List your company

Information Retrieval on NIAH

2Delta AccuracyCertified Quantized Attention
-0.080.4611.54May 20, 2026

Evaluation results

MethodLinks
Certified Quantized AttentionContext Length=32K, n=100 paired trials, Backbone=LLaMA 3.1-8B, Key Precision=INT8 (asymmetric), Value Precision=INT4 (g=16)2026.052
Certified Quantized AttentionContext Length=8K, n=100 paired trials, Backbone=LLaMA 3.1-8B, Key Precision=INT8 (asymmetric), Value Precision=INT4 (g=16)2026.050
Certified Quantized AttentionContext Length=64K, n=100 paired trials, Backbone=LLaMA 3.1-8B, Key Precision=INT8 (asymmetric), Value Precision=INT4 (g=16)2026.050