Web Browsing Comprehension on BrowseComp-Plus (BCP) 830-query (evaluation)
23.44Raw KV ReadsSnapKV
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| SnapKVBackbone=Qwen2.5-14B, KV Cache Budget (C)=8k2026.06 | 23.44 | 32.89 | 7.71 | |
| H2OBackbone=Qwen2.5-14B, KV Cache Budget (C)=16k2026.06 | 22.24 | 68.8 | 15.3 | |
| StreamingLLMBackbone=Qwen2.5-14B, KV Cache Budget (C)=16k2026.06 | 21.61 | 74.7 | 16.14 | |
| SnapKVBackbone=Qwen2.5-14B, KV Cache Budget (C)=16k2026.06 | 21.41 | 73.13 | 15.66 | |
| H2OBackbone=Qwen3-8B, KV Cache Budget (C)=8k2026.06 | 21.37 | 45.66 | 9.76 | |
| TrimKVBackbone=Qwen3-8B, KV Cache Budget (C)=8k2026.06 | 21.25 | 51.57 | 10.96 | |
| Full-cacheBackbone=Qwen2.5-14B, KV Cache Budget (C)=16k2026.06 | 21.03 | 97.95 | 20.6 | |
| IntentKV-Phase-2Backbone=Qwen2.5-14B, KV Cache Budget (C)=8k2026.06 | 20.53 | 90.36 | 18.55 | |
| IntentKV-Phase-2Backbone=Qwen2.5-14B, KV Cache Budget (C)=16k2026.06 | 20.48 | 71.2 | 14.58 | |
| SnapKVBackbone=Qwen3-8B, KV Cache Budget (C)=8k2026.06 | 20.06 | 45.66 | 9.16 | |
| H2OBackbone=Qwen3-8B, KV Cache Budget (C)=16k2026.06 | 19.36 | 71.57 | 13.86 | |
| H2OBackbone=Qwen2.5-14B, KV Cache Budget (C)=8k2026.06 | 19.3 | 27.47 | 5.3 | |
| TrimKVBackbone=Qwen3-8B, KV Cache Budget (C)=16k2026.06 | 18.42 | 71.93 | 13.25 | |
| StreamingLLMBackbone=Qwen3-8B, KV Cache Budget (C)=16k2026.06 | 18.1 | 79.88 | 14.46 | |
| StreamingLLMBackbone=Qwen3-8B, KV Cache Budget (C)=8k2026.06 | 18.05 | 65.42 | 11.81 | |
| SnapKVBackbone=Qwen3-8B, KV Cache Budget (C)=16k2026.06 | 17.29 | 73.86 | 12.77 | |
| IntentKV-Phase-2Backbone=Qwen3-8B, KV Cache Budget (C)=16k2026.06 | 17.08 | 87.47 | 14.94 | |
| Full-cacheBackbone=Qwen3-8B, KV Cache Budget (C)=16k2026.06 | 16.92 | 89.04 | 15.06 | |
| IntentKV-Phase-2Backbone=Qwen3-8B, KV Cache Budget (C)=8k2026.06 | 16.67 | 84.58 | 14.1 | |
| StreamingLLMBackbone=Qwen2.5-14B, KV Cache Budget (C)=8k2026.06 | 16.54 | 49.52 | 8.19 |