Unverified50% confidenceFactExact time
GRPO's original paper and mainstream reproductions all use the within-group mean as the baseline, without exploring alternatives in depth.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
10/28/2026
First Seen
Valid until: 1/26/2027
Sources
Related Entities
Related Claims
UnverifiedThe core innovation of GRPO is the complete removal of the value network; it samples multiple responses to the same question to form a group, and uses the mean return within the group as the baseline.67% similarUnverifiedGRPO can be summarized as PPO minus the critic, plus an intra-group mean baseline.67% similarUnverifiedIn on-policy rollouts, incorrect pseudo-label words rarely appear, so the group-relative advantage always aligns with clean transcriptions; this is the mechanistic explanation for GRPO's noise robustness.66% similarUnverifiedGRPO's supervision granularity is sequence-level, with the entire response sharing a single advantage value, so it cannot indicate which step in the reasoning chain went wrong.66% similarUnverifiedIn GRPO baseline replacement experiments, the group size must be fixed; otherwise the variance of the mean baseline changes with sample size and confounds the results.65% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1109169API
curl https://kongchang.com/api/v1/knowledge/claims/1109169MCP
get_claim(id=1109169)