Long-generation reasoning on MATH500 (pass@1)
0.93pass@1ParisKV
Evaluation Results
| Method | Links | |
|---|---|---|
| ParisKVBackbone=Qwen-3-8B2026.02 | 0.93 | |
| ParisKVBackbone=Qwen-3-4B2026.02 | 0.928 | |
| Qwen-3-4BAttention=Full attention2026.02 | 0.886 | |
| Qwen-3-8BAttention=Full attention2026.02 | 0.874 | |
| ParisKVBackbone=DS-R1-Llam-8B2026.02 | 0.844 | |
| DS-R1-Llam-8BAttention=Full attention2026.02 | 0.804 | |
| PQCacheBackbone=Qwen-3-8B2026.02 | 0.6921 | |
| PQCacheBackbone=DS-R1-Llam-8B2026.02 | 0.6675 | |
| PQCacheBackbone=Qwen-3-4B2026.02 | 0.588 | |
| MagicPIGBackbone=DS-R1-Llam-8B2026.02 | 0.49 | |
| MagicPIGBackbone=Qwen-3-4B2026.02 | 0.464 | |
| MagicPIGBackbone=Qwen-3-8B2026.02 | 0.458 |