Self-Consistency Prompting

Introduction

Self-consistency runs the same CoT prompt N times with high temperature, then takes the majority answer. It's a simple ensemble that often beats single-shot reasoning.

Beginner analogy: Ask 5 students the same hard question; pick the answer most agree on. Crowd-sources the right answer from one model.

Understanding the topic

Core concepts to understand:

• Run prompt N times (typically N = 5–10) with temperature ≥ 0.7.
• Aggregate by majority vote on the final answer.
• Works best when answers are discrete (numbers, classifications).
• Costs N× as much; only use for high-value queries.

Syntax reference

Visual workflow / architecture:

Same prompt × N

Run 1: answer = 42
Run 2: answer = 42
Run 3: answer = 7
Run 4: answer = 42
Run 5: answer = 42

Majority → 42

Real-world use

Used in math benchmarks, financial calculations, and any high-stakes single-number answer where the cost of being wrong >> the cost of running 5 prompts.

Best practices

• Pair with CoT (it's CoT × N).
• Use only on tasks with discrete answers.
• Cache the aggregated answer to avoid re-running.

Hands-on exercise

Interview preparation — practice these questions:

• Q1. What is self-consistency?
• Q2. Why use temperature > 0 here?
• Q3. When NOT to use self-consistency?
• Q4. How does it differ from beam search?