Prompting and Prompt Engineering

Asking Models to Reason

---

Reasoning required

Some tasks require a bit of reasoning from the model. For example, we might want to really bake the chocolate cookies that the model suggested in the chapter focused on Clear Instructions.

Let’s assume that the recipe requires 200g of chocolate, 150g of butter, 2 dl of sugar, 1 egg, 3 dl of wheat flour, and a pinch of salt to produce 50 cookies, but we are short on some of the ingredients.

We can ask the model to calculate the amount of cookies that we can make based on the amount of ingredients that we have. To receive only a numeric response, we use few-shot prompting.

The above is not quite right, as the model fails to reason about the amount of butter.

Chain-of-thought prompting

Chain-of-thought prompting is a technique where large language models are provided some information of how to determine the answer. This can lead to improved reasoning about the inputs.

For additional details, see Chain-of-Thought Prompting Elicits Reasoning in Large Language Models.

Let’s adjust the prompt so that in each answer, we also provide reasoning that leads to the output.

That’s more like it! The model is now able to reason about the amount of butter.

In the above examples, we provided the reasoning for the model to determine the answer. Chain-of-thought prompting can also be used without explicit examples. In such a case, the prompt is added an additional instruction to think through the solution step by step.

This is highlighted in the following example.

As we’ve also observed before, the prompt can influence the output considerably. Specific prompts or parts of prompts can yield more meaningful results than others. As an example, researchers have found that simply adding “Let’s think step by step” before each answer can lead to reasoning in zero-shot prompting as well.

For additional details, see Large Language Models are Zero-Shot Reasoners.