[KongchangAI]
Unverified50% confidenceSolutionExact time

In GRPO, when training instability or abnormal advantage estimates are observed, one can check the skewness and variance of within-group rewards to determine whether the mean is dominated by extreme values.

1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
10/28/2026
First Seen
Valid until: 1/26/2027

Sources

Related Entities

Related Claims

Cite This Claim

Stable URI
https://kongchang.com/claim/1109173
API
curl https://kongchang.com/api/v1/knowledge/claims/1109173
MCP
get_claim(id=1109173)