Zero-shot

AI concepts
About 1 min read

The ability of a model to immediately perform a new task using only pre-trained knowledge and natural language instructions (prompts), without separate training data or examples (few-shot).

Also known as
Zero-shot LearningZero-shot

Detailed explanation

Zero-shot is a technique where an AI model identifies the user's intent to generate outcomes without referencing specific examples for a given task. This is based on the relationships between concepts and logical reasoning capabilities accumulated by Large Language Models (LLMs) through training on massive data. A prime example is when a user asks to "translate this sentence into English" without providing translation examples. Since no separate examples (shots) are included, it consumes fewer tokens and yields faster response times, but performance may be limited in specialized fields outside the model's pre-trained scope or tasks requiring complex output formats.

Why it matters in tool selection

Zero-shot performance represents the baseline strength of the AI model. Selecting a model with excellent zero-shot performance reduces the hassle of constructing examples, minimizes prompt length to save API costs (tokens), and shortens latency in real-time services.

What to check

  • Does it accurately follow the desired output format (JSON, Markdown, etc.) based only on instructions?
  • Does the model sufficiently understand the common sense or basic concepts of the domain?
  • Is a Chain-of-Thought technique needed instead of Zero-shot when complex reasoning is required?

Example

Input: "Analyze the sentiment of the following sentence: 'The weather is so nice today, it's perfect for a walk.'" Output: "Positive" (Note: No example sentences for the 'Positive' outcome were provided)

Confusing terms

Few-shot

A method that includes about 2 to 5 input-output examples within the prompt to assist in performing the task.

Fine-tuning

The process of optimizing performance by updating the model's own weights with a specific dataset, rather than modifying the prompt.

Related terms

Few-shot LearningPrompt Engineering