Self Consistency

---
title: Self-Consistency Prompting
category: frameworks
tags: [reasoning, voting, multi-path, consistency, reliability]
difficulty: advanced
models: [claude, gpt-4, gemini, llama, mistral]
---

# Self-Consistency Prompting

Self-Consistency prompting generates multiple independent reasoning paths for the
same problem and selects the most consistent answer through majority voting,
significantly improving reliability on math, logic, and factual questions.

## When to Use

- Mathematical and arithmetic word problems where a single chain may err
- Logic puzzles and constraint satisfaction with multiple valid approaches
- Factual questions where hallucination risk is high
- High-stakes decisions where confidence in the answer matters
- Any reasoning task where you need a reliability estimate

**When NOT to use:**
- Creative or open-ended tasks with no single correct answer
- Simple factual retrieval that doesn't require multi-step reasoning
- Latency-sensitive applications — multiple paths multiply inference time

## The Technique

Standard chain-of-thought prompting produces a single reasoning path, which
may contain errors that go undetected. Self-Consistency addresses this by
sampling multiple independent solutions and aggregating the results. If four
out of five paths reach the same answer, confidence is high. If answers are
split, the problem may be ambiguous or harder than expected.

The key insight is that correct reasoning tends to converge on the same answer
even when the intermediate steps differ, while errors tend to scatter across
different wrong answers. This makes majority voting a powerful error filter.

## Template