Why models “hallucinate” less when you give examples

“Hallucination” is the polite word for the moment a model confidently talks nonsense. A fresh paper digs into what practitioners already felt in their gut: give the model a couple of examples, and the made-up answers drop sharply.

And the explanation here isn't mysterious at all. But it does change how you should write prompts.

The short version

Look. Without examples, the model has to guess two things at once: what you want and how it should look. Two guesses. And each example you give collapses one of them into a fact.

• It narrows the format. The model copies the shape of your example instead of inventing one.
• It anchors the facts. Concrete instances pull the answer toward your domain and away from generic filler.
• It signals the bar. A careful example tells the model that “roughly right” isn't good enough.

Examples don't make the model smarter. They make the target smaller — and a smaller target is harder to miss.

What to do with it

And you don't need a research budget for this. The next time an agent hands you a vague or invented answer, don't argue with it. Just show it one good example of what you wanted. More often than not, the second try lands. It's the cheapest way to boost accuracy you've got.