Bài kiểm tra có liên kết hướng dẫn · Trung bình Cấp độ
Chủ nghĩa không xác định trong bài kiểm tra đầu ra LLM
Hiểu lý do tại sao các lệnh gọi LLM lặp lại lại khác nhau và biện pháp kiểm soát nào cải thiện khả năng tái tạo mà không đảm bảo kết quả đầu ra giống hệt nhau trên các hệ thống.
Câu hỏi 1 của 8
Why might an LLM return different text for the same prompt on two runs?
Xem trước mọi câu hỏi8 câu hỏi
Câu trả lời và giải thích xuất hiện khi bạn làm bài kiểm tra.
- Why might an LLM return different text for the same prompt on two runs?
- A prompt can only ever have one possible continuation
- Sampling, backend changes, numerical execution, or different tool results can affect generation
- Temperature zero guarantees identical output everywhere
- A model never uses probabilities
- What does a seed parameter generally provide in a supported API?
- A guarantee of identical output across providers
- A guarantee that an answer is correct
- A way to pin tool results automatically
- Best-effort repeatability under matching request and backend conditions
- What can a system fingerprint help a developer detect?
- A change in backend configuration that may affect reproducibility
- Whether the generated answer is true
- Whether users like the answer
- The prompt’s character count
- Why might temperature zero still fail to guarantee identical hosted outputs?
- Temperature zero disables the model
- The API ignores every prompt
- Infrastructure and numerical execution can vary despite deterministic sampling settings
- Tokens are not probabilistic
- What should an experiment log to help reproduce an output?
- Only the response text
- Model identifier, prompt, sampling settings, seed, fingerprint, and tool state
- Only the user’s account name
- Chỉ có điểm cuối cùng
- When is exact-string matching most appropriate in an evaluation?
- When the task explicitly requires exact formatting or a fixed output
- For every open-ended response
- When model versions change often
- When no reference output exists
- How can teams evaluate variation when exact matching is too brittle?
- Ignore all failures
- Compare only response length
- Assume every paraphrase is wrong
- Use structured fields, semantic criteria, or task outcomes
- Does recording a seed make a system correct or secure?
- Yes, it guarantees factual answers
- Yes, it encrypts prompts
- No; it only helps control generation variation in supported conditions
- No, because seeds have no use