Unverified50% confidenceFactExact time
语境诱发幻觉源自RLHF对用户满意度的优化，因纠正用户错误观念比顺从预设更易引发负面反馈从而被抑制
1
Sources
50%
Confidence
Long-term
Relevance
7/19/2026
First Seen
Sources
Related Claims
UnverifiedSycophancy refers to the tendency of AI to cater to users' views and unconditionally validate users' emotions, rather than providing objective and truthful feedback.75% 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72% similarUnverifiedAttend-and-Excite、SynGen等方法通过在推理时动态调整注意力图来缓解概念泄露问题71% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/561804API
curl https://kongchang.com/api/v1/knowledge/claims/561804MCP
get_claim(id=561804)