Why does telling a model "Don't use the word 'delve'" often fail? What is a more effective way to rewrite a prompt to avoid specific behaviors?

Question

Accepted Answer

LLMs are "probabilistic next-token predictors." Telling them not to do something often increases the "weight" of those tokens in their attention. It is better to provide a Positive Alternative (e.g., instead of "Don't be wordy," use "Give a one-sentence summary").

Why does telling a model "Don't use the word 'delve'" often fail? What is a more effective way to rewrite a prompt to avoid specific behaviors?

Practice Your Response

Similar Questions in Generative AI & LLMs