Zero-shot is when an LLM performs a task purely from the instruction, with no examples in the prompt — like saying "translate this from Turkish to English" and supplying no sample translations. The term was popularized by the GPT-3 paper in 2020, which showed models could perform classification, translation, and summarization at a certain level out of the box. Compared with Few-shot it is more minimal, but with a well-crafted instruction modern frontier models are often competitive on most tasks. It is the limiting case of In-context Learning.
MEVZU N°124ISTANBULYEAR I — VOL. III
Glossary · Beginner · 2020
Zero-shot
When the model performs a task with no examples, given only the instruction.
- EN — English term
- Zero-shot
- TR — Turkish term
- Sıfır-Atış (Zero-shot)