List your company

Mathematical Reasoning on AIME 25 (Accuracy, Inference Latency (s))

81.7AccuracyQwen3-235B
-0.04421.17842.463.622Jun 2, 2026Jun 6, 2026Jun 10, 2026Jun 15, 2026Jun 19, 2026Jun 23, 2026Jun 28, 2026

Evaluation results

MethodLinks
Qwen3-235BMethod Category=Heuristic, Parameters=235B2026.0681.7—0.6179
ENTROROUTERMethod Category=Routing Architecture2026.0678.8—0.2929
FORCMethod Category=Existing Routing Architecture2026.0675.2—0.2647
RandomMethod Category=Heuristic2026.0673.3—0.2809
Qwen3-30BMethod Category=Heuristic, Parameters=30B2026.0672.1—0.2519
RouteLLMMethod Category=Existing Routing Architecture2026.0667.5—0.0408
Qwen3-4BMethod Category=Heuristic, Parameters=4B2026.0667.1—0.0192
Full AttentionBudget Size=Full2026.0663.317.5288—
HARD-KV (RKV)Budget Size=4096, Variant=+ Ours2026.0656.728.2316—
HARD-KV (SnapKV)Budget Size=4096, Variant=+ Ours2026.0642.325.2286—
RKVBudget Size=4096, Variant=Base2026.064023.9443—
HDPOBackbone=Qwen2.5-Math-7B2026.0631.690.82—
xRouterMethod Category=Existing Routing Architecture2026.0631—0.0033
Router-R1Method Category=Existing Routing Architecture2026.0628.8—0.0116
Critique-GRPOBackbone=Qwen2.5-Math-7B2026.0628.440.88—
SnapKVBudget Size=4096, Variant=Base2026.0627.925.6214—
HARD-KV (RKV)Budget Size=2048, Variant=+ Ours2026.0626.731.0686—
LUFFYBackbone=Qwen2.5-Math-7B2026.0626.610.72—
VanillaBudget Size=4096, Variant=Base2026.0624.831.8514—
GRPOBackbone=Qwen2.5-Math-7B2026.0624.550.64—
HARD-KV (SnapKV)Budget Size=2048, Variant=+ Ours2026.0623.330.1524—
RKVBudget Size=2048, Variant=Base2026.0621.927.7067—
HARD-KV (Vanilla)Budget Size=4096, Variant=+ Ours2026.0620.728.7944—
HARD-KV (Vanilla)Budget Size=1024, Variant=+ Ours2026.0620.528.0862—
EDGE-GRPOBackbone=Qwen2.5-Math-7B2026.0619.370.3—
SnapKVBudget Size=2048, Variant=Base2026.0619.326.2915—
HARD-KV (RKV)Budget Size=1024, Variant=+ Ours2026.0617.832.279—
VanillaBudget Size=2048, Variant=Base2026.0616.726.6175—
SimpleRLBackbone=Qwen2.5-Math-7B2026.0616.340.58—
OatBackbone=Qwen2.5-Math-7B2026.0615.830.39—
HARD-KV (Vanilla)Budget Size=2048, Variant=+ Ours2026.0612.927.9295—
RKVBudget Size=1024, Variant=Base2026.0610.831.8795—
VanillaBudget Size=1024, Variant=Base2026.066.532.3097—
HARD-KV (SnapKV)Budget Size=1024, Variant=+ Ours2026.063.331.7963—
SnapKVBudget Size=1024, Variant=Base2026.063.131.0075—