To illustrate the self-consistency method, consider a language model prompted with a Chain-of-Thought query to find the probability of exactly one head in three coin flips. The model generates three reasoning paths via sampling. Prediction 1 calculates a probability of ${}37.5\%$ using basic principles. Prediction 2 correctly applies the binomial probability formula, also arriving at ${}37.5\%$. Prediction 3 incorrectly concludes the probability is ${}50\%$. Although the reasoning paths for the first two predictions are different, they converge on the same final answer of ${}37.5\%$. By counting the frequencies of the final answers, ${}37.5\%$ has the highest count (two out of three) and is selected as the final output.