Self-consistency

Ask the same question five times, keep the answer that shows up most often, and accuracy goes up.

Part of the Getting good answers track on lAItest.

One model, one question, five separate attempts. Three say 47. Two say 52. The vote is the answer.

Why sampling more than once helps.

Different routes, same destination.

A model with sampling turned up does not produce the same chain of steps twice. Run the question several times and you get several routes. Wrong routes tend to go wrong in different directions and scatter; correct routes tend to arrive at the same value. So you count the final answers, keep the most common one, and throw the reasoning away.

Try it

Pull temperature to zero and every run is the same run. Self-consistency needs the spread — drag it up and watch the answers diverge. This step is an interactive widget; open the lesson to use it.

A common misconception

Commonly believed: Majority voting means the model checks its own work.

Actually: It never sees the other attempts. Each run is independent and blind. The vote happens outside the model, in your code, and it only works when there is a single comparable answer to tally. Free-form prose gives you nothing to count.

The cost is exactly the multiplier.

Five samples cost five times one sample. That is the entire trade: accuracy bought with money and latency, in a straight line, with no discount. It earns its keep on short high-value answers that can be compared, such as a number, a label or a decision, and rarely on a long document nobody can vote on.

In one sentence

Self-consistency is a vote held outside the model: sample the same question several times and keep the answer that keeps showing up.