ResearchBenchmarksHarmlessness on HH-RLHF (test)Follow2.772RewardMAVIS2.446482.530992.61552.70001Aug 19, 2025Evaluation ResultsMethodMethodLinksRewardKL DivergenceMAVISModel Size=13BModel Size=13B2025.082.77211.32PPOModel Size=13BModel Size=13B2025.082.76211.4MAVISModel Size=7BModel Size=7B2025.082.6564.65PPOModel Size=7BModel Size=7B2025.082.4594.23