List your company

Long-context retrieval and reasoning on RULER 8k context

100Score 1 (S1)Transformer
9597.5100102.5Mar 20, 2026

Evaluation results

MethodLinks
Transformer2026.0310010095873741440.6372827—
GDN2026.031006624310.424226.7112012—
GDN-GSAKV cache ratio=50%2026.0310010086861262570.2123123—
HAM Fixed τKV cache ratio=50%, Threshold type=Fixed2026.031009412280.4221437282216—
HAM Learned τKV cache ratio=50%, Threshold type=Learned2026.03100904.6380.222260.5322516—
HAM Learned routerKV cache ratio=50%, Routing mechanism=Learned router2026.031001006090456474261.12628—
HAM Learned router + EDAKV cache ratio=50%, Routing mechanism=Learned router, Exponential decay averaging (EDA)=true2026.0310010037791833290.92.62226—
Full AttentionSetting=CPT setting2026.04———————————91.38
Hybrid-KSASetting=CPT setting2026.04———————————90.53
Hybrid-LinearSetting=CPT setting2026.04———————————83.86
Hybrid-SCASetting=CPT setting2026.04———————————84.28
Hybrid-SWASetting=CPT setting2026.04———————————88.03
KSASetting=CPT setting2026.04———————————86.78