List your company

Long-context Language Modeling on LongBench

46.23LongBench Average ScoreAdaKV w/ CriticalKV
16.43424.169531.90539.6405Feb 6, 2025Apr 25, 2025Jul 12, 2025Sep 29, 2025Dec 16, 2025Mar 4, 2026May 22, 2026

Evaluation results

MethodLinks
AdaKV w/ CriticalKVBackbone=Mistral-7B, Cache budget=40%2025.0246.23———————
HeadKV w/ CriticalKVBackbone=Mistral-7B, Cache budget=40%2025.0246.1———————
SnapKV w/ CriticalKVBackbone=Mistral-7B, Cache budget=40%2025.0245.35———————
SnapKV w/ OBCache (Joint)Backbone=Mistral-7B, Cache budget=40%2025.0244.23———————
SnapKVBackbone=Mistral-7B, Cache budget=40%2025.0244.03———————
Full KVKV Compression=Disabled2026.0519.3———————
AMS-TOVAKV Compression=Enabled2026.0519.3———————
AMS-ExpectedKV Compression=Enabled2026.0519.25———————
StreamingLLMKV Compression=Enabled2026.0519.19———————
TOVAKV Compression=Enabled2026.0519.12———————
PyramidKVKV Compression=Enabled2026.0519.02———————
AdaKV-ExpE2KV Compression=Enabled2026.0517.58———————
Qwen3-1.7BAttention Mechanism=Global-Attn2026.01—42.0486.223.2372.6731.1110059.21
Qwen3-1.7BAttention Mechanism=Window-Attn, Window Size (W)=20482026.01—32.5161.5610.4352.6713.086.229.41
Qwen3-1.7BAttention Mechanism=Window-Attn + ROSA2026.01—40.5384.3423.766826.1910057.14