Share your thoughts, 1 month free Claude Pro on usSee more

Needle-in-the-haystack Retrieval on Needle-in-the-haystack 2500 tokens

100Needle Accuracy

Baseline

Updated 1mo ago

Evaluation Results

Method	Links
Baseline 2026.04		100	100
Expected Attention 2026.04		100	62.5
Expected Attention 2026.04		100	75
Sub-token 2026.04		100	62.5
Sub-token 2026.04		100	75
EA + Sub-token 2026.04		100	65.6
EA + Sub-token 2026.04		100	56.2
EA + Sub-token 2026.04		100	37.5
Baseline 2026.04		85	100
Expected Attention 2026.04		85	62.5
Expected Attention 2026.04		85	75
EA + Sub-token 2026.04		85	65.6
EA + Sub-token 2026.04		85	56.2
EA + Sub-token 2026.04		84.5	37.5
Sub-token 2026.04		84	62.5
Sub-token 2026.04		84	75