ResearchBenchmarksMath Reasoning on GSM DEFollow66AccuracyKL-DRTO46.86451.83256.861.768Mar 27, 2026Evaluation ResultsMethodMethodLinksAccuracyKL-DRTOBackbone=Llama3-8BBackbone=Llama3-8B2026.0366χ2-DRTOBackbone=Llama3-8BBackbone=Llama3-8B2026.0365.6PPOBackbone=Llama3-8BBackbone=Llama3-8B2026.0362DPOBackbone=Llama3-8BBackbone=Llama3-8B2026.0362RTOBackbone=Llama3-8BBackbone=Llama3-8B2026.0360.8GRPOBackbone=Llama3-8BBackbone=Llama3-8B2026.0360SFTBackbone=Llama3-8BBackbone=Llama3-8B2026.0347.6