Consider a multi-step reasoning problem like the following:
Question: If a store has 10 apples and 8 oranges, and it sells 6 apples and 4 oranges, how many fruits are left in the store?
Instead of directly answering the question, the chain-of-thought prompting would require the language model to produce a series of short sentences that mimic a human's reasoning process:
The store has 10 apples.
The store has 8 oranges.
6 apples are sold.
4 oranges are sold.
There are 10 - 6 = 4 apples left.
There are 8 - 4 = 4 oranges left.
The store now has 4 apples + 4 oranges = 8 fruits.
Using self-consistency, the language model generates multiple reasoning paths:
Reasoning Path A:
The store has 10 apples.
The store has 8 oranges.
6 apples are sold.
4 oranges are sold.
There are 10 - 6 = 4 apples left.
There are 8 - 4 = 4 oranges left.
The store now has 4 apples + 4 oranges = 8 fruits.
Reasoning Path B:
The store initially has 10 apples and 8 oranges.
It sells 6 apples, so there are 10 - 6 = 4 apples left.
It sells 4 oranges, so there are 8 - 4 = 4 oranges left.
The store now has 4 apples + 4 oranges = 8 fruits.
