Reasoning on GPQA Diamond (Acc, #Tok, ∆%)
58.8AccuracyVanilla
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| VanillaReasoner=Qwen3-8B2026.06 | 58.8 | 9,654 | — | |
| ACTSReasoner=Qwen3-8B2026.06 | 58.6 | 6,615 | 31.5 | |
| ACTSπSFTReasoner=Qwen3-8B2026.06 | 58.4 | 7,761 | 19.6 | |
| CoDReasoner=Qwen3-8B2026.06 | 58.3 | 5,789 | 40 | |
| DEERReasoner=Qwen3-8B2026.06 | 57.1 | 7,671 | 20.5 | |
| NoThinkReasoner=Qwen3-8B2026.06 | 48.1 | 1,628 | 83.1 | |
| BudgetGuidanceReasoner=Qwen3-8B2026.06 | 47.8 | 2,308 | 76.1 | |
| ACTSReasoner=DeepSeek-R1-Distill-Qwen-7B2026.06 | 46.8 | 4,404 | 47.7 | |
| ACTSπSFTReasoner=DeepSeek-R1-Distill-Qwen-7B2026.06 | 44.4 | 4,656 | 44.7 | |
| VanillaReasoner=DeepSeek-R1-Distill-Qwen-7B2026.06 | 38.9 | 8,422 | — | |
| CoDReasoner=DeepSeek-R1-Distill-Qwen-7B2026.06 | 36.7 | 5,402 | 35.9 | |
| BudgetGuidanceReasoner=DeepSeek-R1-Distill-Qwen-7B2026.06 | 36.5 | 4,565 | 45.8 | |
| NoThinkReasoner=DeepSeek-R1-Distill-Qwen-7B2026.06 | 35.4 | 700 | 91.7 | |
| DEERReasoner=DeepSeek-R1-Distill-Qwen-7B2026.06 | 34.9 | 3,825 | 54.6 | |
| ACTSπSFTReasoner=DeepSeek-R1-Distill-Qwen-1.5B2026.06 | 30.1 | 3,758 | 59.5 | |
| ACTSReasoner=DeepSeek-R1-Distill-Qwen-1.5B2026.06 | 30.1 | 3,766 | 59.4 | |
| CoDReasoner=DeepSeek-R1-Distill-Qwen-1.5B2026.06 | 20.9 | 7,993 | 13.9 | |
| NoThinkReasoner=DeepSeek-R1-Distill-Qwen-1.5B2026.06 | 19.2 | 977 | 89.5 | |
| VanillaReasoner=DeepSeek-R1-Distill-Qwen-1.5B2026.06 | 18.2 | 9,285 | — | |
| DEERReasoner=DeepSeek-R1-Distill-Qwen-1.5B2026.06 | 12.3 | 4,769 | 48.6 |