Unverified50% confidenceOpinionExact time
RL causes models to converge on a small number of verified high-scoring strategies, making the models better but narrowing their expression.
1
Sources
50%
Confidence
Long-term
Relevance
9/15/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedRLHF本质上是在模型输出分布上施加统计偏好，而非植入真正的价值理解，导致表面对齐在面对对抗性输入时易崩塌70% similarUnverified指令微调和RLHF强化了模型的指令跟随倾向，使其在提升任务能力的同时更易被外部指令误导70% similarUnverifiedA high R² does not always mean the model is good; in complex scenarios it is still necessary to make a comprehensive judgment by combining residual analysis and other methods.70% 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」70% similarUnverifiedRL对pass@1的提升而不影响pass@k，本质上是在压缩模型的输出分布使其更集中在正确答案附近，即分布锐化69% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/923721API
curl https://kongchang.com/api/v1/knowledge/claims/923721MCP
get_claim(id=923721)