Mathematical Reasoning on AIME25 (Accuracy, Token Count)
87.4AccuracyDeepConf
Evaluation results
| Method | Links | ||
|---|---|---|---|
| DeepConfModel=DeepSeek-8B, Compute Budget=@5122026.02 | 87.4 | 23.7 | |
| CoRefine TreeModel=DeepSeek-8B, Compute Budget=@5122026.02 | 87.3 | 0.59 | |
| CoRefineModel=DeepSeek-8B, Compute Budget=@5122026.02 | 86.7 | 0.39 | |
| CoRefine TreeModel=Qwen3-32B, Compute Budget=@5122026.02 | 83.3 | 0.3 | |
| CoRefineModel=Qwen3-32B, Compute Budget=@5122026.02 | 82.7 | 0.19 | |
| MajorityModel=DeepSeek-8B, Compute Budget=@5122026.02 | 82 | 40.1 | |
| DeepConfModel=Qwen3-32B, Compute Budget=@5122026.02 | 81.3 | 1.61 | |
| MajorityModel=Qwen3-32B, Compute Budget=@5122026.02 | 80.7 | 24.3 | |
| Parallel-ProbeBase Model=Qwen3-8B, Type=Handcrafted2026.05 | 76.9 | 0.0847 | |
| SC @ 64Base Model=Qwen3-8B, Type=Handcrafted2026.05 | 76.7 | 0.1124 | |
| ASCBase Model=Qwen3-8B, Type=Handcrafted2026.05 | 76.7 | 0.0406 | |
| ESCBase Model=Qwen3-8B, Type=Handcrafted2026.05 | 76.7 | 0.0793 | |
| SC @ 64Base Model=Qwen3-4B, Type=Handcrafted2026.05 | 76.6 | 0.1088 | |
| ESCBase Model=Qwen3-4B, Type=Handcrafted2026.05 | 76.6 | 0.0793 | |
| ASCBase Model=Qwen3-4B, Type=Handcrafted2026.05 | 76.4 | 0.0277 | |
| Parallel-ProbeBase Model=Qwen3-4B, Type=Handcrafted2026.05 | 76.1 | 0.0806 | |
| AutoTTS (β=1.0)Base Model=Qwen3-8B, Type=Discovered2026.05 | 75.8 | 0.0672 | |
| AutoTTS (β=1.0)Base Model=Qwen3-4B, Type=Discovered2026.05 | 74.4 | 0.061 | |
| AutoTTS (β=0.5)Base Model=Qwen3-8B, Type=Discovered2026.05 | 74.1 | 0.0361 | |
| AutoTTS (β=0.5)Base Model=Qwen3-4B, Type=Discovered2026.05 | 73.8 | 0.0332 | |
| AutoTTS (β=1.0)Base Model=Qwen3-1.7B, Type=Discovered2026.05 | 49 | 0.0613 | |
| AutoTTS (β=0.5)Base Model=Qwen3-1.7B, Type=Discovered2026.05 | 46.7 | 0.0328 | |
| Parallel-ProbeBase Model=Qwen3-1.7B, Type=Handcrafted2026.05 | 44.7 | 0.0776 | |
| SC @ 64Base Model=Qwen3-1.7B, Type=Handcrafted2026.05 | 44.4 | 0.1054 | |
| ASCBase Model=Qwen3-1.7B, Type=Handcrafted2026.05 | 44.4 | 0.0601 | |
| ESCBase Model=Qwen3-1.7B, Type=Handcrafted2026.05 | 44.4 | 0.0914 | |
| AutoTTS (β=1.0)Base Model=Qwen3-0.6B, Type=Discovered2026.05 | 31.1 | 0.0475 | |
| Parallel-ProbeBase Model=Qwen3-0.6B, Type=Handcrafted2026.05 | 29.7 | 0.0698 | |
| SC @ 64Base Model=Qwen3-0.6B, Type=Handcrafted2026.05 | 28.9 | 0.0891 | |
| ASCBase Model=Qwen3-0.6B, Type=Handcrafted2026.05 | 28.9 | 0.0654 | |
| ESCBase Model=Qwen3-0.6B, Type=Handcrafted2026.05 | 28.9 | 0.0869 | |
| AutoTTS (β=0.5)Base Model=Qwen3-0.6B, Type=Discovered2026.05 | 28.6 | 0.025 | |
| Tool-Star-SFTBackbone=Qwen2.5-7B, Training Paradigm=SFT-only, SFT Trajectory Count=54K2026.01 | 10 | — | |
| AutoTrajBackbone=Qwen2.5-7B, Training Paradigm=SFT–RL TIR, SFT Trajectory Count=13K2026.01 | 10 | — | |
| AutoTIRBackbone=Qwen2.5-7B, Training Paradigm=RL-only2026.01 | 6.67 | — | |
| Tool-StarBackbone=Qwen2.5-7B, Training Paradigm=SFT–RL TIR, SFT Trajectory Count=54K2026.01 | 6.67 | — | |
| ReSearchBackbone=Qwen2.5-7B, Training Paradigm=RL-only2026.01 | 3.33 | — | |
| ToRLBackbone=Qwen2.5-7B, Training Paradigm=RL-only2026.01 | 3.33 | — | |
| Vanilla SFT-RL TIRBackbone=Qwen2.5-7B, Training Paradigm=SFT–RL TIR2026.01 | 3.33 | — | |
| Qwen2.5-7B-InstructBackbone=Qwen2.5-7B, Training Paradigm=Instruct2026.01 | 0 | — | |
| R1-SearcherBackbone=Qwen2.5-7B, Training Paradigm=RL-only2026.01 | 0 | — |