What is it?
Generate several answers, grade each one, keep the top-scoring answer. The grader might be unit tests (for code), a calculator, or a trained reward model.
It only works as well as the grader. A good checker makes it very effective; a poor one rewards answers that merely look good.
Also known as: best of n, best-of-n, BoN, rejection sampling, generate and rerank, sample and rank, N-best reranking, verifier reranking.
Technical definition
Best-of-N (rejection sampling at inference) selects argmax of a scoring function over N samples. Scorers include executable checks, outcome reward models (Outcome Reward Model (ORM)), process reward models (Process Reward Model (PRM)) and LLM judges (LLM-as-a-Judge). Over-optimising against an imperfect learned scorer can degrade true quality as N grows (reward-model over-optimisation, see Reward Hacking and Over-Optimization). It is also used offline to build training data (rejection-sampling fine-tuning).
Entity type: technique · Freshness class: evolving.
When should I use it?
Use when an automatic checker exists or can be trained, and the task is worth N times the inference cost.
Implementation patterns
- Code: generate N programs, run tests, return the first or best passing candidate.
- Math: score with a verifier or a calculator-backed checker.
- Writing: rank with a rubric-based judge, with a human spot-check.
- Training data: keep only the best samples for supervised fine-tuning.
Technologies and frameworks
Commonly associated: unit tests / executors, reward models, LLM-as-judge, vLLM n-sampling. Names are examples of what exists, not recommendations, and projects change quickly.
Practical guide
- Define a scorer you trust on a labelled set.
- Generate N diverse candidates.
- Select top-scoring; log scores.
- Plot accuracy vs N on held-out cases and stop where the curve flattens.
- Spot-check selected answers by hand (How to Evaluate AI Systems and Outputs).
Common failure modes
- Scorer bias: the scorer prefers length, confident tone or its own style.
- Over-optimisation: very large N against a learned scorer selects its blind spots.
- No diversity: low temperature yields near-duplicates.
- Cost grows linearly with N with no measured gain.
Troubleshooting
- Chosen answer looks good but is wrong — Use an objective check (tests, retrieval evidence) in place of or alongside the learned scorer.
- Quality plateaus or falls at large N — Reduce N; retrain or replace the scorer.
- Samples nearly identical — Increase temperature/top-p diversity.
Contrasted with
- Self-Consistency (Sampling and Voting) — Voting needs no scorer but needs comparable answers; best-of-N needs a scorer but works on any output the scorer can grade.
- Search over Reasoning (Tree of Thoughts, MCTS, Beam Search) — Best-of-N scores whole answers; tree search scores partial steps.
Questions this page answers
- What is best-of-N sampling?
- How do I choose the best LLM output automatically?
- Does best-of-N help with code generation?
- What is the risk of best-of-N with a reward model?
Typical goals: pick the best of several model answers; use tests or verifiers to choose outputs; avoid reward over-optimisation.
Sources and verification
Status: not independently verified. This entry was authored on 2026-10-06 from general technical knowledge, without live checking of the sources below. Treat every source as a lead to confirm (identifier, title, authorship, current version) before publication or citation.
Freshness class: evolving — Established techniques whose best practice shifts as research and tooling move. Re-review every 6 months.
Canonical sources (unverified leads):
- Cobbe et al., Training Verifiers to Solve Math Word Problems (2021) — arXiv:2110.14168
- Gao et al., Scaling Laws for Reward Model Overoptimization (2022) — arXiv:2210.10760