Skip to slide
Chapter 1 · Chain of Thought, Teaching Models to Think Out Loud
11 / 53

CHAPTER 01 · Chain of Thought, Teaching Models to Think Out Loud · 5 / 7

A reliability booster: self-consistency

A natural follow-up idea, worth knowing because it is widely used, is self-consistency. Instead of generating one chain of thought, you generate several different ones (the model can reason its way to an answer by more than one path), and then take the answer that comes up most often, a majority vote.

flowchart TD
    Q[Question] --> C1[Reasoning path 1 → 9]
    Q --> C2[Reasoning path 2 → 9]
    Q --> C3[Reasoning path 3 → 8]
    C1 --> V[Majority vote]
    C2 --> V
    C3 --> V
    V --> A[Answer: 9]

The intuition: a correct answer can be reached by many different valid lines of reasoning, while mistakes tend to be scattered and inconsistent. So the most common answer across several attempts is usually the right one. This trades extra computation for extra reliability, which is a theme that returns again and again in this folder.

← → arrow keys work too