Welcome to the next stage of our journey into Advanced Techniques in Prompt Engineering. We will focus on the Chain-of-Thought method. This approach not only enhances the model's output but also aids in attaining precise answers by guiding the model through a logical sequence of reasoning.
The Chain of Thought method is a compelling technique that encourages the model to "think aloud" as it reaches an answer. This technique is particularly useful when dealing with complex problems that can benefit from step-by-step reasoning. Unlike most humans who by adulthood know that the best way to think through a complex problem is to think step-by-step, LLMs don't have this instinctive understanding. Explicitly prompting the model to detail its thought process is proven to lead to more accurate and helpful responses.
Designing prompts that leverage the Chain of Thought method requires clarity in the query and precision in the expectations. The goal is to guide the model away from merely guessing or jumping to conclusions and towards methodically working through the problem. This strategy not only aids in achieving more reliable outputs but also enables users (and developers) to understand the rationale behind the model's answers.
One task that LLMs are notoriously bad at is mathematical computations. Let's see an example:
Markdown1__ASK__ 2 3What's 197*971
While after running this many times you might get the right answer, more often than not the LLM will make a mistake and respond with numbers like 191,087, 191,987, etc, none of which are correct.
Now let's see what happens when we change it to apply Chain-of-Thought. The most standard way to apply this method is to add a requirement to think step-by-step. Other acceptable options include
- begin by carefully analyzing the problem
- show your work before responding
- etc
Markdown1__ASK__ 2 3what's 197*971 4 5__REQUIREMENTS__ 6- Think step-by-step
Sample output:
Plain text1Step 1: Multiply 7 (units place of 197) by 971 2 7 * 971 = 6797 3Step 2: Multiply 9 (tens place of 197) by 971 and add a zero at the end 4 9 * 971 = 8739 => 87390 5Step 3: Multiply 1 (hundreds place of 197) by 971 and add two zeroes at the end 6 1 * 971 = 971 => 97100 7Step 4: Add all the results from steps 1, 2, and 3 8 6797 + 87390 + 97100 = 191287 9 10Therefore, 197 * 971 = 191,287
And that response is spot on.
When it comes to Chain-of-Thought, the more you can help guide the thought process of the LLM in the right direction, the more likely you are to get a high-quality and accurate response. For example, the way we can improve mathematical calculations is to hint that using relevant round numbers could help with computations, like this:
Markdown1__ASK__ 2 3what's 197*971 4 5__REQUIREMENTS__ 6- Think Step by Step 7- Identify more round numbers that can be combined to compute this.
Sample Output:
Plain text1To calculate 197*971, we can break it down into simpler calculations by identifying round numbers that can be combined. 2 3Let's break down 197 and 971 into multiples of 100 and the remaining numbers: 4 5197 = 100 + 97 6971 = 900 + 71 7 8Now, we can multiply these in parts: 9 10(100 + 97) * (900 + 71) 11= 100 * 900 + 100 * 71 + 97 * 900 + 97 * 71 12 13Calculating the individual products: 14 15100 * 900 = 90,000 16100 * 71 = 7,100 1797 * 900 = 87,300 1897 * 71 = 6,887 19 20Now we add these results up: 21 2290,000 + 7,100 + 87,300 + 6,887 = 191,287 23 24Therefore, 197*971 is equal to 191,287.
Having learned how to structure your prompts to efficiently utilize the Chain of Thought method, it's now your turn to experiment with this approach. Remember, the goal is to encourage the model to think through problems, step by step, and conclude with concise and accurate answers. Let's begin the practice!