Interaction between correlated verifier errors and group size

Characterize how correlated verifier errors interact with group size in group-relative policy optimization.

Background

Group Relative Policy Optimization normalizes rewards within groups of sampled responses. The paper emphasizes that verifier errors may be correlated across responses, for example because of shared prompt difficulty or common judge behavior. Existing analyses largely assume simpler noise structures, so the effect of error correlation as the number of responses per group changes remains unresolved.

References

Empirically, verifier error should be measured on the policy's own outputs as training proceeds, the interaction between correlated errors and group size remains unexplored,

— Checking the Checkers: A Synthesis of Verifier Noise and Robustness in RLVR  (2610.17824 - Gupta et al., 13 Oct 2026) in Section 8, “Open Problems”