Long-context retrieval and tracking on RULER
83.01FullKVFullKV
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| FullKVModel=LLaMA-3.1-8B, Average KV Budget=5122026.06 | 83.01 | — | — | — | — | |
| FullKVModel=Qwen3-8B, Average KV Budget=5122026.06 | 80.64 | — | — | — | — | |
| CAKEModel=Qwen3-8B, Average KV Budget=5122026.06 | — | 55.59 | — | — | — | |
| PolyKVModel=LLaMA-3.1-8B, Average KV Budget=5122026.06 | — | — | 58.91 | 51.4 | 0.37 | |
| PolyKVModel=Qwen3-8B, Average KV Budget=5122026.06 | — | — | 51.61 | 33.62 | -3.98 | |
| TOVAModel=LLaMA-3.1-8B, Average KV Budget=5122026.06 | — | 58.54 | — | — | — |