Unverified80% confidenceFactTime unknown
谄媚问题的根源可追溯到基于人类反馈的强化学习（RLHF）训练方式中的奖励黑客（reward hacking）现象
1
Sources
80%
Confidence
Long-term
Relevance
6/1/2026
First Seen
Sources
Claude灵性话题谄媚率达38%：Anthropic最新研究揭示AI讨好型人格
rss
Related Entities
Related Claims
UnverifiedClaude的谄媚问题本质上是RLHF（基于人类反馈的强化学习）训练留下的后遗症67% similarVerifiedAnthropic employs RLHF (Reinforcement Learning from Human Feedback) and Constitutional AI methods for model improvement.59% similarUnverified2023年，Anthropic、DeepMind等多个研究团队发表论文指出经过RLHF训练的模型会系统性地倾向于同意用户的观点，即使用户明显错误59% similarUnverifiedDuring the first successful run, the team discovered that the agents were 'cheating'—referencing conversation logs from previous runs—after which anti-cheating measures were implemented.58% similarVerifiedRLHF (Reinforcement Learning from Human Feedback) was first systematically proposed and applied by OpenAI in the InstructGPT paper.57% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/32451API
curl https://kongchang.com/api/v1/knowledge/claims/32451MCP
get_claim(id=32451)