Narrative QA Evaluation
27.2ScoreReFreeKV
Evaluation Results
| Method | Links | |
|---|---|---|
| ReFreeKVBackbone=Mistral-7B-Instruct, k=1, Real Budget Utilization=78.0%2025.02 | 27.2 | |
| Heavy Hitter Oracle (H2O)Backbone=Llama3-8B-Instruct, KV Cache Budget=90%2025.02 | 24.55 | |
| ReFreeKVBackbone=Llama3-8B-Instruct, k=1, Real Budget Utilization=48.7%2025.02 | 23.44 | |
| ReFreeKVBackbone=Qwen2.5-7B-Instruct, k=1, Real Budget Utilization=65.1%2025.02 | 20.66 | |
| MEMPRModel=Qwen3 30B Think2026.05 | 9.26 | |
| HumanModel=Human2026.05 | 7.96 | |
| COMPACTORModel=Llama 3.3 70B2026.05 | 5.72 |