Unverified85% confidenceFactTime unknown
RLHF对齐存在'过度对齐'问题，模型在完全无害的场景下也会过度谨慎，频繁拒绝正常的创意写作和角色扮演请求
1
Sources
85%
Confidence
Long-term
Relevance
6/3/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedThe over-alignment problem causes models to frequently reject normal creative writing and role-playing requests even in completely harmless scenarios.83% similarUnverifiedAI model over-refusal is a known issue in safety alignment research, where RLHF and Constitutional AI training cause models to incorrectly classify harmless requests as requiring refusal70% similarUnverifiedRLHF中人类评审倾向给流畅自信全面的答案更高分而较少关注事实核查，导致模型获得隐性激励表现得有把握比实际有把握更重要70% similarUnverifiedRLHF's optimization for fluency and confident expression may in turn reinforce confident but incorrect statements.67% similarUnverified最新研究"The Danger of Overthinking"指出能做脑内模拟的模型存在想太多的问题，有些模型反复思考却不行动甚至还没尝试就自我放弃64% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/38474API
curl https://kongchang.com/api/v1/knowledge/claims/38474MCP
get_claim(id=38474)