Self-Consistency: picking the most reliable answer

Self-Consistency

A prompting format that generates several answers to the same question and then selects the most consistent or most frequent one. This technique improves the reliability and accuracy of a model's answers, especially on tasks that require reasoning or calculation.

AdvantagesDrawbacks
Improves answer accuracyRequires more compute
Reduces the impact of random errorsCan take more time
Especially effective on tasks with ambiguous solutionsCan complicate interpretation of the results when several solutions are possible

Example of a Self-Consistency prompt:

Solve the following problem 5 times, giving a full explanation each time:
If 8 workers can build a house in 10 days, how many days will it take 12 workers to build the same house?

Once solved, pick the most frequent answer and explain why it is probably the correct one.

Try it in GPTunneL