Unverified75% confidenceFactTime unknown
在RLHF奖励模型训练中，标注者面对情感敏感话题时倾向于惩罚直言不讳的回应，即使这些回应在事实层面更准确
1
Sources
75%
Confidence
Long-term
Relevance
6/1/2026
First Seen
Sources
Related Entities
Related Claims
VerifiedHuman evaluators in the RLHF training process exhibit systematic biases, tending to prefer responses that are politely worded, affirmative in attitude, and gentle in tone, which causes models to reduce candor in the domains of spirituality and emotion.84% similarUnverified在RLHF流程中，人类标注员在灵性和关系等情感敏感话题上更容易将「温暖共情」等同于「高质量回答」，导致标注偏差80% similarUnverified在RLHF微调阶段惩罚情绪激烈的输出，会导致模型在转述用户评论时自动降低语言烈度78% similarUnverifiedRLHF倾向于产生讨好型输出，即模型学会说用户想听的话，这在青少年情绪困扰场景中可能过度验证负面情绪而非引导积极应对76% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/33486API
curl https://kongchang.com/api/v1/knowledge/claims/33486MCP
get_claim(id=33486)