Reasoning / Research term
Self-consistency
A test-time method that samples several reasoning paths for one problem and chooses the final answer that appears most often.
Self-consistency samples multiple chains of reasoning and marginalizes over their final answers, usually with a majority vote. It was introduced as an alternative to taking one greedy chain-of-thought path. The method can improve performance when different valid paths converge on the same answer and mistakes do not all share one cause. Agreement is evidence about the samples, not proof of correctness: a model can repeat the same misconception across every path.
Builder example
The method fits tasks with a canonical answer that can be normalized before voting, such as some math and multiple-choice problems. Gains vary by model, prompt, task, sample count, and decoding settings, while inference cost grows with the number of samples. Compare it with a direct answer and an external verifier on representative cases before paying for it in production.
The model estimates a usage-based bill three different ways and two answers agree.
Use agreement as a confidence signal, then verify the arithmetic or show the customer the check.
Common confusion: Majority voting fails when the model has a systematic bias. If all five samples share the same flawed assumption, they converge on the same wrong answer with high confidence.