Unverified50% confidenceFactExact time
Research from OpenAI, Anthropic, and other institutions indirectly confirms that alignment training significantly compresses a model's performance space in creative writing, complex reasoning, and open-ended dialogue.
1
Sources
50%
Confidence
Long-term
Relevance
7/24/2026
First Seen
Sources
Gemini 3.0 Flash比3.5更受欢迎？揭秘AI「对齐税」真相
redditr/Bard7/12/2026
Related Claims
UnverifiedOpenAI在InstructGPT论文中证明，少量高质量人类偏好数据对模型对齐的贡献远超大规模无监督训练73% similarUnverified对比学习是 Sentence-BERT、OpenAI text-embedding 系列等主流 embedding 模型的核心训练范式72% similarUnverifiedOpenAI's training of InstructGPT and Anthropic's training of Claude both rely heavily on human-in-the-loop mechanisms.69% similarUnverifiedTraining data in the instruction tuning phase contains a large volume of creative writing, business email templates, and role-playing dialogue samples, enabling the model to master various textual formats69% similarUnverifiedOpenAI、Anthropic等公司在预训练之上叠加RLHF，将输出向人类偏好中心对齐68% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/608542API
curl https://kongchang.com/api/v1/knowledge/claims/608542MCP
get_claim(id=608542)