Mathematical Reasoning on AIME25 (Acc, SU, Thr.)
823Threshold (Thr.)Top-k pruning
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| Top-k pruningSystem setting=+ Top-k pruning2026.06 | 823 | — | — | |
| VANILLA + HARD-KVTop-p budget=0.90, Config=+ Ours2026.06 | 792 | 20.7 | 86.3 | |
| VANILLA + HARD-KVTop-p budget=0.60, Config=+ Ours2026.06 | 698 | 21 | 83.1 | |
| SNAPKV + HARD-KVTop-p budget=0.60, Config=+ Ours2026.06 | 645 | 43.3 | 85.2 | |
| SNAPKV + HARD-KVTop-p budget=0.90, Config=+ Ours2026.06 | 626 | 42.6 | 97.9 | |
| CUDA GraphSystem setting=+ CUDA Graph2026.06 | 467 | — | — | |
| Base SystemSystem setting=System Throughput Baseline2026.06 | 350 | — | — | |
| SNAPKVTop-p budget=0.90, Config=Base2026.06 | 343 | 43.3 | 18.9 | |
| SNAPKVTop-p budget=0.60, Config=Base2026.06 | 305 | 16.7 | 30.1 | |
| VANILLATop-p budget=0.60, Config=Base2026.06 | 291 | 20 | 47.5 | |
| VANILLATop-p budget=0.90, Config=Base2026.06 | 289 | 33.3 | 35.5 |