Unverified50% confidenceFactExact time
Anthropic提出了RLAIF和Constitutional AI，DeepSeek探索了GRPO算法，都是对RLHF原始框架的改进
1
Sources
50%
Confidence
Long-term
Relevance
7/19/2026
First Seen
Sources
Related Claims
UnverifiedDeepSeek探索了GRPO算法，作为对RLHF原始框架的改进尝试84% similarVerifiedDeepSeek-R1通过在代码和数学两个可验证域上大规模运行GRPO算法取得成功72% similarVerifiedGRPO（Group Relative Policy Optimization）由DeepSeek提出，无需单独训练价值网络，已成为当前智能体RL训练的主流算法选择67% similarUnverifiedGRPO和GSPO均是PPO算法的改进变体，DeepSeek-R1凭借GRPO在推理能力上实现大幅跃升67% similarUnverifiedDeepSeek R1通过GRPO算法仅用ORM即可涌现出长程推理能力67% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/561430API
curl https://kongchang.com/api/v1/knowledge/claims/561430MCP
get_claim(id=561430)