Unverified50% confidenceFactExact time
DeepSeek-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.
1
Sources
50%
Confidence
Long-term
Relevance
9/23/2026
First Seen
Sources
Related Entities
Related Claims
VerifiedGRPO 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.76% 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）74% similarVerifiedGRPO由DeepSeek团队系统化提出并在训练DeepSeek-R1系列模型时得到大规模验证74% similarUnverifiedDeepSeek R1的强化学习奖励信号主要来自数学答案的正确性验证，而非工具调用格式的规范性，因此未针对Function Call专项对齐73% 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73% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/944590API
curl https://kongchang.com/api/v1/knowledge/claims/944590MCP
get_claim(id=944590)