| Model |
Window |
LongBench |
LEval |
|
\cdashline3-12[1pt/1pt] |
SQA |
MQA |
Summ |
Few-Shot |
Code |
Avg.
|
Closed |
QA |
Summ |
Avg. |
| LLaMA3-8B-32K |
32K |
|
|
|
|
|
|
|
|
|
|
| Token-selection-based methods |
| SnapKV |
4K |
|
|
|
|
|
|
|
|
|
|
| PyramidKV |
4K |
|
|
|
|
|
|
|
|
|
|
| Quest |
4K |
|
|
|
|
|
|
|
|
|
|
| Token-eviction-based methods |
| LM-Infite |
16+4080 |
|
|
|
|
|
|
|
|
|
|
| StreamingLLM |
16+4080 |
|
|
|
|
|
|
|
|
|
|
|
96+4000 |
|
|
|
|
|
|
|
|
|
|
| WA |
4K |
|
|
|
11.13 |
|
|
|
|
|
|
| WA + CPT |
4K |
|
|
|
|
|
|
|
|
|
|
| Layer-sharing-based methods |
| CLA |
32K |
|
|
|
|
|
|
|
|
|
|
|
\cdashline1-12
PoD (ours) |
16+4080+28K |
|
|
|
|
|
|
|
|
|
|
|
PoD+SnapKV (ours) |
4K |
|
|
|
|
|
|
|
|
|
|