Zero-shot
Asking an AI to do something new without showing it any practice problems or examples first, relying entirely on what it already knows.
A prompting technique where a model is asked to perform a task based solely on the instruction provided, without including any task-specific input-output pairs in the context window.
Performing a task with no task-specific examples in the prompt, relying solely on the model's pre-trained knowledge (contrast few-shot, which supplies a handful of examples).
evolution
- 2009 · historyZero-Shot Learning (ZSL) Formalization
Lampert et al. introduced the concept of learning to recognize object classes not present in the training set by using semantic attribute descriptors.
- 2017 · historyTransformer Architecture
The 'Attention Is All You Need' paper provided the architectural foundation for models to generalize across tasks without task-specific retraining.
- 2020 · historyGPT-3 Emergence
OpenAI demonstrated that large language models could perform a wide variety of tasks via zero-shot prompting without any gradient updates.