ResearchBenchmarksMachine Translation on WMT20Follow0.2RewardRLOO0.10640.13070.1550.1793Mar 5, 2025Evaluation ResultsMethodMethodLinksRewardInference TimeReward per TimePeak Memory (GB)RLOOModel=Llama 3.2 1B, Ha...Model=Llama 3.2 1B, Hardware=4 x A100 (40GB VRAM)2025.030.2360.005—ZOPrOModel=Llama 3.2 1B, Ha...Model=Llama 3.2 1B, Hardware=4 x A100 (40GB VRAM)2025.030.1560.025—PPOModel=Llama 3.2 1B, Ha...Model=Llama 3.2 1B, Hardware=4 x A100 (40GB VRAM)2025.030.111150.0015—