B4 · Sampling and voting

Principle

The same question is put to the model N times. A vote (majority, similarity or judge) picks among the answers. The instances do not talk to each other: each answers alone, and the aggregation brings the reliability.

The problem it solves

A single draw can go wrong in isolation, and on some decisions a single error is costly.

When to use it

• Answers are short and comparable: a number, a class, a verdict.
• A single error is costly: moderation, vulnerability detection. It is also the baseline to beat before building a team of agents.

When to avoid it

• The output is open-ended text that cannot be compared by vote: prefer a generator-evaluator (B5).
• The budget is tight: cost grows with N, and gains fade beyond 5 to 10 samples.

Synonyms by vendor

• Parallelization (voting) (Anthropic)
• Self-consistency (Wang et al. (ICLR 2023))
• Agent Forest (« More agents is all you need ») (Li et al. (TMLR 2024))
• Voting-based cooperation (CSIRO Data61)
• Best-of-N (usage courant)

Neighbouring patterns

• B3 · Parallelization
• B5 · Generator-evaluator
• C3 · Reflection
• E5 · Group chat and debate

Sources

• Self-Consistency Improves Chain of Thought Reasoning in Language Models — X. Wang et al. (ICLR 2023), 2023
• More Agents Is All You Need — J. Li et al. (TMLR 2024), 2024
• Debate or Vote: Which Yields Better Decisions in Multi-Agent Large Language Models? — H. Choi, X. Zhu, S. Li (NeurIPS 2025), 2025