Unexpected prompts can lead to unexpected behavior

You might remember that an LLM is just statistically reciting tokens it has seen before. If your prompt is doing something that the LLM deems unexpected, it might cause it either:

  1. blindly follow your instructions, or

  2. exert behavior that looks like it assumes you made a mistake and just selectively augments part of your prompt.

To illustrate this, in one of our Examples, there used to be an equation for calculating total shipping costs on an E-commerce platform:

- Total Shipping costs: SC = Base Shipping Cost+(2×Size (kg))+(1.5×Weight (kg))

It states the total shipping costs depends on a base cost and weight of the item category. We actually forgot to mention that the number of shipped products should also be included. We found that the LLM ended up being confused by this and exerting non-deterministic behavior, sometimes adding the number of products, sometimes leaving it as is, and sometimes coming up with something else entirely. From its behavior, it seemed that it was somehow expecting the number of items shipped to be part of the equation.

This is where your knowledge of the problem comes in. If the shipping cost indeed should depend on the number of items shipped, it should be reflected in the equation:

- Shipping costs: SC = Base Shipping Cost + ((2×Size (kg))+(1.5×Weight (kg)) * units sold)

If it is somehow not dependent on the number of items shipped, we can improve the prompt a little bit by stating this fact explicitly:

- Shipping costs: SC = Base Shipping Cost+(2×Size (kg))+(1.5×Weight (kg)). The total shipping cost is independent on the number of items shipped.