List your company

Long-generation reasoning on MATH500 (pass@1)

0.93pass@1ParisKV
0.439120.566560.6940.82144Feb 7, 2026

Evaluation results

MethodLinks
ParisKVBackbone=Qwen-3-8B2026.020.93
ParisKVBackbone=Qwen-3-4B2026.020.928
Qwen-3-4BAttention=Full attention2026.020.886
Qwen-3-8BAttention=Full attention2026.020.874
ParisKVBackbone=DS-R1-Llam-8B2026.020.844
DS-R1-Llam-8BAttention=Full attention2026.020.804
PQCacheBackbone=Qwen-3-8B2026.020.6921
PQCacheBackbone=DS-R1-Llam-8B2026.020.6675
PQCacheBackbone=Qwen-3-4B2026.020.588
MagicPIGBackbone=DS-R1-Llam-8B2026.020.49
MagicPIGBackbone=Qwen-3-4B2026.020.464
MagicPIGBackbone=Qwen-3-8B2026.020.458