List your company

Mathematical Reasoning on AIME25 (Acc, Gen. Len)

63.3AccuracyFull Attention
4.22819.56434.950.236Jun 27, 2026Jun 28, 2026Jun 29, 2026Jun 30, 2026Jul 1, 2026

Evaluation results

MethodLinks
Full AttentionBackbone=Qwen3-4B, Constraint=Full (No Budget)2026.0663.317,998.7
HARD-KVBackbone=Qwen3-4B, Selection Method=RKV, Top-k Budget=4096, Top-p Budget=0.902026.0653.428,202.8
DASH2026.0750.8—
DR-GRPO2026.0750.3—
Nemotron-4BParameters=4B2026.0746.1—
GRPO + Brevity Bonus2026.0745.9—
GRPO2026.0745.4—
HARD-KVBackbone=Qwen3-4B, Selection Method=SnapKV, Top-k Budget=4096, Top-p Budget=0.902026.0644.326,548.9
SnapKVBackbone=Qwen3-4B, Selection Method=SnapKV, Top-k Budget=40962026.0643.323,654.5
RKVBackbone=Qwen3-4B, Selection Method=RKV, Top-k Budget=40962026.0630.720,616.6
HARD-KVBackbone=Qwen3-4B, Selection Method=SnapKV, Top-k Budget=2048, Top-p Budget=0.902026.0626.728,254.2
VanillaBackbone=Qwen3-4B, Selection Method=Vanilla, Top-k Budget=40962026.0624.831,851.4
RKVBackbone=Qwen3-4B, Selection Method=RKV, Top-k Budget=20482026.0621.926,836.2
HARD-KVBackbone=Qwen3-4B, Selection Method=Vanilla, Top-k Budget=4096, Top-p Budget=0.902026.0620.729,966.5
HARD-KVBackbone=Qwen3-4B, Selection Method=RKV, Top-k Budget=2048, Top-p Budget=0.902026.0620.731,809.6
VanillaBackbone=Qwen3-4B, Selection Method=Vanilla, Top-k Budget=20482026.0616.726,617.5
RKVBackbone=Qwen3-4B, Selection Method=RKV, Top-k Budget=10242026.0613.727,924.7
HARD-KVBackbone=Qwen3-4B, Selection Method=RKV, Top-k Budget=1024, Top-p Budget=0.902026.0613.231,496.4
HARD-KVBackbone=Qwen3-4B, Selection Method=Vanilla, Top-k Budget=2048, Top-p Budget=0.902026.0612.930,837.8
SnapKVBackbone=Qwen3-4B, Selection Method=SnapKV, Top-k Budget=20482026.0612.330,040.4
SnapKVBackbone=Qwen3-4B, Selection Method=SnapKV, Top-k Budget=10242026.0610.932,229.7
HARD-KVBackbone=Qwen3-4B, Selection Method=Vanilla, Top-k Budget=1024, Top-p Budget=0.902026.0610.530,906.7
HARD-KVBackbone=Qwen3-4B, Selection Method=SnapKV, Top-k Budget=1024, Top-p Budget=0.902026.069.332,768
VanillaBackbone=Qwen3-4B, Selection Method=Vanilla, Top-k Budget=10242026.066.532,309.7