Zero-Shot Learning
Zero-shot learning is a model performing a task it was never explicitly trained on, from a description of the task alone, without labelled examples — asking a language model to classify support tickets into categories it has never seen, or CLIP to recognise a document type from its name. Few-shot learning supplies a handful of examples in the prompt. Large pre-trained models made zero-shot practical, and it decides whether a fintech needs a labelled dataset or a well-written sentence.
- Also known as
- Zero-shot, Few-shot learning, In-context learning, Zero-shot classification
- Used at
- classification, extraction, routing and moderation tasks where labelled data is scarce or the categories change often; rapid prototyping; the first version of nearly every language-model feature before fine-tuning is considered
- Binding
- no rule; but a zero-shot classifier making decisions about customers is still a model under model-risk and the EU AI Act, and its accuracy has to be measured like any other
- Typical mechanics
- pre-training on vast, varied data gives the model general capabilities; a prompt describes the task; the model applies what it knows. Few-shot adds two to ten examples to the prompt; fine-tuning trains the weights on hundreds to thousands; the choice is a trade of data and cost against accuracy and consistency
- Typical accuracy
- zero-shot with a frontier model often matches or beats a supervised model trained on a few thousand examples for text tasks; few-shot adds several points; fine-tuning wins where the task is narrow, the volume is high and consistency matters
- Parties
- the labs whose pre-training makes it possible, the engineers who prompt, and the data teams who decide when labelling is worth it
How zero-shot learning works in practice
A fintech launches in a new market and needs to route support tickets into 14 categories. The supervised route: label 3,000 tickets, train a classifier, deploy — three weeks. The zero-shot route: write the 14 categories with one-line descriptions into a prompt, ask the model to pick one per ticket, deploy — an afternoon. Measured on 500 hand-labelled tickets, the zero-shot classifier scores 84% accuracy; adding three examples per category (few-shot) brings it to 89%; the supervised model, three weeks later, scores 91% and costs a hundredth as much per ticket to run. The zero-shot version shipped on day one and the supervised one replaced it when the volume justified the cost.
The mechanism is that a large pre-trained model has seen so much text that "classify into these categories" is a task it already knows how to do; the prompt merely names the categories. The limits are consistency (the model may drift between runs and phrasings), cost (a frontier model per ticket is dear at scale), and the tasks pre-training did not cover (a company's private codes, a rare document layout), where examples or fine-tuning are needed. The decision framework is simple: start zero-shot, measure, add examples where accuracy is short, fine-tune or train a small model where the volume makes the frontier model's cost the problem.
For AI M&A, zero-shot capability is what commoditised the labelled-data businesses of the 2010s; for fintech, it is why a feature that once needed a data-science project now needs an evaluation set and a prompt.
Example
A lender's fraud team wants to flag loan applications whose stated purpose is inconsistent with the applicant's profile — a category no dataset labels. A zero-shot prompt describing the inconsistency and asking for a yes/no with a reason, run on 20,000 past applications, flags 3.1%; analysts review a sample of 400 flagged and 400 unflagged: precision 61%, recall 78%. Adding eight examples of true and false flags raises precision to 74%. The feature goes live as a review-queue signal, not a decision; after six months and 2,600 analyst verdicts, a small fine-tuned model trained on those verdicts reaches 83% precision at a tenth of the cost and takes over. The zero-shot version had been the labelling programme.
Common mistakes
- Deploying a zero-shot classifier without an accuracy measurement on the company's own data. - Staying zero-shot at a volume where a fine-tuned or small model would be far cheaper. - Fine-tuning before checking whether a better prompt and a few examples suffice. - Treating zero-shot output as deterministic; phrasing and sampling change it. - Using zero-shot for decisions about customers without the governance any model needs.