ResearchBenchmarksMathematical Reasoning on MATH 500 (pass@16 metrics)Follow59.7Mean Accuracy @16TTSFT42.85247.22651.655.974Dec 2, 2025Evaluation ResultsMethodMethodLinksMean Accuracy @16Pass Rate @16Pass Rate @1Total Tokens (M)Token SavingTTSFTBackbone=Qwen2.5-Math-...Backbone=Qwen2.5-Math-1.5B2025.1259.790.863.8——OptPO-SFTBackbone=Qwen2.5-Math-...Backbone=Qwen2.5-Math-1.5B2025.125991.260.2—25.29OptPO-SFTBackbone=Llama-3.1-8B-...Backbone=Llama-3.1-8B-Instruct2025.1243.682.847—19.84TTSFTBackbone=Llama-3.1-8B-...Backbone=Llama-3.1-8B-Instruct2025.1243.581.244.2——