Mathematical Reasoning on AIME 25 (Accuracy, Inference Latency (s))
81.7AccuracyQwen3-235B
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| Qwen3-235BMethod Category=Heuristic, Parameters=235B2026.06 | 81.7 | — | 0.6179 | |
| ENTROROUTERMethod Category=Routing Architecture2026.06 | 78.8 | — | 0.2929 | |
| FORCMethod Category=Existing Routing Architecture2026.06 | 75.2 | — | 0.2647 | |
| RandomMethod Category=Heuristic2026.06 | 73.3 | — | 0.2809 | |
| Qwen3-30BMethod Category=Heuristic, Parameters=30B2026.06 | 72.1 | — | 0.2519 | |
| RouteLLMMethod Category=Existing Routing Architecture2026.06 | 67.5 | — | 0.0408 | |
| Qwen3-4BMethod Category=Heuristic, Parameters=4B2026.06 | 67.1 | — | 0.0192 | |
| Full AttentionBudget Size=Full2026.06 | 63.3 | 17.5288 | — | |
| HARD-KV (RKV)Budget Size=4096, Variant=+ Ours2026.06 | 56.7 | 28.2316 | — | |
| HARD-KV (SnapKV)Budget Size=4096, Variant=+ Ours2026.06 | 42.3 | 25.2286 | — | |
| RKVBudget Size=4096, Variant=Base2026.06 | 40 | 23.9443 | — | |
| HDPOBackbone=Qwen2.5-Math-7B2026.06 | 31.69 | 0.82 | — | |
| xRouterMethod Category=Existing Routing Architecture2026.06 | 31 | — | 0.0033 | |
| Router-R1Method Category=Existing Routing Architecture2026.06 | 28.8 | — | 0.0116 | |
| Critique-GRPOBackbone=Qwen2.5-Math-7B2026.06 | 28.44 | 0.88 | — | |
| SnapKVBudget Size=4096, Variant=Base2026.06 | 27.9 | 25.6214 | — | |
| HARD-KV (RKV)Budget Size=2048, Variant=+ Ours2026.06 | 26.7 | 31.0686 | — | |
| LUFFYBackbone=Qwen2.5-Math-7B2026.06 | 26.61 | 0.72 | — | |
| VanillaBudget Size=4096, Variant=Base2026.06 | 24.8 | 31.8514 | — | |
| GRPOBackbone=Qwen2.5-Math-7B2026.06 | 24.55 | 0.64 | — | |
| HARD-KV (SnapKV)Budget Size=2048, Variant=+ Ours2026.06 | 23.3 | 30.1524 | — | |
| RKVBudget Size=2048, Variant=Base2026.06 | 21.9 | 27.7067 | — | |
| HARD-KV (Vanilla)Budget Size=4096, Variant=+ Ours2026.06 | 20.7 | 28.7944 | — | |
| HARD-KV (Vanilla)Budget Size=1024, Variant=+ Ours2026.06 | 20.5 | 28.0862 | — | |
| EDGE-GRPOBackbone=Qwen2.5-Math-7B2026.06 | 19.37 | 0.3 | — | |
| SnapKVBudget Size=2048, Variant=Base2026.06 | 19.3 | 26.2915 | — | |
| HARD-KV (RKV)Budget Size=1024, Variant=+ Ours2026.06 | 17.8 | 32.279 | — | |
| VanillaBudget Size=2048, Variant=Base2026.06 | 16.7 | 26.6175 | — | |
| SimpleRLBackbone=Qwen2.5-Math-7B2026.06 | 16.34 | 0.58 | — | |
| OatBackbone=Qwen2.5-Math-7B2026.06 | 15.83 | 0.39 | — | |
| HARD-KV (Vanilla)Budget Size=2048, Variant=+ Ours2026.06 | 12.9 | 27.9295 | — | |
| RKVBudget Size=1024, Variant=Base2026.06 | 10.8 | 31.8795 | — | |
| VanillaBudget Size=1024, Variant=Base2026.06 | 6.5 | 32.3097 | — | |
| HARD-KV (SnapKV)Budget Size=1024, Variant=+ Ours2026.06 | 3.3 | 31.7963 | — | |
| SnapKVBudget Size=1024, Variant=Base2026.06 | 3.1 | 31.0075 | — |