What is self-consistency in prompting?
Self-consistency samples several independent reasoning answers to the same question at a non-zero temperature and picks the most common final answer. It improves accuracy on reasoning tasks at the price of multiplying token cost and latency.
Explain it simply
If five friends solve the same puzzle separately and four get 42, you trust 42 more than any single friend's answer. Self-consistency does that with one model.
Run the same chain-of-thought prompt several times with temperature above zero so the reasoning paths differ. Extract the final answer from each, then take a majority vote. Wrong paths tend to disagree with one another, while correct paths tend to land on the same answer.
It works for questions with a short, checkable final answer, such as maths or multiple choice. It is awkward for long free-text outputs, where there is nothing simple to vote on. The cost is direct: five samples cost about five times as much and may be run in parallel to protect latency. Use it only where accuracy justifies that.
Example
Question: 17 x 24 = ?
Sample 1 -> 408
Sample 2 -> 398
Sample 3 -> 408
Sample 4 -> 408
Sample 5 -> 418
Majority answer: 408What to say in the interview
"Self-consistency samples multiple reasoning paths and votes on the final answer. It boosts accuracy on reasoning tasks but multiplies cost, so I reserve it for high-value queries with short verifiable answers."