Ikibazo gihuza ibibazo · Biragoye Urwego

Kureka AI ikavuga ko ntazi Ikibazo

Gerageza icyo kwifata bisobanura, kuki uruhushya rutari kwimenyekanisha wenyine, gusuzuma ibisubizo, inkomoko ishingiye kumurimo, hamwe ningaruka zo kwanga.

Inzira zijyanye no kuyoboraKureka LLMs Vuga Simbizi
Ikibazo 1 Bya 8

What does abstention mean in an AI answer workflow?

Reba buri kibazo8 ibibazo

Ibisubizo nibisobanuro bigaragara mugihe ufata ikibazo.

  1. What does abstention mean in an AI answer workflow?
    • Declining to give a definitive answer when support or sufficient information is missing
    • Deleting the user’s question
    • Always returning an empty string
    • Automatically asking another model
  2. What does an “it is okay to say I don’t know” instruction provide?
    • A database of verified facts
    • A guaranteed internal uncertainty detector
    • A requested response behavior, not dependable self-knowledge
    • Proof that any refusal is correct
  3. Which evaluation set best measures abstention behavior?
    • Only questions where the model refuses
    • A set without answer labels
    • Only answerable trivia questions
    • A mix of answerable and unanswerable cases with clear expected decisions
  4. Why should evaluations track unnecessary abstentions as well as unsupported answers?
    • Because all abstentions are errors
    • Because coverage never matters
    • Because unsupported answers are impossible
    • Because an assistant can over-refuse even while avoiding guesses
  5. How can an application ground an abstention decision in a document assistant?
    • Check whether allowed source passages support the requested answer
    • Ask for a higher confidence percentage
    • Hide the retrieved text from reviewers
    • Assume every retrieval result is sufficient
  6. What does the Abstain-QA study contribute, as described in the guide?
    • A legal standard for refusing questions
    • A black-box evaluation dataset spanning answerable/unanswerable questions and varied domains and task types
    • A universal model that never hallucinates
    • A new prompt syntax required by all APIs
  7. Why is a refusal phrase not proof that an answer is safe?
    • The model cannot produce text after refusing
    • It proves sources were checked
    • The phrase itself can be wrong or inappropriate for the situation
    • A refusal always reveals secrets
  8. What should happen when a consequential question lacks adequate source support?
    • Use a documented abstention or route the case to qualified review
    • Remove the source citation
    • Increase the temperature
    • The system should guess confidently