picked up during training.
Explain like I'm 5
Why was it created?
Providing examples costs effort and context space. Large models turned out to follow plain instructions well, so zero-shot became the simplest way to get work done.
Where is it used?
- Quick classification and labeling
- Summarization and rewriting
- Simple extraction
- Prototyping prompts
Why should developers care?
It's the default, lowest-effort way to prompt — knowing when it's enough (vs. needing examples) saves tokens and time.
How does it work?
You give the model an instruction describing the task and the input, with no worked examples. The model applies general knowledge and instruction-following ability learned in training to produce an answer.
Real-world example
'Classify this review as positive or negative: "The battery dies too fast."' — the model answers correctly with no examples provided.
Common use cases
- Straightforward tasks a model already understands
- Saving context space
- Fast prototyping
- Common summarize/translate/classify jobs
Advantages
- Simplest possible prompting
- No examples to craft
- Saves context and cost
- Instant to try
Disadvantages
- Less reliable on niche formats
- No control over exact style without examples
- Quality varies by task
- Ambiguous instructions hurt more
When should you use it?
For common, well-understood tasks where a clear instruction is enough.
When should you avoid it?
When output must follow a specific format or the task is unusual — add examples (few-shot) or fine-tune.
Alternatives
Related terms
Interview questions
Beginner
- What makes a prompt 'zero-shot'?
- When is zero-shot enough?
Intermediate
- Why can large models do zero-shot tasks at all?
- How do you decide between zero-shot and few-shot?
Senior
- How does instruction tuning improve zero-shot performance?
- What are the failure modes of zero-shot on structured-output tasks?
Common misconceptions
- "Zero-shot means the model was never trained on anything related" — it draws on broad training knowledge.
- "Zero-shot is always worse than few-shot" — for common tasks it's often just as good and cheaper.
Fun facts
- Instruction tuning dramatically improved models' zero-shot ability.
- 'Shots' refers to in-prompt examples — zero-shot simply has none.
Timeline
- 2020 — Large models show useful zero-shot behavior
- 2022 — Instruction tuning boosts zero-shot performance
Learning resources
Quick summary
Zero-shot learning asks a model to perform a task from an instruction alone, no examples, using knowledge learned in training.
Cheat sheet
- Instruction only, no examples
- Default lowest-effort prompt
- Great for common tasks
- Add examples for niche formats