List your company

Long-context Understanding on RULER 4k (test)

95.7RULER 4k ScoreExpectedAttention
9.932.17554.4576.725May 25, 2026

Evaluation results

MethodLinks
ExpectedAttentionModel=Llama3.1-8B, Compression Ratio=10%2026.0595.7
IndexMemModel=Llama3.1-8B, Compression Ratio=10%2026.0595.6
IndexMemModel=Llama3.1-8B, Compression Ratio=25%2026.0595.6
SnapKVModel=Llama3.1-8B, Compression Ratio=10%2026.0595.5
IndexMemModel=Qwen3-8B, Compression Ratio=0%2026.0595.3
KeyDiffModel=Qwen3-8B, Compression Ratio=0%2026.0595.3
TOVAModel=Qwen3-8B, Compression Ratio=0%2026.0595.3
SnapKVModel=Qwen3-8B, Compression Ratio=0%2026.0595.3
PyramidKVModel=Qwen3-8B, Compression Ratio=0%2026.0595.3
IndexMemModel=Llama3.1-8B, Compression Ratio=0%2026.0595.3
ExpectedAttentionModel=Llama3.1-8B, Compression Ratio=0%2026.0595.3
ExpectedAttentionModel=Llama3.1-8B, Compression Ratio=25%2026.0595.3
TOVAModel=Llama3.1-8B, Compression Ratio=0%2026.0595.3
SnapKVModel=Llama3.1-8B, Compression Ratio=0%2026.0595.3
PyramidKVModel=Llama3.1-8B, Compression Ratio=0%2026.0595.3
IndexMemModel=Qwen3-8B, Compression Ratio=10%2026.0593.8
KeyDiffModel=Qwen3-8B, Compression Ratio=10%2026.0593.8
IndexMemModel=Llama3.1-8B, Compression Ratio=50%2026.0593.3
TOVAModel=Llama3.1-8B, Compression Ratio=10%2026.0593.2
SnapKVModel=Qwen3-8B, Compression Ratio=10%2026.0592.6
ExpectedAttentionModel=Llama3.1-8B, Compression Ratio=50%2026.0592.2
IndexMemModel=Mistral-7B, Compression Ratio=0%2026.0592.1
ExpectedAttentionModel=Mistral-7B, Compression Ratio=0%2026.0592.1
TOVAModel=Mistral-7B, Compression Ratio=0%2026.0592.1
SnapKVModel=Mistral-7B, Compression Ratio=0%2026.0592.1
PyramidKVModel=Mistral-7B, Compression Ratio=0%2026.0592.1
PyramidKVModel=Qwen3-8B, Compression Ratio=10%2026.0591.4
IndexMemModel=Qwen3-8B, Compression Ratio=25%2026.0590.3
KeyDiffModel=Qwen3-8B, Compression Ratio=25%2026.0589.4
TOVAModel=Qwen3-8B, Compression Ratio=10%2026.0589
SnapKVModel=Llama3.1-8B, Compression Ratio=25%2026.0588.8
IndexMemModel=Mistral-7B, Compression Ratio=10%2026.0588.6
TOVAModel=Llama3.1-8B, Compression Ratio=25%2026.0587.3
IndexMemModel=Mistral-7B, Compression Ratio=25%2026.0585
PyramidKVModel=Llama3.1-8B, Compression Ratio=10%2026.0584.5
SnapKVModel=Qwen3-8B, Compression Ratio=25%2026.0584
TOVAModel=Mistral-7B, Compression Ratio=10%2026.0583.1
IndexMemModel=Mistral-7B, Compression Ratio=50%2026.0582.8
TOVAModel=Qwen3-8B, Compression Ratio=25%2026.0582.5
SnapKVModel=Llama3.1-8B, Compression Ratio=50%2026.0581.8
PyramidKVModel=Llama3.1-8B, Compression Ratio=25%2026.0579.4
IndexMemModel=Qwen3-8B, Compression Ratio=50%2026.0578.9
KeyDiffModel=Qwen3-8B, Compression Ratio=50%2026.0578.6
TOVAModel=Mistral-7B, Compression Ratio=25%2026.0578.4
IndexMemModel=Llama3.1-8B, Compression Ratio=75%2026.0578.3
TOVAModel=Qwen3-8B, Compression Ratio=50%2026.0577.6
ExpectedAttentionModel=Mistral-7B, Compression Ratio=10%2026.0577.1
TOVAModel=Llama3.1-8B, Compression Ratio=50%2026.0576.2
ExpectedAttentionModel=Llama3.1-8B, Compression Ratio=75%2026.0575.9
TOVAModel=Mistral-7B, Compression Ratio=50%2026.0575.5
PyramidKVModel=Qwen3-8B, Compression Ratio=25%2026.0575.1
IndexMemModel=Qwen3-8B, Compression Ratio=75%2026.0572.9
SnapKVModel=Mistral-7B, Compression Ratio=10%2026.0571.5
PyramidKVModel=Mistral-7B, Compression Ratio=10%2026.0570.8
IndexMemModel=Mistral-7B, Compression Ratio=75%2026.0568.8
ExpectedAttentionModel=Mistral-7B, Compression Ratio=25%2026.0566.3
PyramidKVModel=Llama3.1-8B, Compression Ratio=50%2026.0565.8
KeyDiffModel=Qwen3-8B, Compression Ratio=75%2026.0564.4
TOVAModel=Llama3.1-8B, Compression Ratio=75%2026.0563.3
SnapKVModel=Llama3.1-8B, Compression Ratio=75%2026.0563.2
TOVAModel=Qwen3-8B, Compression Ratio=75%2026.0562.4
TOVAModel=Mistral-7B, Compression Ratio=75%2026.0562
SnapKVModel=Mistral-7B, Compression Ratio=25%2026.0557.6
PyramidKVModel=Mistral-7B, Compression Ratio=25%2026.0557.1
SnapKVModel=Qwen3-8B, Compression Ratio=50%2026.0555.7
IndexMemModel=Llama3.1-8B, Compression Ratio=90%2026.0555.2
IndexMemModel=Qwen3-8B, Compression Ratio=90%2026.0546.2
ExpectedAttentionModel=Mistral-7B, Compression Ratio=50%2026.0544.8
PyramidKVModel=Llama3.1-8B, Compression Ratio=75%2026.0543.6
SnapKVModel=Llama3.1-8B, Compression Ratio=90%2026.0543.4
IndexMemModel=Mistral-7B, Compression Ratio=90%2026.0541.5
SnapKVModel=Mistral-7B, Compression Ratio=50%2026.0541.3
PyramidKVModel=Qwen3-8B, Compression Ratio=50%2026.0540.8
PyramidKVModel=Mistral-7B, Compression Ratio=50%2026.0538.6
KeyDiffModel=Qwen3-8B, Compression Ratio=90%2026.0537.9
TOVAModel=Llama3.1-8B, Compression Ratio=90%2026.0537.5
SnapKVModel=Qwen3-8B, Compression Ratio=75%2026.0533.1
SnapKVModel=Mistral-7B, Compression Ratio=75%2026.0531.9
ExpectedAttentionModel=Llama3.1-8B, Compression Ratio=90%2026.0530.6
PyramidKVModel=Qwen3-8B, Compression Ratio=75%2026.0529.7
PyramidKVModel=Mistral-7B, Compression Ratio=75%2026.0529.6
TOVAModel=Mistral-7B, Compression Ratio=90%2026.0529.2
ExpectedAttentionModel=Mistral-7B, Compression Ratio=75%2026.0528.1
TOVAModel=Qwen3-8B, Compression Ratio=90%2026.0524.7
PyramidKVModel=Llama3.1-8B, Compression Ratio=90%2026.0523.1
PyramidKVModel=Mistral-7B, Compression Ratio=90%2026.0519.9
SnapKVModel=Mistral-7B, Compression Ratio=90%2026.0519.8
PyramidKVModel=Qwen3-8B, Compression Ratio=90%2026.0519.5
SnapKVModel=Qwen3-8B, Compression Ratio=90%2026.0519.2
ExpectedAttentionModel=Mistral-7B, Compression Ratio=90%2026.0513.2