引導式測驗 · 中等 等級
反向 Prompt Engineering 測驗
從範例中推斷可觀察到的提示要求,然後測試新的提示,而無需聲稱恢復唯一的原始提示。
問題 1 的 8
How does reverse prompt engineering work in this guide?
預覽每個問題8 問題
當您參加測驗時,就會出現答案和解釋。
- How does reverse prompt engineering work in this guide?
- Reading the model’s hidden system prompt directly
- Recovering every training document
- Using an example output to infer instructions for a new, similar output
- Changing model weights without examples
- Can an output alone reveal one unique prompt that produced it?
- Usually no; many prompts and settings can create similar outputs
- Yes, every output has exactly one prompt
- Yes, if the output has headings
- No, because outputs contain no information
- What does output-to-prompt research support?
- An exact prompt is always recoverable
- Prompt inversion reveals model weights
- A sample proves the output is factual
- A system may infer a semantically related prompt, not necessarily the original wording
- How should a user begin analyzing a sample output?
- Guess the model’s private instructions
- List observable features such as structure, tone, audience, and length
- Copy every sentence verbatim
- Assume its factual claims are correct
- Why test a reconstructed prompt on multiple unseen examples?
- To prove it matches the hidden prompt exactly
- To remove all model variation
- To see whether its instructions generalize beyond the original sample
- To avoid defining a rubric
- How should factual claims in the target output be handled?
- Assume style matching proves accuracy
- Verify them separately from style or format similarity
- Remove all citations
- Use them as the prompt’s only constraint
- Which use best fits this method?
- Describing reusable structure and tone from a sample
- Claiming access to someone else’s private prompt
- Recovering training data
- Bypassing a model’s safety rules
- What should happen if the prompt is moved to a different model?
- Assume it will behave identically
- Remove the test cases
- Keep only the original output
- Re-test and adjust because behavior may differ