Unverified50% confidenceTradeoffExact time
为对齐用户偏好而进行的优化可能让模型倾向于认同和共情对话的一方，牺牲中立与独立判断能力
1
Sources
50%
Confidence
Long-term
Relevance
9/11/2026
First Seen
Sources
Related Entities
Related Claims
Unverified随着头部模型能力普遍拉高，用户关注点正从模型强弱转向对话体验与用户理解能力76% similarUnverified迎合偏差（sycophancy bias）指模型倾向于给出用户似乎想要听到的答案74% similarUnverifiedCognition的差异化在于自主性这一高阶能力,但能否转化为用户粘性和付费意愿仍需市场验证73% similarUnverified使用同一模型既生成又自评存在自我强化偏差，引入独立评估者能提供更客观的质量判断71% similarUnverified人类标注员倾向于选择语气友好、肯定性强的回复，使奖励模型隐含了'讨好用户=高质量回复'的偏见71% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/898712API
curl https://kongchang.com/api/v1/knowledge/claims/898712MCP
get_claim(id=898712)