Scientific Reasoning on GPQA-D (pass@1 Accuracy, Standard Error)
59.1Pass@1 AccuracyR-KV
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| R-KVBase Model=Qwen3-14B, Cache Compression=~ 4x2026.06 | 59.1 | 3 | |
| VASE-ATTNVBase Model=Qwen3-14B, Cache Compression=~ 4x2026.06 | 57.4 | 3 | |
| VASE-DKVBase Model=Qwen3-14B, Cache Compression=~ 4x2026.06 | 56.8 | 3 | |
| CurDKVBase Model=Qwen3-14B, Cache Compression=~ 4x2026.06 | 48.5 | 3.1 |