Unverified60% confidenceFactTime unknown
OpenAI和Anthropic等公司通过收集人类标注员对模型输出的偏好排序数据，训练奖励模型，再用PPO等强化学习算法微调语言模型进行安全对齐
2
Sources
60%
Confidence
Long-term
Relevance
6/3/2026
First Seen
Sources
高自由度AI伴侣聊天软件实测：无限制角色扮演对话体验
bilibili风铃草不夜侯
Related Entities
Cite This Claim
Stable URI
https://kongchang.com/claim/38467API
curl https://kongchang.com/api/v1/knowledge/claims/38467MCP
get_claim(id=38467)