List your company

Long-Context Retrieval on RULER 32K

86.65RULER 32K Retrieval ScoreHybrid-KSA
61.575668.085374.59581.1047Apr 27, 2026May 4, 2026May 12, 2026May 19, 2026May 27, 2026Jun 3, 2026Jun 11, 2026

Evaluation results

MethodLinks
Hybrid-KSASetting=CPT setting2026.0486.65
Full AttentionSetting=CPT setting2026.0484.74
KSASetting=CPT setting2026.0480.3
Hybrid-SWASetting=CPT setting2026.0478.94
MSA-PTTraining Strategy=from-scratch sparse pretraining2026.0677.5
Hybrid-SCASetting=CPT setting2026.0476.89
Hybrid-LinearSetting=CPT setting2026.0476.48
MSA-CPTTraining Strategy=sparse continued pretraining2026.0675.7
FullTraining Strategy=Full-Attention baseline2026.0675
Hybrid-KSAsetting=Train-from-scratch2026.0472.3
Hybrid-GDNsetting=Train-from-scratch2026.0470.41
Fullsetting=Train-from-scratch2026.0469.06
Hybrid-SWAsetting=Train-from-scratch2026.0467.86
Hybrid-SCAsetting=Train-from-scratch2026.0467.74
KSAsetting=Train-from-scratch2026.0462.54