ResearchBenchmarksMathematical Reasoning on GSM CoTFollow83.2Accuracy (GSM CoT)χ2-DRTO64.06469.0327478.968Mar 27, 2026Evaluation ResultsMethodMethodLinksAccuracy (GSM CoT)χ2-DRTOBackbone=Llama3-8BBackbone=Llama3-8B2026.0383.2KL-DRTOBackbone=Llama3-8BBackbone=Llama3-8B2026.0379.6DPOBackbone=Llama3-8BBackbone=Llama3-8B2026.0376.4GRPOBackbone=Llama3-8BBackbone=Llama3-8B2026.0376PPOBackbone=Llama3-8BBackbone=Llama3-8B2026.0375.2RTOBackbone=Llama3-8BBackbone=Llama3-8B2026.0374.4SFTBackbone=Llama3-8BBackbone=Llama3-8B2026.0364.8