Chain-of-Thought (CoT) Prompting

Chain-of-Thought Prompting is a technique where models generate intermediate reasoning steps to solve complex tasks requiring multi-step reasoning. By explicitly breaking down the problem-solving process into sequential, logical steps, this method enhances model performance on tasks such as mathematical reasoning, commonsense inference, and contextual decision-making. It improves accuracy by guiding the model to focus on intermediate reasoning before arriving at the final answer, making the solution process more interpretable and aligned with the task’s requirements.

Here is an example of Chain-of-Thought (CoT) prompting applied to a math word problem:

Example

Prompt:

"John has one pizza, cut into eight equal slices. John eats three slices, and his friend eats two slices. How many slices are left? Let's think step by step."

Response (AI):
The pizza is divided into 8 slices.
John eats 3 slices, so $$8 - 3 = 5$$.
His friend eats 2 slices, so $$5 - 2 = 3$$.
Therefore, there are 3 slices left.

This step-by-step reasoning helps the AI break down the problem logically and arrive at the correct answer[1][2][3].

Advanced Techniques to Enhance COT

1. Few-Shot Chain-of-Thought Prompting

  • Description: Provide a few examples of reasoning steps in the prompt to guide the model in solving similar problems.

  • Example:

    Example 1:
Q: If a train travels 60 miles in 2 hours, what is its average speed?
A: Let's think step-by-step. First, we calculate the distance traveled (60 miles). Next, we divide by the time taken (2 hours). So, the average speed is 60 ÷ 2 = 30 mph.

Example 2:
Q: A car travels 120 miles in 3 hours. What is its average speed?
A: Let's think step-by-step. First, we calculate the distance traveled (120 miles). Next, we divide by the time taken (3 hours). So, the average speed is 120 ÷ 3 = 40 mph.

Task:
Q: A cyclist travels 90 miles in 3 hours. What is their average speed?

2. Self-Consistency Chain-of-Thought

  • Description: Generate multiple reasoning paths for the same problem and select the most consistent answer.
  • Example:
    Q: If a tree grows 5 feet per year and is currently 10 feet tall, how tall will it be in 4 years?

Reasoning Path 1:
- The tree grows 5 feet per year.
- In 4 years, it will grow an additional 5 × 4 = 20 feet.
- Adding this to its current height of 10 feet gives a total of 10 + 20 = 30 feet.

Reasoning Path 2:
- Current height is 10 feet.
- Growth over four years is calculated as (5 × 4 = 20).
- Total height = current height + growth = 10 + 20 = 30 feet.

Final Answer (most consistent): The tree will be 30 feet tall.

3. Analogical Chain-of-Thought Prompting

  • Description: Use analogies to help the model reason through problems by comparing them to familiar concepts.
  • Example:
    Q: How does a thermostat work?

A: Let's think step-by-step using an analogy. A thermostat is like a referee in a game. 
   - It monitors the temperature (like a referee monitors rules).
   - When the temperature goes above or below a set point, it adjusts heating or cooling (like a referee enforces rules when broken).
   - This ensures balance, just as a referee ensures fair play.

4. Contrastive Chain-of-Thought Prompting

  • Description: Provide both correct and incorrect reasoning examples to teach the model how not to reason.
  • Example:
    Correct Example:
Q: If there are 12 apples and you eat half of them, how many are left?
A: Let's think step-by-step. Half of 12 is calculated as 12 ÷ 2 = 6. So, there are 6 apples left.

Incorrect Example:
Q: If there are 12 apples and you eat half of them, how many are left?
A: Half of them means subtracting half from the total, so it's 12 − (12 ÷ 2) = −6. [Incorrect because subtraction was misapplied.]

Task:
Q: If there are 20 oranges and you eat half of them, how many are left?

5. Faithful Chain-of-Thought Prompting

  • Description: Combine natural language reasoning with symbolic computation (e.g., Python code) to ensure accuracy.
  • Example:
    Q: What is the sum of all even numbers between 1 and 10?

Natural Language Reasoning:
  Step-by-step:
  - Identify even numbers between 1 and 10: [2,4,6,8,10].
  - Sum them up.

Python Code for Verification:
    even_numbers = [2,4,6,8,10] result = sum(even_numbers) print(result)
    
Final Answer: The sum is `30`.

6. Tabular Chain-of-Thought Prompting

  • Description: Use structured formats like tables for reasoning steps to improve clarity.
  • Example:
    Q: A shop sells pens at $2 each and notebooks at $5 each. If you buy three pens and two notebooks, what is the total cost?

| Item      | Quantity | Price per Unit | Total Cost |
|-----------|----------|----------------|------------|
| Pen       |    3     |      $2        |    $6      |
| Notebook  |    2     |      $5        |    $10     |

Final Total Cost = $6 + $10 = $16

Answer: The total cost is $16.

Citations