An inference and alignment technique where a model generates multiple (N) candidate responses to a single prompt, and a separate Reward Model or verifier scores them to select the single best output.
Having the AI brainstorm multiple different answers to the same question, and then using a scoring system to pick the single best, safest, or most accurate one.
Best-of-N (BoN) decoding is a strategy used to improve the quality, safety, and reasoning accuracy of LLM outputs without further training the base model. During inference, the model samples N distinct responses (often using a higher temperature to encourage diversity). A separate, trained Reward Model (or a rule-based verifier) then evaluates each candidate and assigns a score. The response with the highest score is returned to the user. While computationally more expensive at inference time, BoN is highly effective at reducing hallucinations and enforcing safety constraints, and it is a foundational component of advanced reasoning models.
For enterprises, BoN is a powerful, low-risk way to improve AI reliability. Instead of deploying a new, untested model, organizations can apply BoN to their existing models to dramatically reduce the rate of harmful or nonsensical outputs. In security, BoN is often used as a defense mechanism: if the Reward Model is specifically trained to penalize jailbreaks or toxic content, BoN actively filters out adversarial successes before they reach the user.
A user asks a medical AI a complex diagnostic question. The system generates 5 different potential answers. A specialized medical verifier model scores each answer based on factual accuracy and safety. The system discards 3 answers that contain minor hallucinations and presents the highest-scoring, most clinically sound answer to the doctor.