引導式測驗 · 中等 等級

基於熵的抽樣測驗

區分下一個標記熵和真實置信度,解釋熵和變熵,並檢查自適應解碼啟發式的限制。

相關引導路徑基於熵的採樣
問題 1 的 8

在感知熵的解碼器中,下一個 token 的熵測量什麼?

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. 在感知熵的解碼器中,下一個 token 的熵測量什麼?
    • How spread out probability is across the model’s possible next tokens
    • Whether the model’s factual claim is true
    • 模型中參數的數量
    • The total number of tokens already generated
  2. A low-entropy next-token distribution most directly indicates that:
    • The predicted answer is guaranteed factual
    • Probability is concentrated on a small set of tokens
    • The model has no uncertainty about the entire conversation
    • Every token in the vocabulary is equally likely
  3. 下一個標記分佈的香農熵計算如下:
    • 對數總和
    • 最大機率減去最小機率
    • -sum p_i 透過令牌記錄 p_i
    • 詞彙表中標記的數量
  4. 在某些感知熵的解碼方案中,varentropy 測量什麼?
    • The variance of model parameter counts
    • The factual accuracy of the next sentence
    • 訓練紀元數
    • Variation in token-level surprisal values under the distribution
  5. 當感知熵的系統看到高熵的下一個 token 分佈時,它可以合理地得出什麼結論?
    • The distribution is more spread out and may merit a task-specific decoding policy
    • The model is certainly hallucinating
    • The highest-probability token is false
    • The output must be stopped immediately
  6. 使用熵來引導生成有什麼實際限制?
    • Entropy cannot be computed from probabilities
    • A chosen entropy policy does not by itself guarantee fewer factual errors
    • Entropy removes the need for model weights
    • Entropy forces all tokens to have equal probability
  7. 低熵的模型回答同意了使用者的錯誤主張。熵值能確立什麼?
    • That the claim has been independently verified
    • That every other candidate is impossible
    • Only that the model’s next-token distribution is concentrated, not that the claim is true
    • That the answer came from a reliable source
  8. How should a team choose thresholds for an entropy-triggered retrieval policy?
    • Assume one threshold works for all models and contexts
    • Choose the threshold from the model name
    • Treat every high-entropy token as a proven error
    • Evaluate the policy on representative tasks and measure whether retrieval improves outcomes