Mathematical Reasoning on AIME 25 (Avg@32, Pass@8, Dist-N, SemDiv)
86.67Pass@8Qwen-3-4B
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| Qwen-3-4BAttention=Full attention2026.02 | 86.67 | — | — | — | |
| Qwen-3-8BAttention=Full attention2026.02 | 83.33 | — | — | — | |
| ParisKVBackbone=Qwen-3-4B2026.02 | 80 | — | — | — | |
| ParisKVBackbone=Qwen-3-8B2026.02 | 73.33 | — | — | — | |
| CraEGTemperature=0.7, Top-p=0.95, Backbone=Qwen3-1.7B2026.01 | 60.37 | 37.29 | 44.46 | 14.5 | |
| CraEGTemperature=1, Top-p=1, Backbone=Qwen3-1.7B2026.01 | 58.91 | 36.46 | 52.03 | 17.7 | |
| Standard SamplingTemperature=0.7, Top-p=0.95, Backbone=Qwen3-1.7B2026.01 | 58.69 | 35.73 | 43.18 | 14.71 | |
| Standard SamplingTemperature=1, Top-p=1, Backbone=Qwen3-1.7B2026.01 | 57.33 | 35.94 | 50.36 | 16.61 | |
| ParisKVBackbone=DS-R1-Llam-8B2026.02 | 53.3 | — | — | — | |
| DS-R1-Llam-8BAttention=Full attention2026.02 | 50 | — | — | — | |
| PQCacheBackbone=Qwen-3-8B2026.02 | 16.67 | — | — | — | |
| PQCacheBackbone=DS-R1-Llam-8B2026.02 | 13.3 | — | — | — | |
| MagicPIGBackbone=DS-R1-Llam-8B2026.02 | 13.3 | — | — | — | |
| MagicPIGBackbone=Qwen-3-8B2026.02 | 10 | — | — | — | |
| MagicPIGBackbone=Qwen-3-4B2026.02 | 6.67 | — | — | — | |
| PQCacheBackbone=Qwen-3-4B2026.02 | 3.33 | — | — | — | |
| HPDModel Family=Qwen 2.5, Student Model=1.5B2026.04 | — | 9.89 | — | — | |
| HPDModel Family=Qwen 2.5, Student Model=3B2026.04 | — | 18.13 | — | — | |
| HPDModel Family=LLaMA 3, Student Model=1B2026.04 | — | 2.08 | — | — | |
| HPDModel Family=LLaMA 3, Student Model=3B2026.04 | — | 12.71 | — | — | |
| MSModel Family=Qwen 2.5, Student Model=1.5B2026.04 | — | 1.04 | — | — | |
| MSModel Family=Qwen 2.5, Student Model=3B2026.04 | — | 2.5 | — | — | |
| MTModel Family=Qwen 2.5, Model Size=7B, Role=Teacher2026.04 | — | 27.19 | — | — | |
| MTModel Family=LLaMA 3, Model Size=8B, Role=Teacher2026.04 | — | 18.02 | — | — |