List your company

Mathematical Reasoning on UMath

56AccuracyFull Attention
7.95220.42632.945.374Jun 27, 2026

Evaluation results

MethodLinks
Full AttentionSelection Method=Full Attn, Budget=None2026.06569,169
SnapKV + HARD-KVSelection Method=SnapKV, Config=+ Ours, Budget=40962026.065621,778
Full AttentionBackbone=Qwen3-4B, Constraint=Full (No Budget)2026.0653.310,111.5
Vanilla + HARD-KVSelection Method=Vanilla, Config=+ Ours, Budget=40962026.065214,711
RKV + HARD-KVSelection Method=RKV, Config=+ Ours, Budget=40962026.065210,123
HARD-KV (RKV)Budget Size=2048, Variant=+ Ours2026.065224,748.5
RKVBackbone=Qwen3-4B, Selection Method=RKV, Top-k Budget=40962026.065212,825.3
Full AttentionBudget Size=Full2026.0650.49,815.7
HARD-KV (RKV)Budget Size=4096, Variant=+ Ours2026.0650.321,296.5
HARD-KVBackbone=Qwen3-4B, Selection Method=RKV, Top-k Budget=4096, Top-p Budget=0.902026.0650.323,156.9
SnapKVSelection Method=SnapKV, Config=Base, Budget=40962026.065021,097
RKVSelection Method=RKV, Config=Base, Budget=40962026.065010,537
HARD-KV (SnapKV)Budget Size=4096, Variant=+ Ours2026.0649.414,520.4
HARD-KVBackbone=Qwen3-4B, Selection Method=SnapKV, Top-k Budget=4096, Top-p Budget=0.902026.0649.416,209.9
VanillaSelection Method=Vanilla, Config=Base, Budget=40962026.064822,183
SnapKVBudget Size=4096, Variant=Base2026.0646.712,420.2
HARD-KV (SnapKV)Budget Size=2048, Variant=+ Ours2026.0646.222,499.4
HARD-KV (Vanilla)Budget Size=1024, Variant=+ Ours2026.0646.115,363.5
SnapKVBackbone=Qwen3-4B, Selection Method=SnapKV, Top-k Budget=40962026.0644.214,371.1
VanillaBackbone=Qwen3-4B, Selection Method=Vanilla, Top-k Budget=40962026.064412,828
HARD-KVBackbone=Qwen3-4B, Selection Method=SnapKV, Top-k Budget=2048, Top-p Budget=0.902026.064223,663.2
HARD-KVBackbone=Qwen3-4B, Selection Method=RKV, Top-k Budget=2048, Top-p Budget=0.902026.064223,663.2
SnapKVBudget Size=2048, Variant=Base2026.0641.819,021.4
RKVBudget Size=4096, Variant=Base2026.0639.927,429.7
VanillaBudget Size=2048, Variant=Base2026.063816,102.5
HARD-KVBackbone=Qwen3-4B, Selection Method=Vanilla, Top-k Budget=4096, Top-p Budget=0.902026.063824,832.7
HARD-KVBackbone=Qwen3-4B, Selection Method=Vanilla, Top-k Budget=1024, Top-p Budget=0.902026.0636.723,695.8
VanillaBackbone=Qwen3-4B, Selection Method=Vanilla, Top-k Budget=20482026.0636.719,021.4
HARD-KV (Vanilla)Budget Size=2048, Variant=+ Ours2026.0636.216,791.5
HARD-KVBackbone=Qwen3-4B, Selection Method=Vanilla, Top-k Budget=2048, Top-p Budget=0.902026.0636.224,691.8
HARD-KV (Vanilla)Budget Size=4096, Variant=+ Ours2026.0636.116,118.8
RKVBudget Size=2048, Variant=Base2026.0635.117,915.2
RKVBackbone=Qwen3-4B, Selection Method=RKV, Top-k Budget=20482026.0635.117,685.8
VanillaBudget Size=4096, Variant=Base2026.0632.126,314.3
SnapKVBackbone=Qwen3-4B, Selection Method=SnapKV, Top-k Budget=20482026.0631.820,651.4
RKVBackbone=Qwen3-4B, Selection Method=RKV, Top-k Budget=10242026.0630.118,504.6
RKVBudget Size=1024, Variant=Base2026.0628.119,625.3
HARD-KV (RKV)Budget Size=1024, Variant=+ Ours2026.0624.124,346.1
HARD-KVBackbone=Qwen3-4B, Selection Method=SnapKV, Top-k Budget=1024, Top-p Budget=0.902026.0622.529,204.8
HARD-KVBackbone=Qwen3-4B, Selection Method=RKV, Top-k Budget=1024, Top-p Budget=0.902026.0620.131,351.6
SnapKVBudget Size=1024, Variant=Base2026.061822,350.7
SnapKVBackbone=Qwen3-4B, Selection Method=SnapKV, Top-k Budget=10242026.061829,332.5
HARD-KV (SnapKV)Budget Size=1024, Variant=+ Ours2026.0615.527,705.7
VanillaBackbone=Qwen3-4B, Selection Method=Vanilla, Top-k Budget=10242026.061428,859.6
VanillaBudget Size=1024, Variant=Base2026.069.832,768