Expired50% confidenceBenchmarkExact time
The JustGRPO method achieves an accuracy of 89.1% on the GSM8K mathematical reasoning benchmark, surpassing all reinforcement learning methods specifically designed for dLLMs over the past year.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
7/17/2026
First Seen
Valid until: 10/15/2026(expired)
Sources
Related Claims
UnverifiedDeepSeek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）68% 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66% 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66% similarUnverifiedDeepSeek and Peking University jointly open-sourced the DSpark inference framework, which focuses on speculative decoding technology, improving inference speed by 85% and generation speed by 6–8% without replacing model weights.62% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/542329API
curl https://kongchang.com/api/v1/knowledge/claims/542329MCP
get_claim(id=542329)