Chain-of-thought 與 self-consistency 提示

使用 OpenAI API 的提示工程

Fouad Trad

Machine Learning Engineer

Chain-of-thought 提示

要求 LLM 先給出推理步驟（想法）再回答
用於複雜推理任務
有助降低模型錯誤

示意圖：chain-of-thought 提示要求模型逐步解題，輸出包含每一步的推理。

Chain-of-thought 提示

以標準提示解推理題

prompt = """Q: You start with 15 books in your collection. At the bookstore, you 
purchase 8 new books. Then, you lend 3 to your friend and 2 to your cousin. Later, 
you visit another bookstore and buy 5 more books. How many books do you have now? 
A: The answer is"""

print(get_response(prompt))

25 books

Chain-of-thought 提示

以 chain-of-thought 提示解推理題

prompt = """Q: You start with 15 books in your collection. At the bookstore, you 
purchase 8 new books. Then, you lend 3 to your friend and 2 to your cousin. Later, 
you visit another bookstore and buy 5 more books. How many books do you have now? 
A: Let's think step by step""" 
print(get_response(prompt))

Step 1: Start with the number of books in your collection: 15 books
Step 2: Purchase 8 new books at the bookstore: 15 + 8 = 23 books
Step 3: Lend 3 books to your friend: 23 - 3 = 20 books
Step 4: Lend 2 books to your cousin: 20 - 2 = 18 books
Step 5: Visit another bookstore and buy 5 more books: 18 + 5 = 23 books
Therefore, you have 23 books now.

少樣本的 chain-of-thought 提示

example = """
Q: The odd numbers in this group add up to an even number:  9, 10, 13, 4, 2.
A: Adding all the odd numbers (9, 13) gives 22. The answer is True.
"""

question = """
Q: The odd numbers in this group add up to an even number: 15, 13, 82, 7. 
A:
"""

prompt = example + question
print(get_response(prompt))

Adding all the odd numbers (15, 13, 7) gives 35. The answer is False.

Chain-of-thought 與多步驟提示之別

多步驟提示

在提示中內嵌步驟

示意圖：多步驟提示會給模型一連串需依序執行的步驟。

Chain-of-thought 與多步驟提示之別

多步驟提示

在提示中內嵌步驟

示意圖：多步驟提示會給模型一連串需依序執行的步驟。

Chain-of-thought 提示

要求模型生成中間步驟

示意圖：在 chain-of-thought 提示中，步驟會隨輸出一併由模型產生。

Chain-of-thought 的限制

任一步驟失誤 → 結果失敗
引入了 self-consistency 提示

示意圖：chain-of-thought 提示中，只要某一步推理有瑕疵，就可能導致失敗結果。

Self-consistency 提示

透過多次提示生成多條 chain-of-thought
以多數決取得最終輸出

示意圖：self-consistency 提示會產生多條 chain-of-thought，各自有輸出，最後以多數決決定結果。

Self-consistency 提示

可以用定義 多個提示，或讓一個 提示生成多個 回應來達成。

self_consistency_instruction = "Imagine three completely independent experts who 
reason differently are answering this question. The final answer is obtained by 
majority vote. The question is: "


problem_to_solve = "If there are 10 cars in the parking lot and 3 more cars arrive. 
Half the original number of cars leave. Then, half of the current number of cars 
arrive. How many cars are there in the parking?"


prompt = self_consistency_instruction + problem_to_solve

print(get_response(prompt))

Self-consistency 提示範例

Expert 1: Let's go step by step [...] Therefore, the total number of cars in the 
parking lot is 8 + 4 = 12.

Expert 2: First, let's calculate [...] Therefore, the total number of cars in the 
parking lot is now 5 + 2 = 7 cars.

Expert 3: Initially, there are 10 cars [...] Thus, the final answer is 8 + 4 = 12 
cars in the parking lot.

Based on the majority vote, the final answer is that there are 12 cars in the 
parking lot.

一起來練習吧！

使用 OpenAI API 的提示工程