List your company

Long-context language understanding on LongBench-e

53.04Average ScoreDense
16.099225.689635.2844.8704Feb 11, 2025Apr 30, 2025Jul 17, 2025Oct 3, 2025Dec 20, 2025Mar 8, 2026May 26, 2026

Evaluation results

MethodLinks
DenseModel=Qwen2.5-14B-Instruct-1M, sparse token budget=5122025.0653.0444.3210065.9688.4136.2545.5276.3453.7360.6831.9322.8322.1441.41————————————
Exact2026.025369.9699.67—87.5362.1738.5664—43.7134.16—24.3243.7650.0856.4314.67—————————
HATAModel=Qwen2.5-14B-Instruct-1M, sparse token budget=5122025.0652.944.8699.6765.8788.4937.4145.4176.6753.4560.731.2520.522.0241.46————————————
Exact (Baseline)Model=Llama-3.1-8B-Instruct2025.0250.1749.6298.13—91.8542.7342.3671.67—38.631.31—22.0742.8748.5452.0520.37—————————
16bitModel=Qwen3-8B, Precision=16bit2026.0549.56—95.81————71.33——29.56—21.9610.27—13.1——91.9562.47——————
ExpAttnBackbone=Qwen3-4B, KV eviction ratio (r)=0.252026.0548.92—————————————————————————
Full KVBackbone=Qwen3-4B, KV eviction ratio (r)=02026.0548.77—————————————————————————
OScaRModel=Qwen3-8B, Quantization=INT2, Group Size=322026.0548.74—92.78————72——28.33—21.469.94—12.35——91.4661.57——————
NestedKVBackbone=Qwen3-4B, KV eviction ratio (r)=0.252026.0548.69—————————————————————————
SnapKVBackbone=Qwen3-4B, KV eviction ratio (r)=0.252026.0548.45—————————————————————————
OTTModel=Qwen3-8B, Quantization=INT2, Group Size=322026.0548.21—92.39————71——28.56—21.829.26—11.86——91.3659.41——————
KIVIModel=Qwen3-8B, Quantization=INT2, Group Size=322026.0547.95—91.92————71——28.12—21.589.03—11.86——91.0958.96——————
TurboQuant+Model=Qwen3-8B, Quantization=2.5-bit2026.0547.56—88.92————71.33——28.47—21.668.35—12.53——90.9558.23——————
ExpAttnBackbone=Qwen3-4B, KV eviction ratio (r)=0.502026.0547.53—————————————————————————
PyramidKVBackbone=Qwen3-4B, KV eviction ratio (r)=0.252026.0547.5—————————————————————————
SnapKVBackbone=Qwen3-4B, KV eviction ratio (r)=0.502026.0547.35—————————————————————————
COMPRESSKVretained context tokens=first and last 322026.0247.2564.8598—86.1560.4335.3854.33—33.3730.3—22.2633.2338.1343.4314.33—————————
BalanceKVModel=Llama-3.1-8B-Instruct2025.0246.7748.4596.65—90.8841.437.8869—36.2427.09—20.8435.7537.0446.3720.39—————————
NestedKVBackbone=Qwen3-4B, KV eviction ratio (r)=0.502026.0546.75—————————————————————————
ClusterGenModel=Llama-3.1-8B-Instruct2025.0246.6247.2396.57—90.8239.2639.4967.67—37.2421.16—19.3133.9342.3150.8520.2—————————
SnapKV2026.0245.3669.1689.06—88.3256.3337.1549.33—31.829.7—22.125.5230.1344.3616.67—————————
ExpAttnBackbone=Qwen3-4B, KV eviction ratio (r)=0.852026.0545.11—————————————————————————
KeyDiffBackbone=Qwen3-4B, KV eviction ratio (r)=0.252026.0544.97—————————————————————————
SnapKVBackbone=Qwen3-4B, KV eviction ratio (r)=0.752026.0544.26—————————————————————————
Full Cache (16-bit)Backbone=Llama 3.1 8B Instruct, Precision=16-bit2026.0544.15———————————————————19.5316.5430.2868.4156.4161.6
StreamingLLMBackbone=Qwen3-4B, KV eviction ratio (r)=0.252026.0544.07—————————————————————————
BlockQuantBackbone=Llama 3.1 8B Instruct2026.0544.03———————————————————19.5516.3129.7268.4956.4861.42
EDENBackbone=Llama 3.1 8B Instruct2026.0543.87———————————————————19.0216.329.4268.4656.2961.41
NestedKVBackbone=Qwen3-4B, KV eviction ratio (r)=0.752026.0543.68—————————————————————————
ExpAttnBackbone=Qwen3-4B, KV eviction ratio (r)=0.752026.0543.58—————————————————————————
RabitQBackbone=Llama 3.1 8B Instruct2026.0543.52———————————————————19.1215.8829.468.2355.7560.35
TurboQuantBackbone=Llama 3.1 8B Instruct2026.0543.2———————————————————18.615.7929.1168.1955.3159.72
PyramidKVBackbone=Qwen3-4B, KV eviction ratio (r)=0.502026.0543.07—————————————————————————
RotateKVModel=Qwen3-8B, Quantization=INT2, Group Size=322026.0542.95—91.57————71——14.11—10.596.78—12.74——88.8148——————
Uniformretained context tokens=first and last 322026.0242.8564.8463.44—81.7161.3435.3754.33—25.9330.02—21.8626.9137.5138.4615.33—————————
BalanceKVretained context tokens=first and last 322026.0242.7565.4871.67—73.8362.5734.5655—23.7130.27—21.9829.536.5737.8912.67—————————
KeyDiffBackbone=Llama-3.2-3B-Instruct, KV eviction ratio (r)=0.252026.0542.5—————————————————————————
NestedKVBackbone=Llama-3.2-3B-Instruct, KV eviction ratio (r)=0.252026.0542.28—————————————————————————
OScaRModel=Llama-3.1-8B, Quantization=INT2, Group Size=322026.0541.75—44.46————73.33——29.45—5.6811.61—13.68——90.4165.39——————
16bitModel=Llama-3.1-8B, Precision=16bit2026.0541.7—43.57————72.67——29.88—5.8911.86—13.76——90.1565.78——————
PyramidKV2026.0241.6467.461.06—86.7759.238.2743.33—30.2427.12—21.1321.5929.9639.3815.88—————————
NestedKVBackbone=Qwen3-4B, KV eviction ratio (r)=0.852026.0541.53—————————————————————————
KeyDiffBackbone=Llama-3.2-3B-Instruct, KV eviction ratio (r)=0.502026.0541.53—————————————————————————
PyramidKVBackbone=Qwen3-4B, KV eviction ratio (r)=0.752026.0541.2—————————————————————————
ExpAttnBackbone=Llama-3.2-3B-Instruct, KV eviction ratio (r)=0.502026.0541.15—————————————————————————
NestedKVBackbone=Llama-3.2-3B-Instruct, KV eviction ratio (r)=0.502026.0541.15—————————————————————————
PyramidKVBackbone=Llama-3.2-3B-Instruct, KV eviction ratio (r)=0.252026.0541.1—————————————————————————
KeyDiffBackbone=Llama-3.2-3B-Instruct, KV eviction ratio (r)=0.752026.0540.78—————————————————————————
OTTModel=Llama-3.1-8B, Quantization=INT2, Group Size=322026.0540.74—39.39————73——29.45—4.6510.7—13.27——90.0265.45——————
KeyDiffBackbone=Qwen3-4B, KV eviction ratio (r)=0.502026.0540.38—————————————————————————
QuaRotModel=Qwen3-8B, Quantization=INT2, Group Size=322026.0540.13—91.33————33——20.13—15.017.47—11.7——89.5252.88——————
ExpAttnBackbone=Llama-3.2-3B-Instruct, KV eviction ratio (r)=0.252026.0540.11—————————————————————————
TurboQuant+Model=Llama-3.1-8B, Quantization=2.5-bit2026.0540.03—41.51————72.33——26.6—3.1810.04—13.85——89.9162.82——————
SnapKVBackbone=Qwen3-4B, KV eviction ratio (r)=0.852026.0540.02—————————————————————————
StreamingLLMBackbone=Qwen3-4B, KV eviction ratio (r)=0.502026.0540.01—————————————————————————
KIVIModel=Llama-3.1-8B, Quantization=INT2, Group Size=322026.0539.84—38.95————73——24.41—2.710.71—14.66——89.8364.42——————
Full KVBackbone=Llama-3.2-3B-Instruct, KV eviction ratio (r)=02026.0539.21—————————————————————————
SnapKVBackbone=Llama-3.2-3B-Instruct, KV eviction ratio (r)=0.252026.0538.8—————————————————————————
NestedKVBackbone=Llama-3.2-3B-Instruct, KV eviction ratio (r)=0.752026.0538.57—————————————————————————
RotateKVModel=Llama-3.1-8B, Quantization=INT2, Group Size=322026.0537.98—38.83————72.33——17.82—1.2810.97—13.13——90.5658.95——————
StreamingLLMBackbone=Llama-3.2-3B-Instruct, KV eviction ratio (r)=0.252026.0537.95—————————————————————————
QuaRotModel=Llama-3.1-8B, Quantization=INT2, Group Size=322026.0537.94—42.46————73——18.23—1.759.24—13.73——89.6855.46——————
PyramidKVBackbone=Qwen3-4B, KV eviction ratio (r)=0.852026.0537.76—————————————————————————
PyramidKVBackbone=Llama-3.2-3B-Instruct, KV eviction ratio (r)=0.502026.0537.21—————————————————————————
ExpAttnBackbone=Llama-3.2-3B-Instruct, KV eviction ratio (r)=0.752026.0537.2—————————————————————————
SnapKVBackbone=Llama-3.2-3B-Instruct, KV eviction ratio (r)=0.502026.0537.01—————————————————————————
StreamingLLM2026.0236.468.4824.83—75.1756.2235.7952.33—20.8929.65—22.2523.1725.4926.4512.5—————————
NestedKVBackbone=Llama-3.2-3B-Instruct, KV eviction ratio (r)=0.852026.0536.1—————————————————————————
StreamingLLMBackbone=Qwen3-4B, KV eviction ratio (r)=0.752026.0535.32—————————————————————————
NestedKVBackbone=Qwen3-4B, KV eviction ratio (r)=0.952026.0534.89—————————————————————————
KeyDiffBackbone=Llama-3.2-3B-Instruct, KV eviction ratio (r)=0.852026.0534.42—————————————————————————
SnapKVBackbone=Llama-3.2-3B-Instruct, KV eviction ratio (r)=0.752026.0533.95—————————————————————————
StreamingLLMBackbone=Qwen3-4B, KV eviction ratio (r)=0.852026.0533.57—————————————————————————
StreamingLLMBackbone=Llama-3.2-3B-Instruct, KV eviction ratio (r)=0.502026.0533.56—————————————————————————
PyramidKVBackbone=Llama-3.2-3B-Instruct, KV eviction ratio (r)=0.752026.0533.2—————————————————————————
ExpAttnBackbone=Llama-3.2-3B-Instruct, KV eviction ratio (r)=0.852026.0532.81—————————————————————————
ExpAttnBackbone=Qwen3-4B, KV eviction ratio (r)=0.952026.0531.95—————————————————————————
SnapKVBackbone=Qwen3-4B, KV eviction ratio (r)=0.952026.0531.47—————————————————————————
PyramidKVBackbone=Qwen3-4B, KV eviction ratio (r)=0.952026.0531.39—————————————————————————
PyramidKVBackbone=Llama-3.2-3B-Instruct, KV eviction ratio (r)=0.852026.0530.76—————————————————————————
SnapKVBackbone=Llama-3.2-3B-Instruct, KV eviction ratio (r)=0.852026.0530.24—————————————————————————
KeyDiffBackbone=Qwen3-4B, KV eviction ratio (r)=0.752026.0529.73—————————————————————————
StreamingLLMBackbone=Llama-3.2-3B-Instruct, KV eviction ratio (r)=0.752026.0529.02—————————————————————————
NestedKVBackbone=Llama-3.2-3B-Instruct, KV eviction ratio (r)=0.952026.0528.99—————————————————————————
StreamingLLMBackbone=Qwen3-4B, KV eviction ratio (r)=0.952026.0528.33—————————————————————————
KeyDiffBackbone=Llama-3.2-3B-Instruct, KV eviction ratio (r)=0.952026.0527.17—————————————————————————
StreamingLLMBackbone=Llama-3.2-3B-Instruct, KV eviction ratio (r)=0.852026.0527.14—————————————————————————
ExpAttnBackbone=Llama-3.2-3B-Instruct, KV eviction ratio (r)=0.952026.0524.75—————————————————————————
StreamingLLMBackbone=Llama-3.2-3B-Instruct, KV eviction ratio (r)=0.952026.0524.42—————————————————————————
PyramidKVBackbone=Llama-3.2-3B-Instruct, KV eviction ratio (r)=0.952026.0524.15—————————————————————————
SnapKVBackbone=Llama-3.2-3B-Instruct, KV eviction ratio (r)=0.952026.0524.14—————————————————————————
KeyDiffBackbone=Qwen3-4B, KV eviction ratio (r)=0.852026.0523.47—————————————————————————
KeyDiffBackbone=Qwen3-4B, KV eviction ratio (r)=0.952026.0517.52—————————————————————————
DASHBackbone=Llama-3.1-8B, Token reduction ratio=20%2026.04—————————————————46.11————————
DASHBackbone=Qwen-3-8B, Token reduction ratio=20%2026.04—————————————————54.25————————
DASHBackbone=Qwen-2.5-7B-Instruct-1M, Token reduction ratio=20%2026.04—————————————————46.25————————
LazyLLMBackbone=Llama-3.1-8B, Token reduction ratio=20%2026.04—————————————————43.11————————
LazyLLMBackbone=Qwen-3-8B, Token reduction ratio=20%2026.04—————————————————52.12————————
LazyLLMBackbone=Qwen-2.5-7B-Instruct-1M, Token reduction ratio=20%2026.04—————————————————41.37————————
SlimInferBackbone=Llama-3.1-8B, Token reduction ratio=20%2026.04—————————————————47.03————————