Unverified50% confidenceFactExact time
Most mainstream large models today are trained on similar internet corpora and adopt similar architectures and alignment methods (such as RLHF), potentially inheriting the same biases and tendencies
1
Sources
50%
Confidence
Long-term
Relevance
9/14/2026
First Seen
Sources
Related Entities
Related Claims
Unverified不同大语言模型因训练数据分布、RLHF偏好对齐策略和模型架构差异，在同一问题上呈现风格迥异的回应倾向75% similarVerified大语言模型在RLHF训练阶段，因人类标注员倾向于给更详细、更礼貌的回答打高分，导致模型学会了过度冗长的表达模式74% similarUnverifiedRLHF has now become one of the standard training procedures for almost all mainstream large language models.73% similarUnverifiedA 2023 paper by OpenAI researchers found that models trained with RLHF tend to express agreement even when faced with clearly incorrect views (sycophancy bias).72% similarUnverified提示词与模型版本存在深度绑定关系，不同版本模型在RLHF训练数据、指令微调策略上的差异会导致相同提示词产生不同输出，这种现象被称为「提示词漂移」71% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/920827API
curl https://kongchang.com/api/v1/knowledge/claims/920827MCP
get_claim(id=920827)