Mathematical Reasoning on AIME 2026 (accuracy)
93.3AIME 2026 AccuracySU-01
Evaluation Results
| Method | Links | |
|---|---|---|
| SU-01Number of runs=82026.05 | 93.3 | |
| Qwen3.6-35B-A3BNumber of runs=82026.05 | 92.9 | |
| Gemma-4-31BNumber of runs=82026.05 | 91.3 | |
| Nemotron-Cascade-2Number of runs=82026.05 | 90 | |
| P1-30B-A3BNumber of runs=82026.05 | 89.6 | |
| GLM-4.7-FlashNumber of runs=82026.05 | 88.3 | |
| RecursiveMASrecursion round=32026.04 | 86.7 | |
| SeirênesBackbone=Qwen3-30B-A3B, same budget=true2026.05 | 79.7 | |
| Single Agenttraining protocol=Full-SFT2026.04 | 76.7 | |
| TextGradrecursion round=32026.04 | 76.7 | |
| FullModel=Qwen3-14B, KV Cache Budget=100%2026.06 | 76.04 | |
| DAPOBackbone=Qwen3-30B-A3B, same budget=true2026.05 | 75.2 | |
| Qwen3-30B-A3BBackbone=Qwen3-30B-A3B2026.05 | 75.1 | |
| Single Agenttraining protocol=LoRA2026.04 | 73.3 | |
| Recursive-TextMASrecursion round=32026.04 | 73.3 | |
| VASE-ATTNVModel=Qwen3-14B, KV Cache Budget=~25%2026.06 | 72.29 | |
| SeirênesBackbone=Qwen3-4B-Instruct, same budget=true2026.05 | 70.8 | |
| VASE-DKVModel=Qwen3-14B, KV Cache Budget=~25%2026.06 | 68.75 | |
| SeerAttention-RModel=Qwen3-14B, KV Cache Budget=~25%2026.06 | 65.62 | |
| Dr.GRPOBackbone=Qwen3-4B-Instruct, same budget=true2026.05 | 64.9 | |
| InTBackbone=Qwen3-4B-Instruct2026.05 | 64.7 | |
| Mixture-of-Agents (MoA)recursion round=32026.04 | 63.3 | |
| LoopLMrecursion round=32026.04 | 63.3 | |
| FullModel=Qwen3-4B, KV Cache Budget=100%2026.06 | 62.71 | |
| DAPOBackbone=Qwen3-4B-Instruct, same budget=true2026.05 | 62.2 | |
| VASE-ATTNVModel=Qwen3-4B, KV Cache Budget=~25%2026.06 | 62.08 | |
| VASE-DKVModel=Qwen3-4B, KV Cache Budget=~25%2026.06 | 61.88 | |
| CurDKVModel=Qwen3-14B, KV Cache Budget=~25%2026.06 | 61.25 | |
| SeerAttention-RModel=Qwen3-4B, KV Cache Budget=~25%2026.06 | 60.42 | |
| R-KVModel=Qwen3-14B, KV Cache Budget=~25%2026.06 | 60.21 | |
| SAGEBackbone=Qwen3-4B-Instruct2026.05 | 57.8 | |
| Qwen3-4B-InstructBackbone=Qwen3-4B-Instruct2026.05 | 56.8 | |
| R-KVModel=Qwen3-4B, KV Cache Budget=~25%2026.06 | 54.58 | |
| CurDKVModel=Qwen3-4B, KV Cache Budget=~25%2026.06 | 54.58 | |
| Scaf-GRPOBackbone=Qwen3-4B-Instruct2026.05 | 54.3 | |
| SnapKVModel=Qwen3-4B, KV Cache Budget=~25%2026.06 | 53.33 | |
| DAPO w/ EDASBackbone=Qwen3-8B2026.05 | 48.65 | |
| RREDCoTGeneration Length=25k tokens, Training Dataset=Numina-CoT2026.06 | 47.5 | |
| CIKAParameter Count=7B, Status=Frozen2026.05 | 46.7 | |
| DelTABackbone=Qwen3-14B-Base, Fine-tuning (FT)=false2026.05 | 45.21 | |
| DAPO w/ EDASBackbone=Qwen3-4B2026.05 | 44.27 | |
| GRPOGeneration Length=25k tokens, Training Dataset=Numina-CoT2026.06 | 44.2 | |
| FIPOBackbone=Qwen3-14B-Base, Fine-tuning (FT)=false2026.05 | 42.5 | |
| DAPOBackbone=Qwen3-14B-Base, Fine-tuning (FT)=true2026.05 | 41.46 | |
| SAPOBackbone=Qwen3-14B-Base, Fine-tuning (FT)=false2026.05 | 41.46 | |
| DAPOBackbone=Qwen3-4B2026.05 | 40.73 | |
| DAPOBackbone=Qwen3-14B-Base, Fine-tuning (FT)=false2026.05 | 39.79 | |
| DAPOBackbone=Qwen3-8B2026.05 | 39.17 | |
| DAPO w/ En AdvBackbone=Qwen3-8B2026.05 | 33.23 | |
| DelTABackbone=Qwen3-8B-Base, Fine-tuning (FT)=false2026.05 | 28.12 | |
| DAPO w/ En AdvBackbone=Qwen3-4B2026.05 | 26.77 | |
| Qwen3-4B Instruct 2507 (starting point)Generation Length=25k tokens, Training Dataset=Numina-CoT2026.06 | 26.7 | |
| DAPOBackbone=Qwen3-8B-Base, Fine-tuning (FT)=true2026.05 | 26.46 | |
| SAPOBackbone=Qwen3-8B-Base, Fine-tuning (FT)=false2026.05 | 26.25 | |
| DAPOBackbone=Qwen3-8B-Base, Fine-tuning (FT)=false2026.05 | 24.17 | |
| FIPOBackbone=Qwen3-8B-Base, Fine-tuning (FT)=false2026.05 | 23.96 | |
| ConSteer-RLBackbone=Qwen3-8B-Base2026.06 | 23.8 | |
| DAPO w/ EDASBackbone=Qwen3-4B-Base2026.05 | 20.97 | |
| GRPOBackbone=Qwen3-8B-Base2026.06 | 20.7 | |
| ConSteer-RLBackbone=Qwen3-4B-Base2026.06 | 20.2 | |
| GRPO w/ EDASBackbone=Qwen3-8B2026.05 | 19.06 | |
| PKPOBackbone=Qwen3-8B2026.05 | 18.85 | |
| GRPOBackbone=Qwen3-8B2026.05 | 18.43 | |
| BaselineBackbone=Qwen3-4B2026.05 | 18.02 | |
| SeirênesBackbone=Qwen2.5-7B-Instruct, same budget=true2026.05 | 17.8 | |
| ConSteer-RLBackbone=Qwen2.5-Math-7B2026.06 | 16.7 | |
| BaselineBackbone=Qwen3-8B2026.05 | 15.1 | |
| GRPOBackbone=Qwen3-4B-Base2026.06 | 14 | |
| DAPO w/ En AdvBackbone=Qwen3-4B-Base2026.05 | 12.71 | |
| LUFFYBackbone=Qwen2.5-7B-Instruct2026.05 | 12.3 | |
| DAPOBackbone=Qwen3-4B-Base2026.05 | 10.73 | |
| BaselineBackbone=Qwen3-4B-Base2026.06 | 10.7 | |
| GRPOBackbone=Qwen2.5-Math-7B2026.06 | 10 | |
| SAGEBackbone=Qwen2.5-7B-Instruct2026.05 | 9.3 | |
| DAPOBackbone=Qwen2.5-7B-Instruct, same budget=true2026.05 | 9.3 | |
| ExGRPOBackbone=Qwen2.5-7B-Instruct2026.05 | 9 | |
| BaselineBackbone=Qwen2.5-Math-7B2026.06 | 6.7 | |
| BaselineBackbone=Qwen3-8B-Base2026.06 | 6.2 | |
| Qwen2.5-7B-InstructBackbone=Qwen2.5-7B-Instruct2026.05 | 5.3 | |
| BaselineBackbone=Qwen3-4B-Base2026.05 | 2.6 |