covariate-adjustment

This paper studies whether training logs from stochastic runs can improve the precision of model comparisons via arm-specific covariate adjustment, finding that simple adjustments can reduce uncertainty, though careful covariate selection is needed to avoid noise.