Unverified50% confidenceFactExact time
RLHF的核心机制是训练一个奖励模型来预测人类对输出的偏好评分，再以奖励信号驱动策略梯度优化（通常采用PPO算法）
1
Sources
50%
Confidence
Long-term
Relevance
7/18/2026
First Seen
Sources
Related Claims
VerifiedRLHF通过收集人类偏好评分训练奖励模型，再用PPO等强化学习算法微调语言模型，但奖励模型可能被对抗性输入欺骗，且无法穷举所有有害请求表达变体84% similarUnverifiedRL-CAI阶段以AI自身的偏好判断作为奖励信号训练奖励模型，再通过PPO算法优化策略模型82% similarPartially VerifiedThe RLHF reward model is trained by having human annotators rank multiple model outputs to learn human preferences.82% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/555285API
curl https://kongchang.com/api/v1/knowledge/claims/555285MCP
get_claim(id=555285)