引導式測驗 · 中等 等級

聊天機器人模擬使用者測試測驗

與模擬使用者探討多回合聊天機器人行為,同時了解模擬器差異和手動驗證的需求。

問題 1 的 8

What can simulated-user testing add to chatbot evaluation?

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。

  1. What can simulated-user testing add to chatbot evaluation?
    • Multi-turn paths that isolated prompts may miss
    • Proof the chatbot matches every customer preference
    • A replacement for all user studies
    • Removal of model variability
  2. What should a simulated user’s scenario specify?
    • Only age and occupation
    • A goal, relevant context, available information and possible interaction path
    • A positive review of the bot
    • Only the expected final answer
  3. How should a team treat a problem found by a simulated user?
    • As candidate issue to inspect and validate
    • As prevalence data about all users
    • As confirmed production failure
    • As a reason to skip transcript review
  4. What limitation did the 2026 tau-Bench study report?
    • All simulators gave identical results
    • Human participants always performed worse
    • Simulation was reliable for every subgroup
    • Simulator choice changed results and subgroup calibration differed in that study
  5. Why test with more than one simulator setup?
    • To eliminate any need for human research
    • To make every transcript identical
    • To see whether findings depend on one simulator model or prompt
    • To select the run with highest success
  6. After a simulation raises a usability concern, what next step should the team take?
    • Validate important concerns with representative human users
    • Publish the synthetic transcript as a testimonial
    • Assume the simulator represents the audience
    • Delete conflicting transcripts
  7. How should teams combine simulated and human results?
    • Pool them without labels
    • Keep evidence sources distinct and compare simulation with human evidence
    • Replace each person with ten simulated turns
    • Report only the simulator score
  8. Which use best fits simulated-user testing?
    • Estimating population complaint prevalence without representative sampling
    • Determining community preferences without involving people
    • Exploring repeatable conversation paths and generating issues to validate
    • Certifying accessibility for all users