Quiz lié à un guide · Moyen Niveau

Quiz de texte généré par Watermarking LLM

Comparez les filigranes de texte statistiques, avec biais de jeton vert, détection de clé, SynthID-Text, limites d'édition et calcul statistique z travaillé.

Chemins de guidage associésFiligrane du texte généré par LLM
Question 1 de 8

How does the soft green-list watermark in the cited scheme influence token generation?

Prévisualisez chaque question8 des questions

Les réponses et les explications apparaissent au fur et à mesure que vous répondez au quiz.

  1. How does the soft green-list watermark in the cited scheme influence token generation?
    • 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. What must match when applying a detector for a keyed green-token watermark?
    • 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. Why can a longer eligible passage provide more evidence for a statistical text watermark?
    • 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. Which published text watermark uses tournament sampling and is documented for Gemini text?
    • SynthID-Text
    • A standard PDF font renderer
    • The TCP packet checksum
    • An ordinary dictionary spell checker
  5. Quelle est la méthode connue pour affaiblir ou supprimer un filigrane de texte ?
    • Le lire à haute voix
    • L'enregistrer au format PDF
    • L'imprimer dans une couleur différente
    • Paraphrase ou traduction lourde
  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