Skip to content
Home: How language models work

03Book 06 · Chapter 03 of 08· 5 min · Self-consistency

Several tries

Self-consistency means: sample several full step-by-step replies, then keep the final answer that appears most often. Each reply is an ordinary run. Agreement is a count, not a proof.

Write the same question five times with on. You might get 323, 323, 326, 323, 319. Self-consistency means: keep the final answer that showed up most often — here, 323. You are not proving the math. You are counting landings.

That is how self-consistency is built into a chat: not by changing the trained , but by running the same question several times with on, collecting each final answer, and keeping the one that appeared most. The model is not “checking its work” in a new way. A program (or the product wrapping the model) is counting endings.

Why the count can help: a correct path often reuses common facts (17 × 20, then subtract 17), so several runs land on the same number. Wrong paths wander in different directions, so they split the vote. The majority is still just a majority. If most runs write a fluent wrong city, that city wins the count too. Self-consistency does not add a truth filter. It only prefers the ending that showed up most often.

Quiz

6 questions

Before moving forward, test your knowledge of what you’ve just learned!

  1. 1.Self-consistency:

  2. 2.This is Book 2 because:

  3. 3.Wrong hops often scatter because:

  4. 4.Five chains agreeing on a fluent wrong city:

  5. 5.Best of N with a scorer vs majority of answers:

  6. 6.Use majority on math more than on taste because:

0 of 6 answered

3 / 8