引導式測驗 · 中等 等級

即時壓縮測驗

減少提示令牌,同時測量壓縮器開銷、端對端延遲、任務品質以及遺失關鍵上下文的風險。

相關引導路徑即時壓縮
問題 1 的 8

What does prompt compression try to achieve?

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。

  1. What does prompt compression try to achieve?
    • Guarantee that every model answer improves
    • Reduce input tokens while retaining information needed for the task
    • Replace evaluation with a shorter prompt
    • Remove all system instructions
  2. What did LLMLingua research investigate?
    • Guaranteed lossless compression for every task
    • A model that only answers in JSON
    • Removing the need for retrieved context
    • Prompt compression methods intended to reduce inference cost and latency while preserving useful information
  3. Why can compression harm an answer?
    • It may remove an exception, constraint, or rare detail needed for the task
    • It always increases context length
    • It prevents models from reading text
    • It changes only the output format
  4. When can compression increase end-to-end latency?
    • Whenever prompts are compressed
    • Only when the answer is correct
    • When compressor overhead is greater than the time saved in model processing
    • When a prompt contains no text
  5. What should be measured when evaluating a compression method?
    • Only the compression ratio
    • Token reduction, compressor time, cost, latency, and task quality
    • Only model output length
    • Only GPU memory
  6. Which prompt content should a compressor preserve carefully?
    • Task goals, constraints, exceptions, definitions, and needed evidence
    • Repeated filler at all costs
    • Only the first sentence
    • Every punctuation mark regardless of meaning
  7. What may happen if a prompt is already short or cached?
    • The compressor always makes the request free
    • The model stops using its cache
    • The prompt becomes longer by definition
    • Compression overhead may exceed the remaining savings
  8. How should a team verify a compressed prompt?
    • Assume fewer tokens means equal quality
    • Test only the example used to create it
    • Compare it with the original on representative held-out examples
    • Skip version tracking