Reasoning on GPQA-D (avg@k, Token usage)
77.46Avg@kType IV (STOP)
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| Type IV (STOP)Model=GPT-OSS-20B2026.04 | 77.46 | 143.4 | -48.26 | |
| Type IIIModel=GPT-OSS-20B2026.04 | 68.43 | 145.9 | -47.36 | |
| Type IModel=GPT-OSS-20B2026.04 | 66.41 | 151.8 | -45.24 | |
| No pruning (Baseline)Model=GPT-OSS-20B2026.04 | 65.55 | 277.2 | — | |
| Type IV (STOP)Model=DS-Qwen-3-8B2026.04 | 63.32 | 193.5 | -70.35 | |
| Type IIIModel=DS-Qwen-3-8B2026.04 | 58.78 | 196.6 | -69.86 | |
| Type IModel=DS-Qwen-3-8B2026.04 | 57.07 | 201 | -69.2 | |
| No pruning (Baseline)Model=DS-Qwen-3-8B2026.04 | 56.87 | 652.6 | — | |
| Type IV (STOP)Model=DS-Qwen-2.5-7B2026.04 | 55.75 | 165.9 | -62.61 | |
| Type IV (STOP)Model=DS-Qwen-2.5-1.5B2026.04 | 48.42 | 179.4 | -67.43 | |
| Type IModel=DS-Qwen-2.5-7B2026.04 | 47.09 | 173.6 | -60.89 | |
| Type IIIModel=DS-Qwen-2.5-7B2026.04 | 46.02 | 166.7 | -62.43 | |
| No pruning (Baseline)Model=DS-Qwen-2.5-7B2026.04 | 45.95 | 443.8 | — | |
| Type IIIModel=DS-Qwen-2.5-1.5B2026.04 | 34.98 | 180.4 | -67.25 | |
| No pruning (Baseline)Model=DS-Qwen-2.5-1.5B2026.04 | 33.08 | 550.9 | — | |
| Type IModel=DS-Qwen-2.5-1.5B2026.04 | 32.32 | 187.1 | -66.03 |