Science Reasoning on GPQA (Accuracy, Token count)
69.2AccuracyDTSR
Evaluation results
| Method | Links | ||
|---|---|---|---|
| DTSRBackbone=Qwen3-32B2026.04 | 69.2 | 5,677 | |
| VanillaBackbone=Qwen3-32B2026.04 | 68.2 | 6,906 | |
| DEERBackbone=Qwen3-32B2026.04 | 67.2 | 6,395 | |
| NoWAITBackbone=Qwen3-32B2026.04 | 66.7 | 5,436 | |
| DTSRBackbone=Qwen3-14B2026.04 | 66.2 | 5,916 | |
| VanillaBackbone=Qwen3-14B2026.04 | 65.7 | 7,513 | |
| NoWAITBackbone=Qwen3-14B2026.04 | 64.1 | 6,775 | |
| DEERBackbone=Qwen3-14B2026.04 | 62.6 | 7,105 | |
| VanillaBackbone=Qwen3-8B2026.04 | 60.1 | 8,861 | |
| DTSRBackbone=Qwen3-8B2026.04 | 59.6 | 7,942 | |
| DEERBackbone=Qwen3-8B2026.04 | 57.1 | 8,306 | |
| NoWAITBackbone=Qwen3-8B2026.04 | 55.6 | 6,920 | |
| NoThinkingBackbone=Qwen3-14B2026.04 | 55.6 | 1,286 | |
| NoThinkingBackbone=Qwen3-32B2026.04 | 54.6 | 561 | |
| NoThinkingBackbone=Qwen3-8B2026.04 | 53 | 1,271 | |
| CoDaPObackbone=Qwen2.5-Math-1.5B2026.06 | 24.45 | — | |
| GRPObackbone=Qwen2.5-Math-1.5B2026.06 | 19.69 | — | |
| Qwen2.5-Math-1.5B2026.06 | 9.85 | — |