Zero-Shot

Definitions

1. Describing a setting where a model performs a task without being given task-specific examples in the prompt, relying instead on prior training and general instructions. Zero-shot behavior is useful because it shows how well a model generalizes without explicit demonstrations.

   In plain English: Doing a task without being shown examples first.

   Example: "The router started with a zero-shot prompt for intent classification before the team added examples for harder edge cases."

Related Terms

• In-Context Learning
• Chain of Density
• Chain of Verification
• Context Stuffing
• Few-Shot
• Goal Misgeneralization
• Out-of-Distribution
• Soft Prompt
• Step-by-Step
• Think Step by Step
• Unlocking