Verified65% confidenceFactExact time
GRPO is a reinforcement learning algorithm proposed by the DeepSeek team that eliminates the Critic Model from PPO, computing intra-group relative rewards as a baseline by generating a set of responses to the same question.
3
Sources
65%
Confidence
Long-term
Relevance
8/6/2026
First Seen
Sources
Related Claims
UnverifiedQwen2.5后训练首次在在线阶段使用GRPO，GRPO是DeepSeek团队提出的PPO轻量化变体，去掉独立价值网络，用组内平均奖励作为基线79% similarVerifiedGRPO (Group Relative Policy Optimization), proposed by DeepSeek, requires no separate training of a value network and has become the mainstream algorithm choice for current agentic RL training.78% 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78% similarUnverifiedDeepSeek团队报告，仅用GRPO在数学和代码数据上进行强化学习而不经过监督微调，模型便自发出现反思、回溯验证等链式推理行为（aha moment）77% similarUnverifiedDeepSeek-R1's training report documented the practice of combining GRPO with a reward function: correct answers receive +1 point, while format errors or incorrect answers receive 0 or negative points.76% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/697161API
curl https://kongchang.com/api/v1/knowledge/claims/697161MCP
get_claim(id=697161)