ResearchBenchmarksMathematical Reasoning on GSM8K (test) (Checkpoint Accuracy Tracking)Follow37.38Accuracy (Ckpt 1)DPO+YFPO36.433636.679336.92537.1707May 12, 2026Evaluation ResultsMethodMethodLinksAccuracy (Ckpt 1)Accuracy (Ckpt 2)Accuracy (Ckpt 3)Average AccuracyBest AccuracyAvg. Accuracy DeltaBest Accuracy DeltaDPO+YFPOTrain Data=10K, λ=0.05...Train Data=10K, λ=0.05, Base Model=Qwen2-0.5B2026.0537.3835.8637.0736.7737.380.00280.0076DPO+YFPOTrain Data=10K, λ=1.0,...Train Data=10K, λ=1.0, Base Model=Qwen2-0.5B2026.0537.3836.6936.923737.380.00510.0076DPOTrain Data=10K, λ=–, B...Train Data=10K, λ=–, Base Model=Qwen2-0.5B2026.0536.4736.6236.3936.4936.62——