1. Chain-of-thought and self-consistency prompting
In this video, we will discuss prompting techniques to understand the output a model returns.
2. Chain-of-thought prompting
Chain-of-thought prompting requires language models to present reasoning steps, or thoughts, before giving a final answer. This technique is valuable for complex reasoning tasks and helps reduce errors by processing the reasoning step by step.
3. Chain-of-thought prompting
To understand the power of chain-of-thought prompts, let's compare them with standard prompts using a math problem. Suppose we want to determine how many books a person has, given their existing count and their lending and purchasing decisions. A standard prompt gives us a number, but doesn't explain the reasoning. We can't verify its correctness without seeing the steps taken.
4. Chain-of-thought prompting
To address this, use a chain-of-thought prompt, asking the model for a step-by-step explanation. The model solves the problem as a series of five steps, providing the correct answer.
5. Chain-of-thought prompting with few-shots
We can use few-shot prompts to obtain chain-of-thought reasoning. Instead of instructing the model to generate reasoning steps, we provide examples of what the answers should include. For instance, to determine if a group of odd numbers adds up to an even number, we provide an example question and answer. This exemplifies the steps: finding the odd numbers first, then summing them to verify the statement. We then provide a new question, with an "A:" for the model to answer. We combine the example and the question to obtain the final prompt. As a result, the model follows a similar logic in its response.
6. Chain-of-thought versus multi-step prompting
Let's study the difference between multi-step and chain-of-thought prompts.
With multi-step prompts, the various steps of the task are directly incorporated into the prompt itself, guiding the LLM's behavior.
7. Chain-of-thought versus multi-step prompting
Chain-of-thought prompts take a different approach by instructing the model to generate intermediate steps or thoughts in its output as it solves the problem. This helps gain insight into the model's decision-making.
8. Chain-of-thought limitation
One limitation of chain-of-thought prompting is that one thought with flawed reasoning will lead to an unsuccessful outcome. This is where self-consistency prompts come in.
9. Self-consistency prompting
Self-consistency prompting is a technique that generates multiple chain-of-thought responses by prompting the model several times. The final output is determined by a majority vote, selecting the most common response as the result.
10. Self-consistency prompting
To implement self-consistency prompting, we can define multiple prompts, or a prompt where the model imagines several independent answers. Here we ask for several independent experts to solve a mathematical problem determining the number of cars in a parking lot, with the final answer obtained by majority vote. To obtain the final prompt, we combine this instruction with the mathematical problem to solve.
11. Self-consistency prompt
In the output, the model gives the response from each expert and aggregates the results to provide a final answer. Since two of the three experts obtained the number 12, the final answer is 12.
12. Let's practice!
Time to practice!