Unverified60% confidenceFactTime unknown

DeepSeek在训练阶段大量使用GRPO算法替代传统的RLHF来压缩训练成本

2

Sources

60%

Confidence

Long-term

Relevance

5/31/2026

First Seen

Sources

Codex接入DeepSeek教程：零基础搭建AI智能体完整指南

bilibili信息科技练老师

Related Entities

DeepSeekcompany GRPOconcept RLHFconcept

Related Claims

UnverifiedDeepSeek探索了GRPO算法，作为对RLHF原始框架的改进尝试81% similar VerifiedGRPO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80% similar VerifiedGRPO（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79% similar VerifiedDeepSeek-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76% similar

Cite This Claim

Stable URI

https://kongchang.com/claim/11862

API

curl https://kongchang.com/api/v1/knowledge/claims/11862

MCP

get_claim(id=11862)