引導式測驗 · 中等 等級

思考鏈推理測驗

比較工作範例和零樣本提示,以及自我一致性、訓練有素的推理、產生的步驟上下文和解釋的忠實度。

相關引導路徑思考鏈推理
問題 1 的 8

在引用的工作解決方案研究中,鏈結思考提示會引出什麼?

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. 在引用的工作解決方案研究中,鏈結思考提示會引出什麼?
    • A guarantee that every answer is logically correct
    • Only an unchanged database lookup identifier
    • Intermediate generated steps before a final answer
    • Automatic removal of all answer text
  2. 哪個簡單的短語可以在沒有範例的情況下觸發逐步推理(零樣本 CoT)?
    • “立即答覆”
    • “簡短一點”
    • “忽略這個問題”
    • “讓我們一步一步思考”
  3. 原始論文關於規模與鏈結思考收益的結論適用於什麼範圍?
    • They prove every small model can never reason
    • They concern the tasks and model sizes actually evaluated
    • They prove every larger model must improve on every task
    • They establish that prompting changes model weights immediately
  4. 在抽樣不同推理路徑後,引用的自一致性方法會聚合什麼?
    • Their final answers
    • Only the alphabetical order of explanation words
    • Only one required greedy path
    • New model weights computed from each sampled answer
  5. 原始 DeepSeek-R1 報告的訓練研究與提供固定模型加鏈結思考提示範例有何不同?
    • It proves all displayed reasoning is a faithful internal trace
    • It requires no model training of any kind
    • It changes every prompt into a permanent tokenizer entry
    • It trains reasoning behavior through RL and a stated staged pipeline
  6. In an autoregressive generated solution, how can an earlier step affect later generation?
    • It guarantees every later inference is correct
    • Its text can become context for subsequent tokens
    • It removes all attention to the original problem
    • It necessarily changes the model’s original weights
  7. What feature distinguishes few-shot CoT exemplars from answer-only exemplars in the original prompting comparison?
    • Intermediate solution steps are included with the example answers
    • Every example omits its question
    • The prompt must retrain the model weights
    • All exemplars are only random strings with no answers
  8. What did the cited unfaithful-explanation study show about some generated reasoning explanations?
    • They always reveal every true causal influence on the model
    • They prove every generated answer is false
    • They can omit biasing input influences while rationalizing the affected answer
    • They eliminate the need for external task verification