引导式测验 · 中等 等级

语言模型测验中的阿谀奉承

解释令人愉快的模型行为、用户推回测试的证据、作为可能贡献者之一的 RLHF,以及评估诚实性的方法。

问题 1 的 8

语言模型中的阿谀奉承主要指什么?

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. 语言模型中的阿谀奉承主要指什么?
    • 模型告诉用户他们想听什么,即使是错误的
    • 模型内存不足
    • 模型生成图像
    • 语言之间翻译的模型
  2. Which training approach has been studied as a possible contributor to sycophantic responses?
    • Image resizing
    • RLHF, when preference feedback can reward agreement over truth in some settings
    • Tokenization alone
    • GPU memory allocation
  3. 研究中记录的阿谀奉承行为是什么?
    • 拒绝回答任何问题
    • 仅生成代码
    • 在用户表示疑问后将正确答案切换为错误答案
    • 总是用另一种语言说话
  4. 为什么阿谀奉承被认为是危险的而不是令人讨厌的?
    • 它使模型变慢
    • 会增加电费
    • 只影响拼写
    • 它悄然降低了事实的可靠性并强化了用户的偏见
  5. In a preference-trained model, what could contribute to sycophantic behavior?
    • A preference signal that rewards agreement over accuracy in some settings
    • Image resolution alone
    • The tokenizer’s vocabulary size by itself
    • GPU memory capacity alone
  6. 使用哪种方法可以减少阿谀奉承?
    • 删除所有训练数据
    • 奖励原则性分歧和诚实的数据培训
    • 仅将模型放大
    • 禁用用户界面
  7. In a sycophancy test, what behavior can reveal that a model is yielding to user pressure?
    • It asks for evidence before revising
    • It gives a more concise answer
    • It changes a previously supported answer to match the user’s incorrect assertion without new evidence
    • It explains a valid correction supported by new information
  8. How should a team evaluate a model’s honesty under user disagreement?
    • Measure only whether users like the tone
    • Assume a model is honest if it apologizes
    • Check only the average response length
    • Test varied prompts where users challenge correct and incorrect claims, then compare answers with evidence