Establish the statistical significance of RaRE’s GPQA Diamond improvement

Establish, through additional training seeds, whether RaRE’s improvement over single-judge Rubrics as Rewards variants on GPQA Diamond is statistically significant and supports a definitive comparison.

Background

RaRE outperforms the single-judge variants on the reported GPQA Diamond results, but the paper states that this improvement is not statistically significant across the five training seeds used. Additional seeds are identified as a way to narrow the confidence intervals and determine whether the observed advantage reflects a reliable effect rather than sampling variation.