ResearchBenchmarksMathematics on AIME 2025 (Accuracy, Thinking Token Reduction (%))Mathematics on AIME 2025 (Accuracy, Thinking Token Reduction (%))Follow56.7AccuracyDART25.533.641.749.8Jun 22, 2026Evaluation resultsMethodMethodLinksAccuracyThinking Token ReductionDARTModel=Qwen3-8BModel=Qwen3-8B2026.0656.729ATModel=Qwen3-8BModel=Qwen3-8B2026.0653.3—NTModel=DeepSeek-V3.2Model=DeepSeek-V3.22026.0653.3—DARTModel=DeepSeek-V3.2Model=DeepSeek-V3.22026.065028ATModel=DeepSeek-V3.2Model=DeepSeek-V3.22026.0640—NTModel=Qwen3-8BModel=Qwen3-8B2026.0626.7—