Reasoning on AIME 24 (Token Efficiency Metrics)
79.17Average @k AccuracyType IV (STOP)
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| Type IV (STOP)Model=DS-Qwen-3-8B2026.04 | 79.17 | 279,000 | -79.51 | |
| Type IIModel=DS-Qwen-3-8B2026.04 | 78.75 | 398,400 | -70.73 | |
| Type IIIModel=DS-Qwen-3-8B2026.04 | 78.33 | 284,300 | -79.11 | |
| Type IModel=DS-Qwen-3-8B2026.04 | 77.5 | 290,400 | -78.67 | |
| Type IV (STOP)Model=GPT-OSS-20B2026.04 | 77.5 | 184,400 | -68.98 | |
| No pruning (Baseline)Model=DS-Qwen-3-8B2026.04 | 76.93 | 1,361,000 | — | |
| Type IIModel=GPT-OSS-20B2026.04 | 76.25 | 299,800 | -49.55 | |
| No pruning (Baseline)Model=GPT-OSS-20B2026.04 | 75.26 | 594,200 | — | |
| Type IIIModel=GPT-OSS-20B2026.04 | 75 | 187,000 | -68.54 | |
| Type IModel=GPT-OSS-20B2026.04 | 73.33 | 196,600 | -66.91 | |
| Type IV (STOP)Model=DS-Qwen-2.5-7B2026.04 | 61.67 | 189,000 | -71.63 | |
| Type IIIModel=DS-Qwen-2.5-7B2026.04 | 55 | 197,400 | -70.38 | |
| No pruning (Baseline)Model=DS-Qwen-2.5-7B2026.04 | 54.69 | 666,200 | — | |
| Type IIModel=DS-Qwen-2.5-7B2026.04 | 54.58 | 312,500 | -53.09 | |
| Type IModel=DS-Qwen-2.5-7B2026.04 | 53.75 | 202,700 | -69.58 | |
| Type IV (STOP)Model=DS-Qwen-2.5-1.5B2026.04 | 37.92 | 204,300 | -73.88 | |
| Type IIIModel=DS-Qwen-2.5-1.5B2026.04 | 32.92 | 210,600 | -73.08 | |
| Type IIModel=DS-Qwen-2.5-1.5B2026.04 | 32.5 | 325,900 | -58.34 | |
| No pruning (Baseline)Model=DS-Qwen-2.5-1.5B2026.04 | 30.1 | 782,300 | — | |
| Type IModel=DS-Qwen-2.5-1.5B2026.04 | 26.25 | 218,300 | -72.09 |