#Prompt wording matters less with larger models

Stop overthinking how you word your ChatGPT prompts for general questions. With the latest models, it doesn’t matter anymore.

A new study shows that the bigger LLMs get, the less they care about purely cosmetic variations in the prompt.

In plain terms: polishing the phrasing of your question won’t do much for the quality of the answer. Same for instructions like “Act as a math teacher”.

One trick still works, though: giving a few examples. In prompt engineering, this comes in three flavors:

• Zero-shot: no example at all
• One-shot: one example of an input and the expected answer
• Few-shot: a handful of examples, usually 2 to 5

The gains are biggest on complex questions, creative tasks and coding help.

Smaller models (under 20B parameters), the ones built to run locally, are a different story: they still react to the exact wording of the prompt.

Last year, “prompt engineer” looked like the job of the future. A year from now, who knows whether the specialty will still exist.

Link to the study: arXiv 2410.12405