Rejection sampling is originally a statistical technique in which samples are drawn from an easy-to-sample distribution and those that fail a condition are discarded, yielding samples from a target distribution. In large language model work the term borrows this idea: the model generates many answers or solutions to a problem, and only those that pass a verifier or a reward model's criterion are kept.
In post-training, retraining on the filtered correct solutions as supervised fine-tuning data, known as rejection sampling fine-tuning, is a simple way to improve reasoning ability. Its limitations are that all information in failed attempts is thrown away and that unnecessary tokens inside correct solutions are reinforced just as strongly, which is why reinforcement learning methods such as GRPO, which learn from relative comparisons within a group, are used to complement it. Best-of-n sampling, which generates several candidates at inference time and picks the best, follows the same principle.