Rejection sampling is generating many candidate outputs and keeping only those that pass a quality filter (reward model, verifier, or rule), used both for training data curation and inference-time improvement. This simple but effective technique leverages the observation that model quality varies across samples, and selecting the best yields significantly better results.