引导式测验 · 硬 等级

推测性抽样验证测验

测试精确的推测性抽样、接受和纠正概率、草稿成本以及对其性能声明的限制。

相关引导路径推测性抽样验证
问题 1 的 8

推测性抽样背后的核心思想是什么？

预览每个问题8 问题

当您参加测验时，就会出现答案和解释。

  1. 推测性抽样背后的核心思想是什么？
    • 使用小型草稿模型提出多个令牌，并由大型模型并行验证
    • 训练一个巨大的模型来动态地跳过层
    • 提前缓存每一个可能的句子
    • 用贪婪解码代替采样以节省时间
  2. 为什么推测抽样不会降低输出质量？
    • 因为草稿模型总是正确的
    • 因为验证规则使得最终的分布与纯目标采样相同
    • 因为它只适用于简短的提示
    • 因为它平均了两个模型的输出
  3. 在精确推测采样中，当草稿标记被拒绝时会发生什么？
    • It outputs the rejected draft token unchanged
    • It discards the entire conversation and restarts
    • The algorithm samples from a corrected residual distribution at that position
    • It permanently changes the target model
  4. 在精确推测采样下，关于输出分布有什么保证？
    • It is always deterministic
    • It matches the draft model’s distribution instead
    • It is more diverse than every target-model output
    • It matches the target model’s sampling distribution
  5. 目标验证过程在精确推测采样中提供什么信息？
    • Target probabilities for proposed positions, used for acceptance tests and residual resampling
    • A guarantee that no target-model computation is needed
    • Only the draft model’s next-token distribution
    • A corrected token sequence without any sampling
  6. Which description correctly distinguishes two self-speculative approaches?
    • Both use only an unrelated small draft model
    • Medusa adds prediction heads, while EAGLE uses feature-space proposals
    • Medusa is a tokenizer and EAGLE is a reward model
    • Both remove the target model from decoding
  7. The draft assigns probability 0.8 to a token and the target assigns 0.4. What is the speculative acceptance probability?
    • 0.2
    • 0.4
    • 0.5
    • 1.0
  8. What most limits speculative sampling’s realized speedup?
    • The color of the model’s interface
    • The target model’s name length
    • Whether the prompt contains punctuation
    • The draft-model cost and how much proposed work the target accepts