List your company

Long-context Understanding on RULER 8K context

54.15CWETraining–Inference Consistent Segmented Execution
-0.60613.609527.82542.0405Sep 10, 2024Dec 20, 2024Apr 1, 2025Jul 11, 2025Oct 21, 2025Jan 30, 2026May 12, 2026

Evaluation results

MethodLinks
Training–Inference Consistent Segmented ExecutionBackbone=LLaMA2-7B-32K2026.0554.15——38.67—
MInferenceBackbone=LLaMA2-7B-32K2026.0539.4——36.17—
LLaMA2-7B-32K (Vanilla Self-Attention)Backbone=LLaMA2-7B-32K, Attention=Vanilla Self-Attention2026.0539.25——35.83—
DuoAttentionBackbone=LLaMA2-7B-32K2026.0538.8——42.67—
CCA-attentionBackbone=LLaMA2-7B-32K2026.0536.25——39.5—
LLoCOContext Length=8K2024.0924.10.1901728.5
E2LLM-C+RContext Length=8K2024.0921.160.668.6125.5
StreamingLLMBackbone=LLaMA2-7B-32K2026.0520.45——38.83—
LongLoRAContext Length=8K2024.0913.438.192.210.3344
YaRNContext Length=8K2024.0910.342.211.811.6734.5
E2LLM-CContext Length=8K2024.096.33.50.611.6731
LongLLMLinguaContext Length=8K2024.095.730.444.224.3316
E2LLM-RContext Length=8K2024.091.59.50.69.3333.5