Arithmetic Reasoning on KVFundaBench AR
87.79Accuracy (Exact Match)FullKV
Evaluation Results
| Method | Links | |
|---|---|---|
| FullKVBackbone=Qwen3-4B-Base2025.02 | 87.79 | |
| ChunkKVBackbone=Qwen3-4B-Base, Compression Ratio=30%, KV Cache Management Level=chunk-level2025.02 | 80.41 | |
| H2OBackbone=Qwen3-4B-Base, Compression Ratio=30%, KV Cache Management Level=token-level2025.02 | 73.78 | |
| ChunkKVBackbone=Qwen3-4B-Base, Compression Ratio=10%, KV Cache Management Level=chunk-level2025.02 | 64.41 | |
| H2OBackbone=Qwen3-4B-Base, Compression Ratio=10%, KV Cache Management Level=token-level2025.02 | 49.2 |