ResearchBenchmarksMathematical Reasoning on AIME 2025 (Pass@1, Timing, and Steps)Follow67.9Pass@1PS-ada56.66859.58462.565.416May 6, 2026Evaluation ResultsMethodMethodLinksPass@1Eval E2E Time (h)Train E2E Time (h)Eval StepsTrain StepsAvg Step Time (s)Validation AccuracyPS-adaModel=Qwen3-8BModel=Qwen3-8B2026.0567.936.931.319016169977.7PS-adaModel=Qwen3-4BModel=Qwen3-4B2026.0566.229.523.114011075780.2BaselineModel=Qwen3-4BModel=Qwen3-4B2026.055949.545.423021177566BaselineModel=Qwen3-8BModel=Qwen3-8B2026.0557.145.443.826025162861