List your company

Long-context Language Understanding on LongBench (Full Metric Set)

58.4Average ScoreOriginal*
17.73628.29338.8549.407Apr 5, 2026Apr 19, 2026May 3, 2026May 18, 2026Jun 1, 2026Jun 15, 2026Jun 30, 2026

Evaluation results

MethodLinks
Original*SWA rho ratio=0.25, Model Size=14B, Evaluation Mode=Zero-shot2026.0458.4——————————————————————
Original*SWA rho ratio=0.25, Model Size=8B, Evaluation Mode=Zero-shot2026.0457.1——————————————————————
BOSCHSWA rho ratio=0.25, Model Size=14B, Evaluation Mode=Zero-shot2026.0456.2——————————————————————
FISHERSWA rho ratio=0.25, Model Size=14B, Evaluation Mode=Zero-shot2026.0455.3——————————————————————
Original*SWA rho ratio=0.25, Model Size=30B, Evaluation Mode=Zero-shot2026.0454.9——————————————————————
FullBackbone=LLaMA-3.1-70B-Instruct2026.0654.62——————————————————————
RaBitQCacheBackbone=LLaMA-3.1-70B-Instruct2026.0654.58——————————————————————
SparQBackbone=LLaMA-3.1-70B-Instruct2026.0653.76——————————————————————
APLSWA rho ratio=0.25, Model Size=14B, Evaluation Mode=Zero-shot2026.0453.2——————————————————————
Full-contextCompression Strategy=Full-context, Prefill (%)=100%, KV (%)=100%2026.0452.7350.16—36.5463.4163.85—29.4156.53—34.8231.8512.53——48.2227.8370.5749.8376.5591.52—100
DCAMSWA rho ratio=0.25, Model Size=14B, Evaluation Mode=Zero-shot2026.0452.5——————————————————————
StructKVCompression Strategy=Prefill-Aware, Prefill (%)=60%, KV (%)=20%2026.0452.4449.93—36.2562.8563.15—29.2456.12—34.4631.7512.43——48.1327.7669.8249.6676.1391.35—100
BOSCH-layerSWA rho ratio=0.25, Model Size=14B, Evaluation Mode=Zero-shot2026.0452.3——————————————————————
RAZORSWA rho ratio=0.25, Model Size=14B, Evaluation Mode=Zero-shot2026.0452.2——————————————————————
BOSCHSWA rho ratio=0.25, Model Size=8B, Evaluation Mode=Zero-shot2026.0452.2——————————————————————
StructKVCompression Strategy=Prefill-Aware, Prefill (%)=60%, KV (%)=10%2026.0451.8749.82—36.1561.5462.76—28.5355.23—33.2631.6212.16——48.1427.2466.8449.5475.8291.26—100
FastKVCompression Strategy=Prefill-Aware, Prefill (%)=60%, KV (%)=20%2026.0451.1248.92—35.9259.2462.25—28.1454.43—32.5531.2211.53——47.8626.8364.1249.1275.5290.84—99.5
APLSWA rho ratio=0.25, Model Size=8B, Evaluation Mode=Zero-shot2026.0451.1——————————————————————
BOSCH-multiSWA rho ratio=0.25, Model Size=14B, Evaluation Mode=Zero-shot2026.0451.1——————————————————————
SnapKVCompression Strategy=Decoding-Only, Prefill (%)=100%, KV (%)=20%2026.045149.63—35.8459.8661.88—28.2354.52—32.4530.9311.85——47.8426.5561.2649.5475.0291.06—99.5
QuestBackbone=LLaMA-3.1-70B-Instruct2026.0650.69——————————————————————
FastKVCompression Strategy=Prefill-Aware, Prefill (%)=60%, KV (%)=10%2026.0450.5348.21—35.5258.5461.53—27.6353.92—31.8630.8411.08——47.5326.2463.2648.5474.0590.26—99.5
PROXYSWA rho ratio=0.25, Model Size=8B, Evaluation Mode=Zero-shot2026.0450.5——————————————————————
SnapKVCompression Strategy=Decoding-Only, Prefill (%)=100%, KV (%)=10%2026.0450.4349.13—35.4259.2561.55—27.9353.86—31.5430.5411.23——47.5225.8660.5849.2573.5590.24—99.5
BOSCH-singleSWA rho ratio=0.25, Model Size=14B, Evaluation Mode=Zero-shot2026.0450.3——————————————————————
PROXYSWA rho ratio=0.25, Model Size=14B, Evaluation Mode=Zero-shot2026.0449.8——————————————————————
StreamingLLMCompression Strategy=Decoding-Only, Prefill (%)=100%, KV (%)=20%2026.0449.4148.85—35.8255.6362.18—27.9152.86—29.5430.1210.25——47.1425.9361.2448.9675.0491.02—88
StreamingLLMCompression Strategy=Decoding-Only, Prefill (%)=100%, KV (%)=10%2026.0448.6848.33—35.1254.2161.25—27.4452.14—28.4528.539.54——46.8225.2660.5648.6374.0590.52—88
RANDSWA rho ratio=0.25, Model Size=14B, Evaluation Mode=Zero-shot2026.0448.2——————————————————————
GemFilterCompression Strategy=Prefill-Aware, Prefill (%)=70%, KV (%)=20%2026.0447.8747.85—34.6353.5260.52—26.8548.94—31.2427.158.24——46.8325.4250.2346.5474.5590.83—92.5
BOSCH-layerSWA rho ratio=0.25, Model Size=8B, Evaluation Mode=Zero-shot2026.0447.7——————————————————————
BOSCHSWA rho ratio=0.5, Model Size=14B, Evaluation Mode=Zero-shot2026.0447——————————————————————
GemFilterCompression Strategy=Prefill-Aware, Prefill (%)=60%, KV (%)=10%2026.0446.8446.94—33.9252.8659.86—26.5447.55—29.5326.247.55——46.1224.6348.5245.8373.0590.22—90
BOSCH-multiSWA rho ratio=0.25, Model Size=8B, Evaluation Mode=Zero-shot2026.0446.7——————————————————————
Diagonal-Tiled Mixed-Precision AttentionBackbone=LLaMA3.1-8B2026.0446.4332.5432.9334.8950.4956.9644.527.0752.6558.8126.8126.76.229685.543.0524.675443.9971.587.8917.92—
BOSCHSWA rho ratio=0.25, Model Size=30B, Evaluation Mode=Zero-shot2026.0446.2——————————————————————
BOSCH-singleSWA rho ratio=0.25, Model Size=8B, Evaluation Mode=Zero-shot2026.0445.5——————————————————————
Original*SWA rho ratio=0.25, Model Size=1.7B, Evaluation Mode=Zero-shot2026.0445.4——————————————————————
PyramidInferCompression Strategy=Prefill-Aware, Prefill (%)=60%, KV (%)=60%2026.0444.7844.53—29.5445.1660.84—21.8643.25—24.2527.545.83——39.8223.4261.5247.5670.0589.24—82
FISHERSWA rho ratio=0.25, Model Size=30B, Evaluation Mode=Zero-shot2026.0444.2——————————————————————
Native (SDPA)Backbone=LLaMA3.1-8B2026.0444.1139.1528.433.8749.8949.4641.526.1652.2657.5725.8225.284.029992.1442.6423.4742.2336.66577.0114.8—
QADASWA rho ratio=0.25, Model Size=14B, Evaluation Mode=Zero-shot2026.0443.9——————————————————————
INTRSWA rho ratio=0.25, Model Size=14B, Evaluation Mode=Zero-shot2026.0443.4——————————————————————
INTRSWA rho ratio=0.25, Model Size=8B, Evaluation Mode=Zero-shot2026.0443.3——————————————————————
DCAMSWA rho ratio=0.25, Model Size=8B, Evaluation Mode=Zero-shot2026.0443——————————————————————
INTRSWA rho ratio=0.25, Model Size=30B, Evaluation Mode=Zero-shot2026.0442.3——————————————————————
APLSWA rho ratio=0.25, Model Size=30B, Evaluation Mode=Zero-shot2026.0441.8——————————————————————
BOSCHSWA rho ratio=0.5, Model Size=8B, Evaluation Mode=Zero-shot2026.0441.8——————————————————————
RAZORSWA rho ratio=0.25, Model Size=8B, Evaluation Mode=Zero-shot2026.0441.2——————————————————————
FISHERSWA rho ratio=0.25, Model Size=8B, Evaluation Mode=Zero-shot2026.0440.9——————————————————————
DCAMSWA rho ratio=0.25, Model Size=30B, Evaluation Mode=Zero-shot2026.0440.8——————————————————————
BOSCH-multiSWA rho ratio=0.25, Model Size=30B, Evaluation Mode=Zero-shot2026.0440.8——————————————————————
H2OCompression Strategy=Decoding-Only, Prefill (%)=100%, KV (%)=20%2026.0440.7323.83—34.9224.1262.54—27.5632.85—18.5426.849.06——19.5624.2358.2647.1272.0492.15—78
RANDSWA rho ratio=0.25, Model Size=8B, Evaluation Mode=Zero-shot2026.0440.2——————————————————————
RANDSWA rho ratio=0.25, Model Size=30B, Evaluation Mode=Zero-shot2026.0440.2——————————————————————
QADASWA rho ratio=0.25, Model Size=8B, Evaluation Mode=Zero-shot2026.0440.1——————————————————————
BOSCH-layerSWA rho ratio=0.25, Model Size=30B, Evaluation Mode=Zero-shot2026.0439.9——————————————————————
H2OCompression Strategy=Decoding-Only, Prefill (%)=100%, KV (%)=10%2026.0439.1119.54—35.2220.8759.83—28.1430.55—13.2625.428.54——16.2325.6557.5646.5371.5591.82—75
PROXYSWA rho ratio=0.25, Model Size=30B, Evaluation Mode=Zero-shot2026.0438.4——————————————————————
BOSCHSWA rho ratio=0.25, Model Size=1.7B, Evaluation Mode=Zero-shot2026.0438——————————————————————
BOSCHSWA rho ratio=0.75, Model Size=14B, Evaluation Mode=Zero-shot2026.0438——————————————————————
Diagonal-Tiled Mixed-Precision AttentionBackbone=LLaMA3.2-3B2026.0437.236.3333.0433.0448.5945.5430.7521.7847.8753.6225.7725.441.53710.539.6223.6350.1542.7469.588.1316.69—
BOSCH-layerSWA rho ratio=0.25, Model Size=1.7B, Evaluation Mode=Zero-shot2026.0437.1——————————————————————
APLSWA rho ratio=0.25, Model Size=1.7B, Evaluation Mode=Zero-shot2026.0436.8——————————————————————
QADASWA rho ratio=0.25, Model Size=1.7B, Evaluation Mode=Zero-shot2026.0436.4——————————————————————
BOSCH-singleSWA rho ratio=0.25, Model Size=1.7B, Evaluation Mode=Zero-shot2026.0436.3——————————————————————
FISHERSWA rho ratio=0.25, Model Size=1.7B, Evaluation Mode=Zero-shot2026.0436.1——————————————————————
BOSCHSWA rho ratio=0.875, Model Size=14B, Evaluation Mode=Zero-shot2026.0436.1——————————————————————
INTRSWA rho ratio=0.25, Model Size=1.7B, Evaluation Mode=Zero-shot2026.0436——————————————————————
BOSCH-multiSWA rho ratio=0.25, Model Size=1.7B, Evaluation Mode=Zero-shot2026.0436——————————————————————
BOSCHSWA rho ratio=0.5, Model Size=30B, Evaluation Mode=Zero-shot2026.0436——————————————————————
Native (SDPA)Backbone=LLaMA3.2-3B2026.0435.8430.7528.1132.2646.1842.9826.7525.5646.8448.0620.3519.763.4808.2532.1322.5244.1434.7962.583.6813.65—
QADASWA rho ratio=0.25, Model Size=30B, Evaluation Mode=Zero-shot2026.0435.1——————————————————————
RAZORSWA rho ratio=0.25, Model Size=1.7B, Evaluation Mode=Zero-shot2026.0435.1——————————————————————
RAZORSWA rho ratio=0.25, Model Size=30B, Evaluation Mode=Zero-shot2026.0434.5——————————————————————
PROXYSWA rho ratio=0.25, Model Size=1.7B, Evaluation Mode=Zero-shot2026.0434.4——————————————————————
BOSCH-singleSWA rho ratio=0.25, Model Size=30B, Evaluation Mode=Zero-shot2026.0434——————————————————————
RANDSWA rho ratio=0.25, Model Size=1.7B, Evaluation Mode=Zero-shot2026.0433.7——————————————————————
BMESWA rho ratio=0.25, Model Size=14B, Evaluation Mode=Zero-shot2026.0432.6——————————————————————
DCAMSWA rho ratio=0.25, Model Size=1.7B, Evaluation Mode=Zero-shot2026.0432.5——————————————————————
BOSCHSWA rho ratio=0.5, Model Size=1.7B, Evaluation Mode=Zero-shot2026.0432.1——————————————————————
BOSCHSWA rho ratio=0.75, Model Size=8B, Evaluation Mode=Zero-shot2026.0431.6——————————————————————
BMESWA rho ratio=0.25, Model Size=1.7B, Evaluation Mode=Zero-shot2026.0431.5——————————————————————
BMESWA rho ratio=0.25, Model Size=8B, Evaluation Mode=Zero-shot2026.0431——————————————————————
BOSCHSWA rho ratio=0.875, Model Size=8B, Evaluation Mode=Zero-shot2026.0428.6——————————————————————
BMESWA rho ratio=0.25, Model Size=30B, Evaluation Mode=Zero-shot2026.0428.2——————————————————————
BOSCHSWA rho ratio=0.75, Model Size=1.7B, Evaluation Mode=Zero-shot2026.0426——————————————————————
BOSCHSWA rho ratio=0.75, Model Size=30B, Evaluation Mode=Zero-shot2026.0424.8——————————————————————
BOSCHSWA rho ratio=0.875, Model Size=1.7B, Evaluation Mode=Zero-shot2026.0423.6——————————————————————
BOSCHSWA rho ratio=0.875, Model Size=30B, Evaluation Mode=Zero-shot2026.0419.3——————————————————————