引導式測驗 · 中等 等級

當思維鏈受傷時測驗

測試思想鏈的定義、有限的研究結果、特定於模型的指導以及基於任務的比較(包括延遲)。

相關引導路徑當思想鏈受傷時
問題 1 的 8

What does chain-of-thought prompting ask a model to produce?

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 chain-of-thought prompting ask a model to produce?
    • Intermediate reasoning text before its answer
    • Only an image input
    • A guaranteed proof of correctness
    • A list of training examples
  2. What did the 2025 ICML study report across its six selected tasks?
    • The study tested every task category
    • CoT reduced performance on every task
    • Performance drops on three tasks and mixed effects on the others
    • CoT always improved performance
  3. What does that study establish about CoT across all AI tasks?
    • It guarantees better performance for reasoning models
    • It applies only to image generation
    • It is universally harmful
    • It can reduce performance in some settings but does not establish a universal rule
  4. Why might an explicit rationale add operational cost?
    • It reduces all input context
    • It automatically stops tool use
    • It eliminates evaluation needs
    • It can use additional output tokens and increase latency
  5. How should a model-generated explanation be treated as evidence?
    • As proof the answer is correct
    • As insufficient by itself to prove correctness or faithfully trace the internal process
    • As confirmation the model used a source
    • As a guarantee of transparency
  6. What does OpenAI’s current guidance recommend for its API reasoning models?
    • Never state constraints
    • Always demand a long chain of thought
    • Generally avoid explicit “think step by step” requests
    • Use identical prompts across every model
  7. How should a team decide whether CoT helps its own application?
    • Compare different models and prompts at once
    • Judge by which answer is longer
    • Compare it with direct prompting on representative cases using defined scoring
    • Assume the research paper covers its task
  8. Which factors should a team include when comparing prompt variants?
    • Output quality and relevant costs such as latency or token use
    • Only model marketing claims
    • Whether the explanation sounds confident
    • Only response length