Moyan AI Training Institution LogoMoyan AI

Practice · Established · Advanced

Self-Consistency

Sampling several independent reasoning paths for the same question and taking the most common answer.

What Self-Consistency is

Because sampling is stochastic, a model can reach the right answer by different routes. Agreement across routes is a useful proxy for correctness.

How it works

Generate n completions at non-zero temperature, extract the final answers, and pick the majority. Cost scales linearly with n, so it is reserved for high-value questions.

Why it matters

It is one of the simplest reliability upgrades available for reasoning tasks and needs no model changes.

Common uses

  • Maths and logic evaluation
  • High-stakes extraction
  • Confidence estimation

Strengths

  • Meaningful accuracy improvement
  • Gives a rough confidence signal

Watch for

  • Multiplies cost and latency
  • Majority can be confidently wrong

Continue exploring

More in this collection

Browse all AI Concepts