Unverified50% confidenceTradeoffExact time
When the rewards of a group's outputs are entirely identical (all correct or all incorrect), GRPO cannot compute a meaningful relative advantage within the group, the gradient is zero, and these samples contribute nothing to model parameter updates.
1
Sources
50%
Confidence
Long-term
Relevance
10/7/2026
First Seen
Sources
Related Entities
Related Claims
VerifiedGRPO eliminates the dependency on an independent critic network by sampling a group of outputs for the same question and using the within-group average reward as the baseline, compressing memory requirements from approximately 4× that of PPO to approximately 2×.67% similarUnverified对比学习中的推远操作只要求负样本相似度低于正样本，并不要求达到-1的极端值67% similarUnverifiedIf a difficult problem fails across all samples (pass@k ≈ 0), the reward variance is zero, the gradient signal nearly vanishes, and the parameter update approaches zero.66% similarUnverifiedWhen the success rate is extremely low, almost all sample rewards are identical, resulting in extremely high policy gradient variance and very low signal-to-noise ratio, preventing the model from obtaining effective update directions.63% similarUnverifiedThe model's output tokens are zero, because it only returns probabilities rather than generating text.62% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/982892API
curl https://kongchang.com/api/v1/knowledge/claims/982892MCP
get_claim(id=982892)