KV cache reuse efficiency on WildChat
50.2Match RateCachePrune
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| CachePrunePrivacy setting=disabled, Reuse granularity=variable-length fine-grained, Model=Mistral-7B2026.05 | 50.2 | 134 | |
| EPICReuse granularity=fixed-size chunks, Chunk size=128 tokens, Model=Mistral-7B2026.05 | 25.1 | 165 | |
| CacheBlendReuse granularity=fixed-size chunks, Chunk size=128 tokens, Model=Mistral-7B2026.05 | 22.7 | 171 | |
| CacheCraftReuse granularity=fixed-size chunks, Chunk size=128 tokens, Model=Mistral-7B2026.05 | 22.5 | 178 |