Glossary
Zero-Shot Learning
Asking a model to do something it was never explicitly trained on, with no examples, just a description of the task. Large language models turned this from a research niche into the default way software uses AI.
General definition
Zero-Shot Learning describes a model handling a task or class without any labelled examples of it. The term comes from computer vision, where a classifier trained on some animals could be asked to recognise a new one from a textual description of its attributes. With large language models it became mainstream: because the model has read an enormous amount of text about almost every task, an instruction such as “Classify this support ticket as billing, technical or account” is usually enough to get a sensible answer with no training data at all.
It sits on a spectrum of how much task-specific signal you give the model:
- Zero-shot: a description of the task only; fastest to set up, no data needed, quality depends on how well the model already knows the task
- Few-shot (in-context learning): the same description plus a handful of worked examples in the prompt; often a large accuracy gain for the cost of a longer prompt
- Fine-tuning: updating the model’s weights on hundreds or thousands of examples; best for narrow, high-volume tasks, fixed formats and style, at the cost of data collection and a training pipeline
LLMs are strong zero-shot learners because pre-training on web-scale text and instruction tuning teach them the shape of thousands of tasks, so a new one is usually a recombination of things they have seen. Prompt patterns that raise zero-shot accuracy include stating the role and the goal explicitly, defining the output format (a fixed label set, JSON), listing the decision criteria, and asking the model to reason step by step before answering. A well-written system prompt is, in effect, a zero-shot task specification.
Zero-shot fails when the task depends on private knowledge the model never saw (your product names, your policies), when labels are subtle or domain-specific (medical coding, legal categories), when the output must match an exact house format, or when the model fills gaps with plausible hallucinations. The usual escalation path is zero-shot, then few-shot, then retrieval to supply the missing knowledge, and fine-tuning only when volume and stability justify it.
In the Ethora ecosystem
Most agents built on the Ethora AI SDK start zero-shot. An agent is a model plus a system prompt, so a support bot, a triage assistant or a moderation reviewer is specified in plain language and works from the first message without any training data. Where the task needs knowledge the model lacks, the knowledge base supplies it through retrieval rather than fine-tuning: the agent stays zero-shot on the task and grounded on your documents.
Because each agent chooses its own model, teams can test the same prompt against a cloud model and a self-hosted one through an OpenAI-compatible endpoint (the self-hosted LLM path) and see which handles the task zero-shot well enough; smaller local models often need a few examples in the prompt where a large cloud model does not. The AI auto-moderation in the Trust and Safety system is a zero-shot classifier in practice: a policy written in words, applied to every message.
