Long-generation reasoning on GPQA Diamond
72.22pass@1ParisKV
Evaluation Results
| Method | Links | |
|---|---|---|
| ParisKVBackbone=Qwen-3-4B, Top-K=1002026.02 | 72.22 | |
| Qwen-3-4BAttention=Full attention2026.02 | 64.14 | |
| ParisKVBackbone=DS-R1-Llam-8B, Top-K=1002026.02 | 57.07 | |
| ParisKVBackbone=Qwen-3-8B, Top-K=1002026.02 | 55.05 | |
| Qwen-3-8BAttention=Full attention2026.02 | 54.54 | |
| DS-R1-Llam-8BAttention=Full attention2026.02 | 49.49 | |
| PQCacheBackbone=Qwen-3-8B, Compression Ratio=20%2026.02 | 47.85 | |
| PQCacheBackbone=Qwen-3-4B, Compression Ratio=20%2026.02 | 38.38 | |
| MagicPIGBackbone=Qwen-3-8B, Retrieval Policy=Dynamic2026.02 | 33.84 | |
| MagicPIGBackbone=Qwen-3-4B, Retrieval Policy=Dynamic2026.02 | 32.32 | |
| MagicPIGBackbone=DS-R1-Llam-8B, Retrieval Policy=Dynamic2026.02 | 27.78 | |
| PQCacheBackbone=DS-R1-Llam-8B, Compression Ratio=20%2026.02 | 23.81 |