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。
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。
- 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
- 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
- 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
- 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
- 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
- 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
- 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
- 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