Zero-Shot Learning

In one line

Ask the model to do a task with no examples — just a description.

What it actually means

Zero-shot means the model has never seen labeled examples for this specific task at training time (and you’re not giving it any at inference time). For LLMs, this looks like: “Classify this email as urgent or not urgent: …” — no example pairs, just the instruction. For CLIP and related models, zero-shot image classification works by encoding candidate labels as text prompts (“a photo of a cat”, “a photo of a dog”) and picking the closest to the image embedding. Quality depends on how well the task description matches something the model saw during pretraining.

Why it matters

Zero-shot is the cheapest, fastest way to try an idea. No dataset, no training, no fine-tuning — just a prompt. The baseline for any new task should be “what does zero-shot already get me?” before you invest in labeling or fine-tuning. For many tasks modern LLMs are already good enough zero-shot that you ship without any further training.

Example

Prompt: "Is the following review positive or negative? Review: The battery dies in two hours. Answer:"
Output: "negative"

You’ll hear it when

• Prototyping a classification task with an LLM.
• Evaluating CLIP or open-vocabulary detectors.
• Comparing zero-shot, few-shot, and fine-tuned baselines.
• Justifying a quick proof-of-concept before a labeling project.