Decoder. plain-English AI glossary

Rejection sampling

▲ Rising

Also called Rejection Sampling

Generate several candidate answers, keep only the good ones, and train on those.

Think of it like

Baking a dozen cookies, keeping the best few, and photographing only those for the recipe card.

Example

A model produces eight solutions per math problem; a checker keeps the correct ones, and those become new training data.

How it actually works

Rejection sampling filters a model's own generations by some quality bar — a verifier, a reward model, or an exact-answer check — and fine-tunes on the survivors. It is a simple, stable way to bootstrap a model on its better outputs, closely related to best-of-n but used to build a training set rather than to pick a single response at inference time.

For product teams

A cheap way to squeeze more quality out of a model using its own best work.

For engineers

Sample n candidates, keep those passing a quality/verifier filter, and fine-tune on the accepted set.

Best-of-n — The inference-time cousin that keeps just one.
Sampling — The generation step it filters.

Read anything AI without the jargon

Look up any term in plain English, or save terms as you read with the free Chrome extension.

Open Decoder Add to Chrome