跳到主要内容Skip to main content

Self-Consistency(自一致性)Self-Consistency

Self-Consistency

  • 智能体Agents
  • 第三站 · 用深Station 3 · Going Deeper
  • 引用核验 2026-11-19refs checked 2026-11-19
解决什么问题What problem it solves

让模型对同一道题采样多条推理路径,再按多数答案投票,用冗余换准确率。Sample several reasoning paths for the same question and take a majority vote, trading extra compute for accuracy.

这是什么What it is

先让模型用思维链生成多条不同的推理路径,而不是只取一条贪心解码结果。

再对最终答案做多数投票;路径不同但答案一致,说明这条答案更稳。

它不训练新模型,也不改权重,只在推理阶段多做几次采样。

Generate multiple chain-of-thought paths instead of one greedy decode.

Take a majority vote over final answers; agreement across different paths is a stronger signal.

No new model and no weight change — only extra sampling at inference.

为什么需要知道Why it matters

当单次思维链偶尔算错时,自一致性能把「偶然对一次」变成「稳定对多数」。它是测试时计算里最便宜的一种。When a single chain of thought occasionally slips, self-consistency turns “right once by luck” into “right in most samples”. It is the cheapest form of test-time compute.

出处Sources
  • Self-Consistency 论文(arXiv)Self-Consistency Improves Chain of Thought Reasoning (arXiv 2203.11171) ↗