prompt attack

[zl-comment]

eg
"Determine if the given pair of sentences displays entailment, neutral, or contradiction. Respond with 'entailment', 'neutral', or 'contradiction'\nQuestion: {content}\nAnswer:"
The prompt words Answer and Question cannot be modified. If they are attacked, the accuracy will drop a lot. I did not see this in the prompt words in your conclusion.

[Immortalise]

Yes, in our setup, terms like ‘entailment’ and ‘Question’ are fixed and unmodifiable because we believe changing them is unnecessary and would cause a significant shift in meaning compared to the original clean prompts.

In this example, you can define your own unmodifiable words.