List your company

Mathematical Reasoning on AIME 24 (Accuracy and Token Count)

79.9AccuracyFull Attention
3.25223.15143.0562.949Jun 2, 2026Jun 6, 2026Jun 10, 2026Jun 14, 2026Jun 18, 2026Jun 22, 2026Jun 27, 2026

Evaluation results

MethodLinks
Full AttentionBudget Size=Full2026.0679.913,735.7
FP16Backbone=Phi-4-14B, K/V=16/16, bits/elem=16.0, UP (Uniform Precision)=true2026.0662.211
KVarNBackbone=Phi-4-14B, K/V=2/2, bits/elem=2.3, UP (Uniform Precision)=true2026.0661.711
FP16Backbone=Qwen3-4B, K/V=16/16, bits/elem=16.0, UP (Uniform Precision)=true2026.0661.112
KVarNBackbone=Qwen3-4B, K/V=2/2, bits/elem=2.3, UP (Uniform Precision)=true2026.066012
PolarQuantBackbone=Phi-4-14B, K/V=4/2, bits/elem=3.3, UP (Uniform Precision)=true2026.066012
QuaRotBackbone=Phi-4-14B, K/V=2/2, bits/elem=2.3, UP (Uniform Precision)=true2026.0658.912
HARD-KV (RKV)Budget Size=4096, Variant=+ Ours2026.0657.928,094
KIVIBackbone=Phi-4-14B, K/V=2/2, bits/elem=2.3, UP (Uniform Precision)=true2026.0657.812
SnapKVBudget Size=4096, Variant=Base2026.0657.719,157.4
QuaRotBackbone=Qwen3-4B, K/V=2/2, bits/elem=2.3, UP (Uniform Precision)=true2026.0656.712
HARD-KV (SnapKV)Budget Size=4096, Variant=+ Ours2026.0656.721,616.5
KVQuant-1%Backbone=Phi-4-14B, K/V=2/2, bits/elem=2.4, UP (Uniform Precision)=false2026.0655.613
KittyBackbone=Phi-4-14B, K/V=2/2, bits/elem=2.4, UP (Uniform Precision)=false2026.0655.612
KIVIBackbone=Qwen3-4B, K/V=2/2, bits/elem=2.3, UP (Uniform Precision)=true2026.0655.512
KittyBackbone=Qwen3-4B, K/V=2/2, bits/elem=2.4, UP (Uniform Precision)=false2026.0653.313
HARD-KV (RKV)Budget Size=2048, Variant=+ Ours2026.0653.327,819.6
PolarQuantBackbone=Qwen3-4B, K/V=4/2, bits/elem=3.3, UP (Uniform Precision)=true2026.0652.213
TurboQuantBackbone=Phi-4-14B, K/V=3/3, bits/elem=4.5, UP (Uniform Precision)=true, Implementation=vLLM2026.0652.213
TurboQuantBackbone=Qwen3-4B, K/V=3/3, bits/elem=4.6, UP (Uniform Precision)=true, Implementation=vLLM2026.0648.913
RKVBudget Size=4096, Variant=Base2026.0643.323,281.5
HARD-KV (SnapKV)Budget Size=2048, Variant=+ Ours2026.0640.228,471.4
KVQuant-1%Backbone=Qwen3-4B, K/V=2/2, bits/elem=2.4, UP (Uniform Precision)=false2026.064014
HARD-KV (Vanilla)Budget Size=2048, Variant=+ Ours2026.0637.324,158.1
HARD-KV (Vanilla)Budget Size=4096, Variant=+ Ours2026.0633.523,449.8
VanillaBudget Size=2048, Variant=Base2026.0633.123,530.6
VanillaBudget Size=4096, Variant=Base2026.0629.731,843.5
HARD-KV (Vanilla)Budget Size=1024, Variant=+ Ours2026.062424,158.1
HARD-KV (RKV)Budget Size=1024, Variant=+ Ours2026.0623.430,480.8
RKVBudget Size=2048, Variant=Base2026.0623.327,968
SnapKVBudget Size=2048, Variant=Base2026.0621.323,860.3
HARD-KV (SnapKV)Budget Size=1024, Variant=+ Ours2026.0613.332,768
RKVBudget Size=1024, Variant=Base2026.0613.330,423.2
VanillaBudget Size=1024, Variant=Base2026.061032,161
SnapKVBudget Size=1024, Variant=Base2026.066.231,323.1