KV cache compression on 5 sequences Seq 1 (hard)
0.117KL Divergence (Seq 1 hard)KVSCULPT
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| KVSCULPTCompression ratio (r)=0.3, Retain zone (m)=256, Optimization method=L-BFGS, Synthetic future queries=128, V solver=ridge regression2026.03 | 0.117 | 1.2 | |
| Select+FitCompression ratio (r)=0.3, Retain zone (m)=256, V solver=ridge regression2026.03 | 0.139 | — |