There are many confusing aspects of the experiment for me. #39
Description
Dear author,
For this paper, I have a general understanding of the method you proposed, but I have many unclear or confused points about the experimental details. I hope you can answer the:
- Through the data you provided, I can't combine different stages. For example, for defense, the prompts in the defense file are different from those of normal tasks. How should they be combined? Secondly, for different baselines, the experimental data is different? So is the comparison reasonable in this case?
- For the method, I don't quite understand what the "seed" in the method is? As you mentioned GPTFuzzer in the article, their "seed" is the jailbreak template, and the target is different harmful instructions. However, in your method and experiment, I can't clearly distinguish these points.
- Finally, do you have some examples that can reflect the working process or the whole life cycle of PromptFuzz? This will be of great help to readers' understanding.
The above are some of my questions. I would be very grateful if you can answer them.
Looking forward to your reply!