Unverified50% confidenceTradeoffExact time
GRPO and its variants often encounter issues such as training collapse, reward oscillation, and late-stage performance degradation in actual training.
1
Sources
50%
Confidence
Long-term
Relevance
9/12/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedCompared to PPO, GRPO eliminates the value network (Critic) and directly compares rewards among samples within a group to estimate the advantage function, significantly reducing training computation costs.70% similarUnverifiedAfter adding the DINO objective, CAPI's originally smooth training loss curve became noticeably more erratic, and training stability decreased.69% similarUnverifiedGAIL借鉴GAN思想用对抗训练隐式恢复专家行为分布，理论基础是将模仿学习转化为占用度量匹配问题，但受GAN模式崩塌问题影响训练稳定性69% similarUnverified针对特定工具做专项强化训练的副作用是对非训练分布工具的泛化能力下降，学术上称为分布外泛化失败68% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/910345API
curl https://kongchang.com/api/v1/knowledge/claims/910345MCP
get_claim(id=910345)