Reward-wise QA fairness and alignment on OQA benchmark leaderboard · SOTA2 Research

Reward-wise QA fairness and alignment on OQA

0.927JS Divergence (FI)BASE

Evaluation Results

Method	Links
BASE2026.04		0.927	0.9221	0.8886	0.614	0.681	0.629	0.499	0.561	0.477	0.989
SFT2026.04		0.9283	0.9247	0.8983	0.654	0.722	0.683	0.539	0.601	0.532	0.988
BASE2026.04		0.9285	0.9308	0.9016	0.65	0.763	0.679	0.531	0.658	0.551	0.99
BASE2026.04		0.9309	0.9274	0.9095	0.684	0.741	0.695	0.571	0.623	0.555	0.967
SFT2026.04		0.9318	0.9337	0.9288	0.768	0.835	0.835	0.62	0.714	0.7	0.979
SFT2026.04		0.9329	0.9384	0.9105	0.704	0.781	0.699	0.601	0.653	0.583	0.967
PPO/Was.2026.04		0.961	0.9596	0.9448	0.722	0.786	0.742	0.675	0.734	0.625	0.966
PPO/Was.2026.04		0.9621	0.9644	0.9632	0.797	0.86	0.876	0.735	0.823	0.819	0.973
PPO/Was.2026.04		0.9621	0.9645	0.9617	0.801	0.859	0.877	0.735	0.824	0.806	0.996
PPO/Was.2026.04		0.9902	0.9926	0.9828	0.748	0.823	0.812	0.685	0.781	0.707	0.921
APPA2026.04		0.9909	0.9887	0.9717	0.722	0.78	0.72	0.686	0.735	0.596	1
PPO/Was.2026.04		0.9918	0.9937	0.9902	0.803	0.856	0.878	0.732	0.811	0.792	1
APPA2026.04		0.9922	0.994	0.9923	0.816	0.872	0.872	0.765	0.841	0.815	1
PPO/Was.2026.04		0.9923	0.9946	0.9925	0.796	0.858	0.869	0.736	0.828	0.816	0.991
APPA2026.04		0.9924	0.9942	0.9939	0.814	0.872	0.871	0.766	0.842	0.823	0.985