为 LLM 生成的文本测验添加水印

当您参加测验时，就会出现答案和解释。

1. 引用的方案中的软绿名单水印如何影响词元生成？
   • By requiring a visible watermark sentence in every output
   • By encrypting every word into unreadable bytes
   • By adding a preference bias to green-token logits
   • By guaranteeing that every generated token belongs to the green list
2. 应用基于密钥的绿色词元水印检测器时，什么必须匹配？
   • The full original training dataset in every case
   • Physical access to the generation GPU
   • Every prompt the provider has ever received
   • The candidate text, scheme settings, tokenization and relevant key
3. 为什么更长的符合条件的段落能够为统计文本水印提供更多证据？
   • A law requires every long passage to be flagged
   • More scored tokens can make a persistent signal easier to distinguish from null variation
   • Every longer passage must have better factual accuracy
   • All short texts are necessarily unwatermarked
4. 哪种已发布的文本水印使用了锦标赛采样（tournament sampling）并针对 Gemini 文本进行了记录？
   • SynthID-Text
   • A standard PDF font renderer
   • The TCP packet checksum
   • An ordinary dictionary spell checker
5. 削弱或删除文本水印的已知方法是什么？
   • 大声朗读
   • 将其另存为 PDF
   • 用不同的颜色打印
   • 大量释义或翻译
6. What input helps reproduce green-list selection in the original context-dependent scheme?
   • A required live clock reading
   • The relevant preceding token context under the scheme’s seeding rule
   • Only the author’s IP address
   • Only the displayed font color
7. Using z=(G−gamma×T)/sqrt(T×gamma×(1−gamma)), what is z for T=100, gamma=0.5 and G=60?
   • 2
   • 1
   • 5
   • 10
8. Why does failing to detect one particular watermark not establish that a passage is human-written?
   • All machine-generated text must contain that exact keyed signature
   • Every negative score proves human authorship mathematically
   • Machine-generated text may not carry the tested watermark or matching configuration
   • The detector has necessarily recovered the original prompt