In the Evo-GDPO clipped surrogate policy objective, what does the reference distribution $p_{\text{ref}}$ represent?

In the Evo-GDPO objective, candidate policies are rewarded solely on the basis of relative within-group ranking.

In the Evo-GDPO clipped surrogate loss formula, what specific quantity is denoted by $|\mathbf{h}_i|$?

Explain how the importance weight ratio $\rho_{i,j}(\theta)$ is constructed in the Evo-GDPO clipped surrogate objective, identifying the probability distributions in its numerator and denominator.

Match each mathematical component of the Evo-GDPO objective function with its role or definition.

The Evo-GDPO objective regularizes policy updates using token-level ___ divergence computed against a frozen reference checkpoint.

Order the mathematical operations performed to compute candidate $i$'s token-level clipped surrogate term before averaging across the harness length:

How does Evo-GDPO treat Candidate A relative to Candidate B, and how does this behavior contrast with standard GRPO?