Self-Consistency

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.

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.

问题：如果一家商店有 10 个苹果和 8 个橙子，此时店里卖出了 6 个苹果和 4 个橙子，那么店里还剩下多少水果？

思维链提示（chain-of-thought prompting）不是直接回答问题，而是要求语言模型生成一系列模仿人类推理过程的短句：

商店有 10 个苹果。
店里有 8 个橙子。
卖了 6 个苹果。
卖了 4 个橙子。
还剩下 10 - 6 = 4 个苹果。
剩下 8 - 4 = 4 个橙子。
商店现在有 4 个苹果 + 4 个橙子 = 8 个水果。

推理路径A：

商店有 10 个苹果。
店里有 8 个橙子。
卖了 6 个苹果。
卖了 4 个橙子。
还剩下 10 - 6 = 4 个苹果。
剩下 8 - 4 = 4 个橙子。
商店现在有 4 个苹果 + 4 个橙子 = 8 个水果。

推理路径 B：

商店最初有 10 个苹果和 8 个橙子。
商店卖了 6 个苹果，所以还剩下 10 - 6 = 4 个苹果。
商店卖了 4 个橙子，所以还剩下 8 - 4 = 4 个橙子。
商店现在有 4 个苹果 + 4 个橙子 = 8 个水果。