Mathematical Reasoning on AIME 26 (pass@1 accuracy, standard error)
72.3Pass@1 AccuracyVASE-ATTNV
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| VASE-ATTNVBase Model=Qwen3-14B, Cache Compression=~ 4x2026.06 | 72.3 | 6.5 | |
| VASE-DKVBase Model=Qwen3-14B, Cache Compression=~ 4x2026.06 | 68.8 | 6.9 | |
| CurDKVBase Model=Qwen3-14B, Cache Compression=~ 4x2026.06 | 61.3 | 7.3 | |
| R-KVBase Model=Qwen3-14B, Cache Compression=~ 4x2026.06 | 60.2 | 7.4 |