Mathematical Reasoning on MATH500 (Acc, ΔTok)
97.2AccuracyTHINKBRAKE
Evaluation results
| Method | Links | ||
|---|---|---|---|
| THINKBRAKEBase Model=Qwen3-32B2025.10 | 97.2 | -1.4 | |
| Qwen3-32BStrategy=Baseline2025.10 | 97 | — | |
| THINKBRAKEBase Model=Qwen3-14B2025.10 | 96.9 | -7.7 | |
| Qwen3-14BStrategy=Baseline2025.10 | 96.8 | — | |
| THINKBRAKEBase Model=Qwen3-4B-Thinking2025.10 | 96.6 | -20.4 | |
| NoWaitBase Model=Qwen3-4B-Thinking2025.10 | 96.4 | -17.9 | |
| Qwen3-4BStrategy=Baseline2025.10 | 96 | — | |
| NoWaitBase Model=Qwen3-32B2025.10 | 95.8 | -20.6 | |
| THINKBRAKEBase Model=Qwen3-4B2025.10 | 95.4 | -14.4 | |
| Qwen3-4B-ThinkingStrategy=Baseline2025.10 | 95.3 | — | |
| NoWaitBase Model=Qwen3-14B2025.10 | 95.2 | -25.1 | |
| NoWaitBase Model=Qwen3-4B2025.10 | 94.6 | -31.4 | |
| DEERBase Model=Qwen3-4B-Thinking2025.10 | 94.4 | -14.2 | |
| DeepSeek-R1-7BStrategy=Baseline2025.10 | 93.8 | — | |
| THINKBRAKEBase Model=DeepSeek-R1-7B2025.10 | 92 | -23.2 | |
| THINKBRAKE-pBase Model=Qwen3-4B-Thinking2025.10 | 91.2 | -57.9 | |
| NoWaitBase Model=DeepSeek-R1-7B2025.10 | 91.2 | -31 | |
| Dynasor-CoTBase Model=Qwen3-4B-Thinking2025.10 | 87.2 | -42.1 | |
| Phi-4-ReasoningStrategy=Baseline2025.10 | 71.1 | — | |
| NoWaitBase Model=Phi-4-Reasoning2025.10 | 70.8 | -17.7 | |
| THINKBRAKEBase Model=Phi-4-Reasoning2025.10 | 70.4 | -17.2 | |
| ThinkLessBase Model=Qwen3-4B-Thinking2025.10 | 56.4 | -100 |