Mathematical Reasoning on AIME25 (Acc, Gen. Len)
63.3AccuracyFull Attention
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Full AttentionBackbone=Qwen3-4B, Constraint=Full (No Budget)2026.06 | 63.3 | 17,998.7 | |
| HARD-KVBackbone=Qwen3-4B, Selection Method=RKV, Top-k Budget=4096, Top-p Budget=0.902026.06 | 53.4 | 28,202.8 | |
| DASH2026.07 | 50.8 | — | |
| DR-GRPO2026.07 | 50.3 | — | |
| Nemotron-4BParameters=4B2026.07 | 46.1 | — | |
| GRPO + Brevity Bonus2026.07 | 45.9 | — | |
| GRPO2026.07 | 45.4 | — | |
| HARD-KVBackbone=Qwen3-4B, Selection Method=SnapKV, Top-k Budget=4096, Top-p Budget=0.902026.06 | 44.3 | 26,548.9 | |
| SnapKVBackbone=Qwen3-4B, Selection Method=SnapKV, Top-k Budget=40962026.06 | 43.3 | 23,654.5 | |
| RKVBackbone=Qwen3-4B, Selection Method=RKV, Top-k Budget=40962026.06 | 30.7 | 20,616.6 | |
| HARD-KVBackbone=Qwen3-4B, Selection Method=SnapKV, Top-k Budget=2048, Top-p Budget=0.902026.06 | 26.7 | 28,254.2 | |
| VanillaBackbone=Qwen3-4B, Selection Method=Vanilla, Top-k Budget=40962026.06 | 24.8 | 31,851.4 | |
| RKVBackbone=Qwen3-4B, Selection Method=RKV, Top-k Budget=20482026.06 | 21.9 | 26,836.2 | |
| HARD-KVBackbone=Qwen3-4B, Selection Method=Vanilla, Top-k Budget=4096, Top-p Budget=0.902026.06 | 20.7 | 29,966.5 | |
| HARD-KVBackbone=Qwen3-4B, Selection Method=RKV, Top-k Budget=2048, Top-p Budget=0.902026.06 | 20.7 | 31,809.6 | |
| VanillaBackbone=Qwen3-4B, Selection Method=Vanilla, Top-k Budget=20482026.06 | 16.7 | 26,617.5 | |
| RKVBackbone=Qwen3-4B, Selection Method=RKV, Top-k Budget=10242026.06 | 13.7 | 27,924.7 | |
| HARD-KVBackbone=Qwen3-4B, Selection Method=RKV, Top-k Budget=1024, Top-p Budget=0.902026.06 | 13.2 | 31,496.4 | |
| HARD-KVBackbone=Qwen3-4B, Selection Method=Vanilla, Top-k Budget=2048, Top-p Budget=0.902026.06 | 12.9 | 30,837.8 | |
| SnapKVBackbone=Qwen3-4B, Selection Method=SnapKV, Top-k Budget=20482026.06 | 12.3 | 30,040.4 | |
| SnapKVBackbone=Qwen3-4B, Selection Method=SnapKV, Top-k Budget=10242026.06 | 10.9 | 32,229.7 | |
| HARD-KVBackbone=Qwen3-4B, Selection Method=Vanilla, Top-k Budget=1024, Top-p Budget=0.902026.06 | 10.5 | 30,906.7 | |
| HARD-KVBackbone=Qwen3-4B, Selection Method=SnapKV, Top-k Budget=1024, Top-p Budget=0.902026.06 | 9.3 | 32,768 | |
| VanillaBackbone=Qwen3-4B, Selection Method=Vanilla, Top-k Budget=10242026.06 | 6.5 | 32,309.7 |