Mathematical Reasoning on AIME 25 (Acc, Gen. Len)
70AccuracyFull Attention
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Full AttentionSelection Method=Full Attn, Budget=None2026.06 | 70 | 16,852 | |
| RKV + HARD-KVSelection Method=RKV, Config=+ Ours, Budget=40962026.06 | 50 | 21,427 | |
| SnapKV + HARD-KVSelection Method=SnapKV, Config=+ Ours, Budget=40962026.06 | 46.7 | 22,901 | |
| Vanilla + HARD-KVSelection Method=Vanilla, Config=+ Ours, Budget=40962026.06 | 43.3 | 24,474 | |
| RKVSelection Method=RKV, Config=Base, Budget=40962026.06 | 40 | 21,275 | |
| VanillaSelection Method=Vanilla, Config=Base, Budget=40962026.06 | 33.3 | 28,701 | |
| SnapKVSelection Method=SnapKV, Config=Base, Budget=40962026.06 | 33.3 | 24,256 |