Best-of-N Sampling

An inference-time search strategy that independently generates N candidate samples, scores all candidates with a reward model or other scoring function, and selects the highest-scoring result.