Glossary term
Glossary term
Agentic Systems
Instructing a model to perform a task without any in-prompt examples.
A type of machine learning training where the model infers a prediction for a task that it was not specifically already trained on. In other words, the model is given zero task-specific training examples but asked to do inference for that task.
OpenAI's GPT-4 is used zero-shot for medical coding (ICD-10) by a health-tech startup, achieving 84% accuracy on common diagnoses solely from the task description in the system prompt.
Google Gemini is deployed zero-shot by a media company for multilingual news summarisation - the prompt specifies the task and output format but no examples, relying entirely on the model's pretraining.
Anthropic Claude is used zero-shot by a legal-tech company to extract parties, dates, and obligations from contracts - the structured system prompt alone provides sufficient guidance for >90% field accuracy.
Created for this library
A retail e-commerce team uses zero-shot learning to classify new product categories before any labeled examples exist.
A research team uses zero-shot learning to evaluate a model's generalization to tasks it was not explicitly trained on.
A SaaS team uses zero-shot learning to launch a new ticket category in a day without any labeled examples for that class.
Definition source: Google for Developers Machine Learning Glossary | Creative Commons Attribution 4.0 License